projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/decode: Hoist shader-db stats to shared decode
[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_short(enum mali_attr_mode mode)
418 {
419 switch(mode) {
420 /* TODO: Combine to just "instanced" once this can be done
421 * unambiguously in all known cases */
422 case MALI_ATTR_POT_DIVIDE:
423 return "instanced_pot";
424 case MALI_ATTR_MODULO:
425 return "instanced_mod";
426 case MALI_ATTR_NPOT_DIVIDE:
427 return "instanced_npot";
428 case MALI_ATTR_IMAGE:
429 return "image";
430 case MALI_ATTR_INTERNAL:
431 return "internal";
432 default:
433 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
434 return "";
435 }
436 }
437
438 #undef DEFINE_CASE
439
440 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
441 static char *
442 pandecode_wrap_mode(enum mali_wrap_mode op)
443 {
444 switch (op) {
445 DEFINE_CASE(REPEAT);
446 DEFINE_CASE(CLAMP_TO_EDGE);
447 DEFINE_CASE(CLAMP_TO_BORDER);
448 DEFINE_CASE(MIRRORED_REPEAT);
449
450 default:
451 pandecode_msg("XXX: invalid wrap mode %X\n", op);
452 return "";
453 }
454 }
455 #undef DEFINE_CASE
456
457 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
458 static char *
459 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
460 {
461 switch (fmt) {
462 DEFINE_CASE(TILED);
463 DEFINE_CASE(UNKNOWN);
464 DEFINE_CASE(LINEAR);
465 DEFINE_CASE(AFBC);
466
467 default:
468 unreachable("Invalid case");
469 }
470 }
471 #undef DEFINE_CASE
472
473 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
474 static char *
475 pandecode_exception_access(enum mali_exception_access access)
476 {
477 switch (access) {
478 DEFINE_CASE(NONE);
479 DEFINE_CASE(EXECUTE);
480 DEFINE_CASE(READ);
481 DEFINE_CASE(WRITE);
482
483 default:
484 unreachable("Invalid case");
485 }
486 }
487 #undef DEFINE_CASE
488
489 /* Midgard's tiler descriptor is embedded within the
490 * larger FBD */
491
492 static void
493 pandecode_midgard_tiler_descriptor(
494 const struct midgard_tiler_descriptor *t,
495 unsigned width,
496 unsigned height,
497 bool is_fragment)
498 {
499 pandecode_log(".tiler = {\n");
500 pandecode_indent++;
501
502 if (t->hierarchy_mask == MALI_TILER_DISABLED)
503 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
504 else
505 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
506
507 /* We know this name from the kernel, but we never see it nonzero */
508
509 if (t->flags)
510 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
511
512 MEMORY_PROP(t, polygon_list);
513
514 /* The body is offset from the base of the polygon list */
515 assert(t->polygon_list_body > t->polygon_list);
516 unsigned body_offset = t->polygon_list_body - t->polygon_list;
517
518 /* It needs to fit inside the reported size */
519 assert(t->polygon_list_size >= body_offset);
520
521 /* Check that we fit */
522 struct pandecode_mapped_memory *plist =
523 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
524
525 assert(t->polygon_list_size <= plist->length);
526
527 /* Now that we've sanity checked, we'll try to calculate the sizes
528 * ourselves for comparison */
529
530 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
531 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
532
533 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
534 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
535 ref_header, ref_size);
536 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
537 pandecode_msg("body offset %d\n", body_offset);
538 }
539
540 /* The tiler heap has a start and end specified -- it should be
541 * identical to what we have in the BO. The exception is if tiling is
542 * disabled. */
543
544 MEMORY_PROP(t, heap_start);
545 assert(t->heap_end >= t->heap_start);
546
547 struct pandecode_mapped_memory *heap =
548 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
549
550 unsigned heap_size = t->heap_end - t->heap_start;
551
552 /* Tiling is enabled with a special flag */
553 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
554 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
555
556 bool tiling_enabled = hierarchy_mask;
557
558 if (tiling_enabled) {
559 /* When tiling is enabled, the heap should be a tight fit */
560 unsigned heap_offset = t->heap_start - heap->gpu_va;
561 if ((heap_offset + heap_size) != heap->length) {
562 pandecode_msg("XXX: heap size %d (expected %d)\n",
563 heap_size, heap->length - heap_offset);
564 }
565
566 /* We should also have no other flags */
567 if (tiler_flags)
568 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
569 } else {
570 /* When tiling is disabled, we should have that flag and no others */
571
572 if (tiler_flags != MALI_TILER_DISABLED) {
573 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
574 tiler_flags);
575 }
576
577 /* We should also have an empty heap */
578 if (heap_size) {
579 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
580 heap_size);
581 }
582
583 /* Disabled tiling is used only for clear-only jobs, which are
584 * purely FRAGMENT, so we should never see this for
585 * non-FRAGMENT descriptors. */
586
587 if (!is_fragment)
588 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
589 }
590
591 /* We've never seen weights used in practice, but we know from the
592 * kernel these fields is there */
593
594 bool nonzero_weights = false;
595
596 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
597 nonzero_weights |= t->weights[w] != 0x0;
598 }
599
600 if (nonzero_weights) {
601 pandecode_log(".weights = {");
602
603 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
604 pandecode_log("%d, ", t->weights[w]);
605 }
606
607 pandecode_log("},");
608 }
609
610 pandecode_indent--;
611 pandecode_log("}\n");
612 }
613
614 static void
615 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
616 {
617 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
618 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
619
620 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
621 pandecode_indent++;
622
623 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
624 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
625
626 pandecode_log(".format = ");
627 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
628 pandecode_log_cont(",\n");
629
630 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
631 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
632
633 MEMORY_PROP(s, framebuffer);
634 pandecode_prop("stride = %d", s->stride);
635
636 /* Earlier in the actual commandstream -- right before width -- but we
637 * delay to flow nicer */
638
639 pandecode_log(".clear_flags = ");
640 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
641 pandecode_log_cont(",\n");
642
643 if (s->depth_buffer | s->depth_buffer_enable) {
644 MEMORY_PROP(s, depth_buffer);
645 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
646 }
647
648 if (s->stencil_buffer | s->stencil_buffer_enable) {
649 MEMORY_PROP(s, stencil_buffer);
650 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
651 }
652
653 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
654 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
655 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
656 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
657 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
658 }
659
660 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
661 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
662 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
663 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
664 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
665 }
666
667 if (s->clear_stencil) {
668 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
669 }
670
671 MEMORY_PROP(s, unknown_address_0);
672 const struct midgard_tiler_descriptor t = s->tiler;
673 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
674
675 pandecode_indent--;
676 pandecode_log("};\n");
677
678 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
679 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
680 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
681 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
682
683 printf(".zero3 = {");
684
685 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
686 printf("%X, ", s->zero3[i]);
687
688 printf("},\n");
689
690 printf(".zero6 = {");
691
692 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
693 printf("%X, ", s->zero6[i]);
694
695 printf("},\n");
696 }
697
698 static void
699 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
700 {
701 pandecode_log(".unknown%d = {", name);
702
703 for (int i = 0; i < count; ++i)
704 printf("%X, ", slide[i]);
705
706 pandecode_log("},\n");
707 }
708
709 #define SHORT_SLIDE(num) \
710 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
711
712 static void
713 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
714 {
715 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
716 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
717
718 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
719 pandecode_indent++;
720
721 SHORT_SLIDE(1);
722
723 pandecode_indent--;
724 printf("},\n");
725 }
726
727 /* Extracts the number of components associated with a Mali format */
728
729 static unsigned
730 pandecode_format_component_count(enum mali_format fmt)
731 {
732 /* Mask out the format class */
733 unsigned top = fmt & 0b11100000;
734
735 switch (top) {
736 case MALI_FORMAT_SNORM:
737 case MALI_FORMAT_UINT:
738 case MALI_FORMAT_UNORM:
739 case MALI_FORMAT_SINT:
740 return ((fmt >> 3) & 3) + 1;
741 default:
742 /* TODO: Validate */
743 return 4;
744 }
745 }
746
747 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
748
749 static unsigned
750 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
751 {
752 unsigned mask = 0;
753 assert(MALI_CHANNEL_RED == 0);
754
755 for (unsigned c = 0; c < 4; ++c) {
756 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
757
758 if (chan <= MALI_CHANNEL_ALPHA)
759 mask |= (1 << chan);
760 }
761
762 return mask;
763 }
764
765 /* Validates that a (format, swizzle) pair is valid, in the sense that the
766 * swizzle doesn't access any components that are undefined in the format.
767 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
768 * omitted */
769
770 static bool
771 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
772 {
773 unsigned nr_comp = pandecode_format_component_count(fmt);
774 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
775 unsigned valid_mask = (1 << nr_comp) - 1;
776 unsigned invalid_mask = ~valid_mask;
777
778 if (access_mask & invalid_mask) {
779 pandecode_msg("XXX: invalid components accessed\n");
780 return false;
781 }
782
783 /* Check for the default non-swizzling swizzle so we can suppress
784 * useless printing for the defaults */
785
786 unsigned default_swizzles[4] = {
787 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
788 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
789 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
790 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
791 };
792
793 return (swizzle == default_swizzles[nr_comp - 1]);
794 }
795
796 /* Maps MALI_RGBA32F to rgba32f, etc */
797
798 static void
799 pandecode_format_short(enum mali_format fmt, bool srgb)
800 {
801 /* We want a type-like format, so cut off the initial MALI_ */
802 char *format = pandecode_format(fmt);
803 format += strlen("MALI_");
804
805 unsigned len = strlen(format);
806 char *lower_format = calloc(1, len + 1);
807
808 for (unsigned i = 0; i < len; ++i)
809 lower_format[i] = tolower(format[i]);
810
811 /* Sanity check sRGB flag is applied to RGB, per the name */
812 if (srgb && lower_format[0] != 'r')
813 pandecode_msg("XXX: sRGB applied to non-colour format\n");
814
815 /* Just prefix with an s, so you get formats like srgba8_unorm */
816 if (srgb)
817 pandecode_log_cont("s");
818
819 pandecode_log_cont("%s", lower_format);
820 free(lower_format);
821 }
822
823 static void
824 pandecode_swizzle(unsigned swizzle, enum mali_format format)
825 {
826 /* First, do some validation */
827 bool trivial_swizzle = pandecode_validate_format_swizzle(
828 format, swizzle);
829
830 if (trivial_swizzle)
831 return;
832
833 /* Next, print the swizzle */
834 pandecode_log_cont(".");
835
836 static const char components[] = "rgba01";
837
838 for (unsigned c = 0; c < 4; ++c) {
839 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
840
841 if (chan >= MALI_CHANNEL_RESERVED_0) {
842 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
843 continue;
844 }
845 pandecode_log_cont("%c", components[chan]);
846 }
847 }
848
849 static void
850 pandecode_rt_format(struct mali_rt_format format)
851 {
852 pandecode_log(".format = {\n");
853 pandecode_indent++;
854
855 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
856 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
857 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
858
859 pandecode_prop("block = %s",
860 pandecode_mfbd_block_format(format.block));
861
862 /* TODO: Map formats so we can check swizzles and print nicely */
863 pandecode_log("swizzle");
864 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
865 pandecode_log_cont(",\n");
866
867 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
868 MALI_NEGATIVE(format.nr_channels));
869
870 pandecode_log(".flags = ");
871 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
872 pandecode_log_cont(",\n");
873
874 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
875 * which is a faster hardware-based alternative to the wallpaper method
876 * to preserve framebuffer contents across frames. In practice, MFBD
877 * preload is buggy on Midgard, and so this is a chicken bit. If this
878 * bit isn't set, most likely something broke unrelated to preload */
879
880 if (!format.no_preload) {
881 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
882 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
883 }
884
885 if (format.zero)
886 pandecode_prop("zero = 0x%" PRIx32, format.zero);
887
888 pandecode_indent--;
889 pandecode_log("},\n");
890 }
891
892 static void
893 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
894 {
895 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
896 pandecode_indent++;
897
898 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
899 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
900 struct pandecode_mapped_memory *mem =
901 pandecode_find_mapped_gpu_mem_containing(rt_va);
902 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
903
904 pandecode_log("{\n");
905 pandecode_indent++;
906
907 pandecode_rt_format(rt->format);
908
909 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
910 pandecode_log(".afbc = {\n");
911 pandecode_indent++;
912
913 char *a = pointer_as_memory_reference(rt->afbc.metadata);
914 pandecode_prop("metadata = %s", a);
915 free(a);
916
917 pandecode_prop("stride = %d", rt->afbc.stride);
918 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
919
920 pandecode_indent--;
921 pandecode_log("},\n");
922 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
923 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
924 rt->afbc.metadata,
925 rt->afbc.stride,
926 rt->afbc.unk);
927 }
928
929 MEMORY_PROP(rt, framebuffer);
930 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
931
932 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
933 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
934 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
935 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
936 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
937 }
938
939 if (rt->zero1 || rt->zero2 || rt->zero3) {
940 pandecode_msg("XXX: render target zeros tripped\n");
941 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
942 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
943 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
944 }
945
946 pandecode_indent--;
947 pandecode_log("},\n");
948 }
949
950 pandecode_indent--;
951 pandecode_log("};\n");
952 }
953
954 static unsigned
955 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
956 {
957 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
958 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
959
960 if (fb->sample_locations) {
961 /* The blob stores all possible sample locations in a single buffer
962 * allocated on startup, and just switches the pointer when switching
963 * MSAA state. For now, we just put the data into the cmdstream, but we
964 * should do something like what the blob does with a real driver.
965 *
966 * There seem to be 32 slots for sample locations, followed by another
967 * 16. The second 16 is just the center location followed by 15 zeros
968 * in all the cases I've identified (maybe shader vs. depth/color
969 * samples?).
970 */
971
972 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
973
974 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
975
976 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
977 pandecode_indent++;
978
979 for (int i = 0; i < 32 + 16; i++) {
980 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
981 }
982
983 pandecode_indent--;
984 pandecode_log("};\n");
985 }
986
987 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
988 pandecode_indent++;
989
990 pandecode_prop("unk0 = 0x%x", fb->unk0);
991
992 if (fb->sample_locations)
993 pandecode_prop("sample_locations = sample_locations_%d", job_no);
994
995 /* Assume that unknown1 was emitted in the last job for
996 * now */
997 MEMORY_PROP(fb, unknown1);
998
999 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1000 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1001 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1002 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1003
1004 pandecode_prop("unk1 = 0x%x", fb->unk1);
1005 pandecode_prop("unk2 = 0x%x", fb->unk2);
1006 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1007 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1008
1009 pandecode_log(".mfbd_flags = ");
1010 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1011 pandecode_log_cont(",\n");
1012
1013 if (fb->clear_stencil)
1014 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1015
1016 if (fb->clear_depth)
1017 pandecode_prop("clear_depth = %f", fb->clear_depth);
1018
1019 /* TODO: What is this? Let's not blow up.. */
1020 if (fb->unknown2 != 0x1F)
1021 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1022
1023 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1024 MEMORY_PROP(fb, scratchpad);
1025 const struct midgard_tiler_descriptor t = fb->tiler;
1026 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
1027
1028 if (fb->zero3 || fb->zero4) {
1029 pandecode_msg("XXX: framebuffer zeros tripped\n");
1030 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1031 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1032 }
1033
1034 pandecode_indent--;
1035 pandecode_log("};\n");
1036
1037 gpu_va += sizeof(struct bifrost_framebuffer);
1038
1039 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
1040 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1041 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1042
1043 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1044 pandecode_indent++;
1045
1046 MEMORY_PROP(fbx, checksum);
1047
1048 if (fbx->checksum_stride)
1049 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1050
1051 pandecode_log(".flags = ");
1052 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
1053 pandecode_log_cont(",\n");
1054
1055 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
1056 pandecode_log(".ds_afbc = {\n");
1057 pandecode_indent++;
1058
1059 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1060 pandecode_prop("depth_stencil_afbc_stride = %d",
1061 fbx->ds_afbc.depth_stencil_afbc_stride);
1062 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1063
1064 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1065 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1066 pandecode_prop("zero1 = 0x%" PRIx32,
1067 fbx->ds_afbc.zero1);
1068 pandecode_prop("padding = 0x%" PRIx64,
1069 fbx->ds_afbc.padding);
1070 }
1071
1072 pandecode_indent--;
1073 pandecode_log("},\n");
1074 } else {
1075 pandecode_log(".ds_linear = {\n");
1076 pandecode_indent++;
1077
1078 if (fbx->ds_linear.depth) {
1079 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1080 pandecode_prop("depth_stride = %d",
1081 fbx->ds_linear.depth_stride);
1082 }
1083
1084 if (fbx->ds_linear.stencil) {
1085 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1086 pandecode_prop("stencil_stride = %d",
1087 fbx->ds_linear.stencil_stride);
1088 }
1089
1090 if (fbx->ds_linear.depth_stride_zero ||
1091 fbx->ds_linear.stencil_stride_zero ||
1092 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1093 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1094 pandecode_prop("depth_stride_zero = 0x%x",
1095 fbx->ds_linear.depth_stride_zero);
1096 pandecode_prop("stencil_stride_zero = 0x%x",
1097 fbx->ds_linear.stencil_stride_zero);
1098 pandecode_prop("zero1 = 0x%" PRIx32,
1099 fbx->ds_linear.zero1);
1100 pandecode_prop("zero2 = 0x%" PRIx32,
1101 fbx->ds_linear.zero2);
1102 }
1103
1104 pandecode_indent--;
1105 pandecode_log("},\n");
1106 }
1107
1108 if (fbx->zero3 || fbx->zero4) {
1109 pandecode_msg("XXX: fb_extra zeros tripped\n");
1110 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1111 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1112 }
1113
1114 pandecode_indent--;
1115 pandecode_log("};\n");
1116
1117 gpu_va += sizeof(struct bifrost_fb_extra);
1118 }
1119
1120 if (is_fragment)
1121 pandecode_render_target(gpu_va, job_no, fb);
1122
1123 /* Passback the render target count */
1124 return MALI_NEGATIVE(fb->rt_count_1);
1125 }
1126
1127 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1128 * count */
1129
1130 static void
1131 pandecode_padded_vertices(unsigned shift, unsigned k)
1132 {
1133 unsigned odd = 2*k + 1;
1134 unsigned pot = 1 << shift;
1135 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1136 }
1137
1138 /* Given a magic divisor, recover what we were trying to divide by.
1139 *
1140 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1141 * 0 <= m < 2^N, for N bits in m.
1142 *
1143 * Let q represent the number we would like to divide by.
1144 *
1145 * By definition of a magic divisor for N-bit unsigned integers (a number you
1146 * multiply by to magically get division), m is a number such that:
1147 *
1148 * (m * x) & (2^N - 1) = floor(x/q).
1149 * for all x on Z where 0 <= x < 2^N
1150 *
1151 * Ignore the case where any of the above values equals zero; it is irrelevant
1152 * for our purposes (instanced arrays).
1153 *
1154 * Choose x = q. Then:
1155 *
1156 * (m * x) & (2^N - 1) = floor(x/q).
1157 * (m * q) & (2^N - 1) = floor(q/q).
1158 *
1159 * floor(q/q) = floor(1) = 1, therefore:
1160 *
1161 * (m * q) & (2^N - 1) = 1
1162 *
1163 * Recall the identity that the bitwise AND of one less than a power-of-two
1164 * equals the modulo with that power of two, i.e. for all x:
1165 *
1166 * x & (2^N - 1) = x % N
1167 *
1168 * Therefore:
1169 *
1170 * mq % (2^N) = 1
1171 *
1172 * By definition, a modular multiplicative inverse of a number m is the number
1173 * q such that with respect to a modulos M:
1174 *
1175 * mq % M = 1
1176 *
1177 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1178 *
1179 */
1180
1181 static void
1182 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1183 {
1184 #if 0
1185 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1186 * shift) the most lame way possible... just repeatedly add.
1187 * Asymptoptically slow but nobody cares in practice, unless you have
1188 * massive numbers of vertices or high divisors. */
1189
1190 unsigned inverse = 0;
1191
1192 /* Magic implicitly has the highest bit set */
1193 magic |= (1 << 31);
1194
1195 /* Depending on rounding direction */
1196 if (extra)
1197 magic++;
1198
1199 for (;;) {
1200 uint32_t product = magic * inverse;
1201
1202 if (shift) {
1203 product >>= shift;
1204 }
1205
1206 if (product == 1)
1207 break;
1208
1209 ++inverse;
1210 }
1211
1212 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1213
1214 /* Recall we're supposed to divide by (gl_level_divisor *
1215 * padded_num_vertices) */
1216
1217 unsigned padded_num_vertices = inverse / orig_divisor;
1218
1219 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1220 #endif
1221 }
1222
1223 static void
1224 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1225 mali_ptr addr, int job_no, char *suffix,
1226 int count, bool varying)
1227 {
1228 char *prefix = varying ? "varying" : "attribute";
1229
1230 if (!addr) {
1231 pandecode_msg("no %s\n", prefix);
1232 return;
1233 }
1234
1235 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1236
1237 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1238 pandecode_indent++;
1239
1240 for (int i = 0; i < count; ++i) {
1241 enum mali_attr_mode mode = attr[i].elements & 7;
1242
1243 if (mode == MALI_ATTR_UNUSED)
1244 pandecode_msg("XXX: unused attribute record\n");
1245
1246 pandecode_make_indent();
1247
1248 /* For non-linear records, we need to print the type of record */
1249 if (mode != MALI_ATTR_LINEAR)
1250 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1251
1252 /* Print the name to link with attr_meta */
1253 pandecode_log_cont("%s_%d", prefix, i);
1254
1255 /* Print the stride and size */
1256 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1257
1258 /* Check: the size must be divisible by the stride */
1259 if (attr[i].size % attr[i].stride)
1260 pandecode_msg("XXX: size not divisible by stride\n");
1261
1262 /* TODO: Sanity check the quotient itself -- it should equal
1263 * vertex count (or something computed from it for instanced)
1264 * which means we can check and elide */
1265
1266 /* Finally, print the pointer */
1267 mali_ptr raw_elements = attr[i].elements & ~7;
1268 char *a = pointer_as_memory_reference(raw_elements);
1269 pandecode_log_cont(" = (%s);\n", a);
1270 free(a);
1271
1272 /* Check the pointer */
1273 pandecode_validate_buffer(raw_elements, attr[i].size);
1274
1275 /* shift/extra_flags exist only for instanced */
1276 if (attr[i].shift | attr[i].extra_flags) {
1277 if (mode == MALI_ATTR_LINEAR)
1278 pandecode_msg("XXX: instancing fields set for linear\n");
1279
1280 pandecode_prop("shift = %d", attr[i].shift);
1281 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1282 }
1283
1284 /* Decode further where possible */
1285
1286 if (mode == MALI_ATTR_MODULO) {
1287 pandecode_padded_vertices(
1288 attr[i].shift,
1289 attr[i].extra_flags);
1290 }
1291
1292 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1293 i++;
1294 pandecode_log("{\n");
1295 pandecode_indent++;
1296 pandecode_prop("unk = 0x%x", attr[i].unk);
1297 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1298 if (attr[i].zero != 0)
1299 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1300 pandecode_prop("divisor = %d", attr[i].divisor);
1301 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1302 pandecode_indent--;
1303 pandecode_log("}, \n");
1304 }
1305
1306 }
1307
1308 pandecode_indent--;
1309 pandecode_log("};\n");
1310 }
1311
1312 static mali_ptr
1313 pandecode_shader_address(const char *name, mali_ptr ptr)
1314 {
1315 /* TODO: Decode flags */
1316 mali_ptr shader_ptr = ptr & ~15;
1317
1318 char *a = pointer_as_memory_reference(shader_ptr);
1319 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1320 free(a);
1321
1322 return shader_ptr;
1323 }
1324
1325 static void
1326 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1327 {
1328 unsigned any_nonzero =
1329 stencil->ref | stencil->mask | stencil->func |
1330 stencil->sfail | stencil->dpfail | stencil->dppass;
1331
1332 if (any_nonzero == 0)
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 /* constant/equation is in a union */
1415 if (!blend->shader)
1416 return 0;
1417
1418 pandecode_log(".blend = {\n");
1419 pandecode_indent++;
1420
1421 if (is_shader) {
1422 pandecode_shader_address("shader", blend->shader);
1423 } else {
1424 pandecode_blend_equation(&blend->equation);
1425 pandecode_prop("constant = %f", blend->constant);
1426 }
1427
1428 pandecode_indent--;
1429 pandecode_log("},\n");
1430
1431 /* Return blend shader to disassemble if present */
1432 return is_shader ? (blend->shader & ~0xF) : 0;
1433 }
1434
1435 static mali_ptr
1436 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1437 {
1438 struct midgard_blend_rt *b =
1439 ((struct midgard_blend_rt *) descs) + rt_no;
1440
1441 /* Flags determine presence of blend shader */
1442 bool is_shader = (b->flags & 0xF) >= 0x2;
1443
1444 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1445 pandecode_indent++;
1446
1447 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1448
1449 union midgard_blend blend = b->blend;
1450 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1451
1452 pandecode_indent--;
1453 pandecode_log("};\n");
1454
1455 return shader;
1456 }
1457
1458 /* Attributes and varyings have descriptor records, which contain information
1459 * about their format and ordering with the attribute/varying buffers. We'll
1460 * want to validate that the combinations specified are self-consistent.
1461 */
1462
1463 static int
1464 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1465 {
1466 char base[128];
1467 char *prefix = varying ? "varying" : "attribute";
1468 unsigned max_index = 0;
1469 snprintf(base, sizeof(base), "%s_meta", prefix);
1470
1471 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1472 pandecode_indent++;
1473
1474 struct mali_attr_meta *attr_meta;
1475 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1476
1477 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1478
1479 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1480 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1481 sizeof(*attr_mem));
1482
1483 /* If the record is discard, it should be zero for everything else */
1484
1485 if (attr_meta->format == MALI_VARYING_DISCARD) {
1486 uint64_t zero =
1487 attr_meta->index |
1488 attr_meta->unknown1 |
1489 attr_meta->unknown3 |
1490 attr_meta->src_offset;
1491
1492 if (zero)
1493 pandecode_msg("XXX: expected empty record for varying discard\n");
1494
1495 /* We want to look for a literal 0000 swizzle -- this
1496 * is not encoded with all zeroes, however */
1497
1498 enum mali_channel z = MALI_CHANNEL_ZERO;
1499 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1500 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1501
1502 if (!good_swizzle)
1503 pandecode_msg("XXX: expected zero swizzle for discard\n");
1504
1505 if (!varying)
1506 pandecode_msg("XXX: cannot discard attribute\n");
1507
1508 /* If we're all good, omit the record */
1509 if (!zero && varying && good_swizzle) {
1510 pandecode_log("/* discarded varying */\n");
1511 continue;
1512 }
1513 }
1514
1515 if (attr_meta->index > max_index)
1516 max_index = attr_meta->index;
1517
1518 if (attr_meta->unknown1 != 0x2) {
1519 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1520 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1521 }
1522
1523 if (attr_meta->unknown3) {
1524 pandecode_msg("XXX: unexpected unknown3 set\n");
1525 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1526 }
1527
1528 pandecode_make_indent();
1529 pandecode_format_short(attr_meta->format, false);
1530 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1531
1532 if (attr_meta->src_offset)
1533 pandecode_log_cont("[%u]", attr_meta->src_offset);
1534
1535 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1536
1537 pandecode_log_cont(";\n");
1538 }
1539
1540 pandecode_indent--;
1541 pandecode_log("};\n");
1542
1543 return count ? (max_index + 1) : 0;
1544 }
1545
1546 static void
1547 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1548 {
1549 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1550
1551 if (imem) {
1552 /* Indices are literally just a u32 array :) */
1553
1554 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1555
1556 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1557 pandecode_indent++;
1558
1559 for (unsigned i = 0; i < (index_count + 1); i += 3)
1560 pandecode_log("%d, %d, %d,\n",
1561 indices[i],
1562 indices[i + 1],
1563 indices[i + 2]);
1564
1565 pandecode_indent--;
1566 pandecode_log("};\n");
1567 }
1568 }
1569
1570 /* return bits [lo, hi) of word */
1571 static u32
1572 bits(u32 word, u32 lo, u32 hi)
1573 {
1574 if (hi - lo >= 32)
1575 return word; // avoid undefined behavior with the shift
1576
1577 return (word >> lo) & ((1 << (hi - lo)) - 1);
1578 }
1579
1580 static void
1581 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1582 {
1583 pandecode_log_cont("{\n");
1584 pandecode_indent++;
1585
1586 /* Decode invocation_count. See the comment before the definition of
1587 * invocation_count for an explanation.
1588 */
1589
1590 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1591 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1592 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1593
1594 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1595 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1596 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1597
1598 /* Even though we have this decoded, we want to ensure that the
1599 * representation is "unique" so we don't lose anything by printing only
1600 * the final result. More specifically, we need to check that we were
1601 * passed something in canonical form, since the definition per the
1602 * hardware is inherently not unique. How? Well, take the resulting
1603 * decode and pack it ourselves! If it is bit exact with what we
1604 * decoded, we're good to go. */
1605
1606 struct mali_vertex_tiler_prefix ref;
1607 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1608
1609 bool canonical =
1610 (p->invocation_count == ref.invocation_count) &&
1611 (p->size_y_shift == ref.size_y_shift) &&
1612 (p->size_z_shift == ref.size_z_shift) &&
1613 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1614 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1615 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1616 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1617
1618 if (!canonical) {
1619 pandecode_msg("XXX: non-canonical workgroups packing\n");
1620 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1621 ref.invocation_count,
1622 ref.size_y_shift,
1623 ref.size_z_shift,
1624 ref.workgroups_x_shift,
1625 ref.workgroups_y_shift,
1626 ref.workgroups_z_shift,
1627 ref.workgroups_x_shift_2);
1628
1629 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1630 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1631 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1632 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1633 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1634 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1635 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1636 }
1637
1638 /* Regardless, print the decode */
1639 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1640 size_x, size_y, size_z,
1641 groups_x, groups_y, groups_z);
1642
1643 /* TODO: Decode */
1644 if (p->unknown_draw)
1645 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1646
1647 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1648
1649 if (p->draw_mode != MALI_DRAW_NONE)
1650 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1651
1652 /* Index count only exists for tiler jobs anyway */
1653
1654 if (p->index_count)
1655 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1656
1657 if (p->offset_bias_correction)
1658 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1659
1660 /* TODO: Figure out what this is. It's not zero */
1661 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1662
1663 pandecode_indent--;
1664 pandecode_log("},\n");
1665 }
1666
1667 static void
1668 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1669 {
1670 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1671 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1672
1673 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1674 pubufs, job_no);
1675 pandecode_indent++;
1676
1677 for (int i = 0; i < ubufs_count; i++) {
1678 pandecode_log("{\n");
1679 pandecode_indent++;
1680
1681 unsigned size = (ubufs[i].size + 1) * 16;
1682 mali_ptr addr = ubufs[i].ptr << 2;
1683
1684 pandecode_validate_buffer(addr, size);
1685
1686 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1687 pandecode_prop("size = %u", size);
1688 pandecode_prop("ptr = (%s) >> 2", ptr);
1689 pandecode_indent--;
1690 pandecode_log("},\n");
1691 free(ptr);
1692 }
1693
1694 pandecode_indent--;
1695 pandecode_log("};\n");
1696 }
1697
1698 static void
1699 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1700 {
1701
1702 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1703
1704 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1705
1706 if (scratchpad->zero) {
1707 pandecode_msg("XXX: scratchpad zero tripped");
1708 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1709 }
1710
1711 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1712 pandecode_indent++;
1713
1714 pandecode_prop("flags = 0x%x", scratchpad->flags);
1715 MEMORY_PROP(scratchpad, gpu_scratchpad);
1716
1717 pandecode_indent--;
1718 pandecode_log("};\n");
1719 }
1720
1721 static unsigned shader_id = 0;
1722
1723 static struct midgard_disasm_stats
1724 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1725 bool is_bifrost, unsigned nr_regs)
1726 {
1727 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1728 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1729
1730 /* Compute maximum possible size */
1731 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1732
1733 /* Print some boilerplate to clearly denote the assembly (which doesn't
1734 * obey indentation rules), and actually do the disassembly! */
1735
1736 printf("\n\n");
1737
1738 struct midgard_disasm_stats stats;
1739
1740 if (is_bifrost) {
1741 disassemble_bifrost(code, sz, false);
1742
1743 /* TODO: Extend stats to Bifrost */
1744 stats.texture_count = -1;
1745 stats.sampler_count = -1;
1746 stats.attribute_count = -1;
1747 stats.varying_count = -1;
1748 stats.uniform_count = -1;
1749 stats.uniform_buffer_count = -1;
1750 stats.work_count = -1;
1751
1752 stats.instruction_count = 0;
1753 stats.bundle_count = 0;
1754 stats.quadword_count = 0;
1755 } else {
1756 stats = disassemble_midgard(code, sz);
1757 stats.work_count = nr_regs;
1758 }
1759
1760 /* Print shader-db stats */
1761
1762 unsigned nr_threads =
1763 (stats.work_count <= 4) ? 4 :
1764 (stats.work_count <= 8) ? 2 :
1765 1;
1766
1767 printf("shader%d - %s shader: "
1768 "%u inst, %u bundles, %u quadwords, "
1769 "%u registers, %u threads, 0 loops\n\n\n",
1770 shader_id++,
1771 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX",
1772 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1773 stats.work_count, nr_threads);
1774
1775
1776 return stats;
1777 }
1778
1779 static void
1780 pandecode_texture(mali_ptr u,
1781 struct pandecode_mapped_memory *tmem,
1782 unsigned job_no, unsigned tex)
1783 {
1784 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1785
1786 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1787 pandecode_indent++;
1788
1789 struct mali_texture_format f = t->format;
1790
1791 /* See the definiton of enum mali_texture_type */
1792
1793 bool is_cube = f.type == MALI_TEX_CUBE;
1794 unsigned dimension = is_cube ? 2 : f.type;
1795
1796 pandecode_make_indent();
1797
1798 /* TODO: Are there others? */
1799 bool is_zs = f.format == MALI_Z32_UNORM;
1800
1801 /* Recall Z/S switched the meaning of linear/tiled .. */
1802 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1803 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1804
1805 /* Print the layout. Default is linear; a modifier can denote AFBC or
1806 * u-interleaved/tiled modes */
1807
1808 if (f.layout == MALI_TEXTURE_AFBC)
1809 pandecode_log_cont("afbc");
1810 else if (f.layout == MALI_TEXTURE_TILED)
1811 pandecode_log_cont(is_zs ? "linear" : "tiled");
1812 else if (f.layout == MALI_TEXTURE_LINEAR)
1813 pandecode_log_cont("linear");
1814 else
1815 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1816
1817 pandecode_swizzle(t->swizzle, f.format);
1818 pandecode_log_cont(" ");
1819
1820 /* Distinguish cube/2D with modifier */
1821
1822 if (is_cube)
1823 pandecode_log_cont("cube ");
1824
1825 pandecode_format_short(f.format, f.srgb);
1826 pandecode_swizzle(f.swizzle, f.format);
1827
1828 /* All four width/height/depth/array_size dimensions are present
1829 * regardless of the type of texture, but it is an error to have
1830 * non-zero dimensions for unused dimensions. Verify this. array_size
1831 * can always be set, as can width. */
1832
1833 if (t->height && dimension < 2)
1834 pandecode_msg("XXX: nonzero height for <2D texture\n");
1835
1836 if (t->depth && dimension < 3)
1837 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1838
1839 /* Print only the dimensions that are actually there */
1840
1841 pandecode_log_cont(": %d", t->width + 1);
1842
1843 if (dimension >= 2)
1844 pandecode_log_cont("x%u", t->height + 1);
1845
1846 if (dimension >= 3)
1847 pandecode_log_cont("x%u", t->depth + 1);
1848
1849 if (t->array_size)
1850 pandecode_log_cont("[%u]", t->array_size + 1);
1851
1852 if (t->levels)
1853 pandecode_log_cont(" mip %u", t->levels);
1854
1855 pandecode_log_cont("\n");
1856
1857 if (f.unknown1 | f.zero) {
1858 pandecode_msg("XXX: texture format zero tripped\n");
1859 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1860 pandecode_prop("zero = %" PRId32, f.zero);
1861 }
1862
1863 if (!f.unknown2) {
1864 pandecode_msg("XXX: expected unknown texture bit set\n");
1865 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
1866 }
1867
1868 if (t->swizzle_zero) {
1869 pandecode_msg("XXX: swizzle zero tripped\n");
1870 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1871 }
1872
1873 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
1874 pandecode_msg("XXX: texture zero tripped\n");
1875 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1876 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1877 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1878 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1879 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1880 }
1881
1882 pandecode_log(".payload = {\n");
1883 pandecode_indent++;
1884
1885 /* A bunch of bitmap pointers follow.
1886 * We work out the correct number,
1887 * based on the mipmap/cubemap
1888 * properties, but dump extra
1889 * possibilities to futureproof */
1890
1891 int bitmap_count = MALI_NEGATIVE(t->levels);
1892
1893 /* Miptree for each face */
1894 if (f.type == MALI_TEX_CUBE)
1895 bitmap_count *= 6;
1896
1897 /* Array of textures */
1898 bitmap_count *= MALI_NEGATIVE(t->array_size);
1899
1900 /* Stride for each element */
1901 if (f.manual_stride)
1902 bitmap_count *= 2;
1903
1904 /* Sanity check the size */
1905 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
1906 assert (bitmap_count <= max_count);
1907
1908 for (int i = 0; i < bitmap_count; ++i) {
1909 /* How we dump depends if this is a stride or a pointer */
1910
1911 if (f.manual_stride && (i & 1)) {
1912 /* signed 32-bit snuck in as a 64-bit pointer */
1913 uint64_t stride_set = t->payload[i];
1914 uint32_t clamped_stride = stride_set;
1915 int32_t stride = clamped_stride;
1916 assert(stride_set == clamped_stride);
1917 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1918 } else {
1919 char *a = pointer_as_memory_reference(t->payload[i]);
1920 pandecode_log("%s, \n", a);
1921 free(a);
1922 }
1923 }
1924
1925 pandecode_indent--;
1926 pandecode_log("},\n");
1927
1928 pandecode_indent--;
1929 pandecode_log("};\n");
1930 }
1931
1932 static void
1933 pandecode_vertex_tiler_postfix_pre(
1934 const struct mali_vertex_tiler_postfix *p,
1935 int job_no, enum mali_job_type job_type,
1936 char *suffix, bool is_bifrost)
1937 {
1938 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1939 struct pandecode_mapped_memory *attr_mem;
1940
1941 unsigned rt_count = 1;
1942
1943 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1944 * are the only things actually needed from the FBD, vertex/tiler jobs
1945 * no longer reference the FBD -- instead, this field points to some
1946 * info about the scratchpad.
1947 */
1948 if (is_bifrost)
1949 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1950 else if (p->framebuffer & MALI_MFBD)
1951 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1952 else if (job_type == JOB_TYPE_COMPUTE)
1953 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1954 else
1955 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1956
1957 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1958 int texture_count = 0, sampler_count = 0;
1959
1960 if (shader_meta_ptr) {
1961 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1962 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1963
1964 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1965 pandecode_indent++;
1966
1967 /* Save for dumps */
1968 attribute_count = s->attribute_count;
1969 varying_count = s->varying_count;
1970 texture_count = s->texture_count;
1971 sampler_count = s->sampler_count;
1972
1973 if (is_bifrost) {
1974 uniform_count = s->bifrost2.uniform_count;
1975 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1976 } else {
1977 uniform_count = s->midgard1.uniform_buffer_count;
1978 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1979 }
1980
1981 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1982
1983 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1984 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1985 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1986 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1987
1988 unsigned nr_registers = 0;
1989
1990 if (is_bifrost) {
1991 pandecode_log(".bifrost1 = {\n");
1992 pandecode_indent++;
1993
1994 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1995 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1996
1997 pandecode_indent--;
1998 pandecode_log("},\n");
1999 } else {
2000 pandecode_log(".midgard1 = {\n");
2001 pandecode_indent++;
2002
2003 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
2004 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
2005 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
2006 nr_registers = s->midgard1.work_count;
2007
2008 pandecode_log(".flags = ");
2009 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2010 pandecode_log_cont(",\n");
2011
2012 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2013
2014 pandecode_indent--;
2015 pandecode_log("},\n");
2016 }
2017
2018 if (s->depth_units || s->depth_factor) {
2019 pandecode_prop("depth_factor = %f", s->depth_factor);
2020 pandecode_prop("depth_units = %f", s->depth_units);
2021 }
2022
2023 if (s->alpha_coverage) {
2024 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2025 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2026
2027 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2028 invert_alpha_coverage ? "~" : "",
2029 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2030 }
2031
2032 if (s->unknown2_3 || s->unknown2_4) {
2033 pandecode_log(".unknown2_3 = ");
2034
2035 int unknown2_3 = s->unknown2_3;
2036 int unknown2_4 = s->unknown2_4;
2037
2038 /* We're not quite sure what these flags mean without the depth test, if anything */
2039
2040 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
2041 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2042 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2043
2044 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2045 }
2046
2047 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2048 pandecode_log_cont(",\n");
2049
2050 pandecode_log(".unknown2_4 = ");
2051 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2052 pandecode_log_cont(",\n");
2053 }
2054
2055 if (s->stencil_mask_front || s->stencil_mask_back) {
2056 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2057 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2058 }
2059
2060 pandecode_stencil("front", &s->stencil_front);
2061 pandecode_stencil("back", &s->stencil_back);
2062
2063 if (is_bifrost) {
2064 pandecode_log(".bifrost2 = {\n");
2065 pandecode_indent++;
2066
2067 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2068 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2069 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2070 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2071
2072 pandecode_indent--;
2073 pandecode_log("},\n");
2074 } else if (s->midgard2.unknown2_7) {
2075 pandecode_log(".midgard2 = {\n");
2076 pandecode_indent++;
2077
2078 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2079 pandecode_indent--;
2080 pandecode_log("},\n");
2081 }
2082
2083 if (s->unknown2_8)
2084 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2085
2086 if (!is_bifrost) {
2087 /* TODO: Blend shaders routing/disasm */
2088
2089 union midgard_blend blend = s->blend;
2090 pandecode_midgard_blend(&blend, false);
2091 }
2092
2093 pandecode_indent--;
2094 pandecode_log("};\n");
2095
2096 /* MRT blend fields are used whenever MFBD is used, with
2097 * per-RT descriptors */
2098
2099 if (job_type == JOB_TYPE_TILER) {
2100 void* blend_base = (void *) (s + 1);
2101
2102 for (unsigned i = 0; i < rt_count; i++) {
2103 mali_ptr shader = 0;
2104
2105 if (is_bifrost)
2106 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2107 else
2108 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2109
2110 if (shader & ~0xF)
2111 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
2112 }
2113 }
2114
2115 if (shader_ptr & ~0xF)
2116 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
2117 } else
2118 pandecode_msg("<no shader>\n");
2119
2120 if (p->viewport) {
2121 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2122 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2123
2124 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2125 pandecode_indent++;
2126
2127 pandecode_prop("clip_minx = %f", f->clip_minx);
2128 pandecode_prop("clip_miny = %f", f->clip_miny);
2129 pandecode_prop("clip_minz = %f", f->clip_minz);
2130 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2131 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2132 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2133
2134 /* Only the higher coordinates are MALI_POSITIVE scaled */
2135
2136 pandecode_prop("viewport0 = { %d, %d }",
2137 f->viewport0[0], f->viewport0[1]);
2138
2139 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2140 f->viewport1[0] + 1, f->viewport1[1] + 1);
2141
2142 pandecode_indent--;
2143 pandecode_log("};\n");
2144 }
2145
2146 if (p->attribute_meta) {
2147 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2148
2149 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2150 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
2151 }
2152
2153 /* Varyings are encoded like attributes but not actually sent; we just
2154 * pass a zero buffer with the right stride/size set, (or whatever)
2155 * since the GPU will write to it itself */
2156
2157 if (p->varying_meta) {
2158 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2159 }
2160
2161 if (p->varyings) {
2162 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2163
2164 /* Number of descriptors depends on whether there are
2165 * non-internal varyings */
2166
2167 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
2168 }
2169
2170 if (p->uniform_buffers) {
2171 if (uniform_buffer_count)
2172 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2173 else
2174 pandecode_msg("XXX: UBOs specified but not referenced\n");
2175 } else if (uniform_buffer_count)
2176 pandecode_msg("XXX: UBOs referenced but not specified\n");
2177
2178 /* We don't want to actually dump uniforms, but we do need to validate
2179 * that the counts we were given are sane */
2180
2181 if (p->uniforms) {
2182 if (uniform_count)
2183 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
2184 else
2185 pandecode_msg("XXX: Uniforms specified but not referenced");
2186 } else if (uniform_count)
2187 pandecode_msg("XXX: UBOs referenced but not specified\n");
2188
2189 if (p->texture_trampoline) {
2190 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2191
2192 if (mmem) {
2193 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2194
2195 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2196 pandecode_indent++;
2197
2198 for (int tex = 0; tex < texture_count; ++tex) {
2199 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2200 char *a = pointer_as_memory_reference(*u);
2201 pandecode_log("%s,\n", a);
2202 free(a);
2203 }
2204
2205 pandecode_indent--;
2206 pandecode_log("};\n");
2207
2208 /* Now, finally, descend down into the texture descriptor */
2209 for (unsigned tex = 0; tex < texture_count; ++tex) {
2210 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2211 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2212 if (tmem)
2213 pandecode_texture(*u, tmem, job_no, tex);
2214 }
2215 }
2216 }
2217
2218 if (p->sampler_descriptor) {
2219 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2220
2221 if (smem) {
2222 struct mali_sampler_descriptor *s;
2223
2224 mali_ptr d = p->sampler_descriptor;
2225
2226 for (int i = 0; i < sampler_count; ++i) {
2227 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2228
2229 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2230 pandecode_indent++;
2231
2232 pandecode_log(".filter_mode = ");
2233 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2234 pandecode_log_cont(",\n");
2235
2236 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2237 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2238
2239 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2240 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2241 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2242
2243 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2244
2245 if (s->zero || s->zero2) {
2246 pandecode_msg("XXX: sampler zero tripped\n");
2247 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2248 }
2249
2250 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2251
2252 pandecode_prop("border_color = { %f, %f, %f, %f }",
2253 s->border_color[0],
2254 s->border_color[1],
2255 s->border_color[2],
2256 s->border_color[3]);
2257
2258 pandecode_indent--;
2259 pandecode_log("};\n");
2260 }
2261 }
2262 }
2263 }
2264
2265 static void
2266 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2267 {
2268 if (!(p->position_varying || p->occlusion_counter || p->flags))
2269 return;
2270
2271 pandecode_log(".postfix = {\n");
2272 pandecode_indent++;
2273
2274 MEMORY_PROP(p, position_varying);
2275 MEMORY_PROP(p, occlusion_counter);
2276
2277 if (p->flags)
2278 pandecode_prop("flags = %d", p->flags);
2279
2280 pandecode_indent--;
2281 pandecode_log("},\n");
2282 }
2283
2284 static void
2285 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2286 {
2287 pandecode_log_cont("{\n");
2288 pandecode_indent++;
2289
2290 pandecode_prop("unk2 = 0x%x", v->unk2);
2291
2292 if (v->zero0 || v->zero1) {
2293 pandecode_msg("XXX: vertex only zero tripped");
2294 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2295 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2296 }
2297
2298 pandecode_indent--;
2299 pandecode_log("}\n");
2300 }
2301
2302 static void
2303 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2304 {
2305
2306 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2307 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2308
2309 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2310 pandecode_indent++;
2311
2312 if (h->zero) {
2313 pandecode_msg("XXX: tiler heap zero tripped\n");
2314 pandecode_prop("zero = 0x%x", h->zero);
2315 }
2316
2317 for (int i = 0; i < 12; i++) {
2318 if (h->zeros[i] != 0) {
2319 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2320 i, h->zeros[i]);
2321 }
2322 }
2323
2324 pandecode_prop("heap_size = 0x%x", h->heap_size);
2325 MEMORY_PROP(h, tiler_heap_start);
2326 MEMORY_PROP(h, tiler_heap_free);
2327
2328 /* this might point to the beginning of another buffer, when it's
2329 * really the end of the tiler heap buffer, so we have to be careful
2330 * here. but for zero length, we need the same pointer.
2331 */
2332
2333 if (h->tiler_heap_end == h->tiler_heap_start) {
2334 MEMORY_PROP(h, tiler_heap_start);
2335 } else {
2336 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2337 pandecode_prop("tiler_heap_end = %s + 1", a);
2338 free(a);
2339 }
2340
2341 pandecode_indent--;
2342 pandecode_log("};\n");
2343 }
2344
2345 static void
2346 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2347 {
2348 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2349 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2350
2351 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2352
2353 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2354 pandecode_indent++;
2355
2356 if (t->zero0 || t->zero1) {
2357 pandecode_msg("XXX: tiler meta zero tripped\n");
2358 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2359 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2360 }
2361
2362 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2363 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2364
2365 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2366 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2367
2368 for (int i = 0; i < 12; i++) {
2369 if (t->zeros[i] != 0) {
2370 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2371 i, t->zeros[i]);
2372 }
2373 }
2374
2375 pandecode_indent--;
2376 pandecode_log("};\n");
2377 }
2378
2379 static void
2380 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2381 {
2382 pandecode_log(".gl_enables = ");
2383
2384 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2385
2386 pandecode_log_cont(",\n");
2387 }
2388
2389 static void
2390 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2391 {
2392 if (u.pointer == 0x0)
2393 return;
2394
2395 pandecode_log(".primitive_size = {\n");
2396 pandecode_indent++;
2397
2398 if (constant) {
2399 pandecode_prop("constant = %f", u.constant);
2400 } else {
2401 MEMORY_PROP((&u), pointer);
2402 }
2403
2404 pandecode_indent--;
2405 pandecode_log("},\n");
2406 }
2407
2408 static void
2409 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2410 {
2411 pandecode_log_cont("{\n");
2412 pandecode_indent++;
2413
2414 /* TODO: gl_PointSize on Bifrost */
2415 pandecode_primitive_size(t->primitive_size, true);
2416
2417 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2418
2419 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2420 || t->zero6 || t->zero7 || t->zero8) {
2421 pandecode_msg("XXX: tiler only zero tripped\n");
2422 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2423 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2424 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2425 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2426 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2427 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2428 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2429 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2430 }
2431
2432 pandecode_indent--;
2433 pandecode_log("},\n");
2434 }
2435
2436 static int
2437 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2438 const struct pandecode_mapped_memory *mem,
2439 mali_ptr payload, int job_no)
2440 {
2441 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2442
2443 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2444
2445 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2446 pandecode_indent++;
2447
2448 pandecode_log(".prefix = ");
2449 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2450
2451 pandecode_log(".vertex = ");
2452 pandecode_vertex_only_bfr(&v->vertex);
2453
2454 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2455
2456 pandecode_indent--;
2457 pandecode_log("};\n");
2458
2459 return sizeof(*v);
2460 }
2461
2462 static int
2463 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2464 const struct pandecode_mapped_memory *mem,
2465 mali_ptr payload, int job_no)
2466 {
2467 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2468
2469 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2470
2471 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2472 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2473
2474 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2475 pandecode_indent++;
2476
2477 pandecode_log(".prefix = ");
2478 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2479
2480 pandecode_log(".tiler = ");
2481 pandecode_tiler_only_bfr(&t->tiler, job_no);
2482
2483 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2484
2485 pandecode_indent--;
2486 pandecode_log("};\n");
2487
2488 return sizeof(*t);
2489 }
2490
2491 static int
2492 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2493 const struct pandecode_mapped_memory *mem,
2494 mali_ptr payload, int job_no)
2495 {
2496 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2497
2498 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2499
2500 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2501
2502 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2503 pandecode_indent++;
2504
2505 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2506 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2507
2508 bool instanced = v->instance_shift || v->instance_odd;
2509 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2510
2511 pandecode_log(".prefix = ");
2512 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2513
2514 pandecode_gl_enables(v->gl_enables, h->job_type);
2515
2516 if (v->instance_shift || v->instance_odd) {
2517 pandecode_prop("instance_shift = 0x%d /* %d */",
2518 v->instance_shift, 1 << v->instance_shift);
2519 pandecode_prop("instance_odd = 0x%X /* %d */",
2520 v->instance_odd, (2 * v->instance_odd) + 1);
2521
2522 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2523 }
2524
2525 if (v->offset_start)
2526 pandecode_prop("offset_start = %d", v->offset_start);
2527
2528 if (v->zero5) {
2529 pandecode_msg("XXX: midgard payload zero tripped\n");
2530 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2531 }
2532
2533 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2534
2535 pandecode_indent--;
2536 pandecode_log("};\n");
2537
2538 return sizeof(*v);
2539 }
2540
2541 static int
2542 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2543 mali_ptr payload, int job_no,
2544 bool is_bifrost)
2545 {
2546 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2547
2548 bool fbd_dumped = false;
2549
2550 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2551 /* Only SFBDs are understood, not MFBDs. We're speculating,
2552 * based on the versioning, kernel code, etc, that the
2553 * difference is between Single FrameBuffer Descriptor and
2554 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2555 * up with multi-framebuffer support being added (T7xx onwards,
2556 * including Gxx). In any event, there's some field shuffling
2557 * that we haven't looked into yet. */
2558
2559 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2560 fbd_dumped = true;
2561 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2562 /* We don't know if Bifrost supports SFBD's at all, since the
2563 * driver never uses them. And the format is different from
2564 * Midgard anyways, due to the tiler heap and scratchpad being
2565 * moved out into separate structures, so it's not clear what a
2566 * Bifrost SFBD would even look like without getting an actual
2567 * trace, which appears impossible.
2568 */
2569
2570 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2571 fbd_dumped = true;
2572 }
2573
2574 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2575 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2576 pandecode_indent++;
2577
2578 /* See the comments by the macro definitions for mathematical context
2579 * on why this is so weird */
2580
2581 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2582 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2583
2584 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2585 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2586 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2587
2588 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2589 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2590 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2591
2592 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2593 * we have to fallback on offsets. */
2594
2595 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2596
2597 /* TODO: Decode */
2598 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2599
2600 if (fbd_dumped)
2601 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2602 fbd_type, extra_flags);
2603 else {
2604 char *a = pointer_as_memory_reference(p);
2605 pandecode_prop("framebuffer = %s | %s | 0x%X", a, fbd_type, extra_flags);
2606 free(a);
2607 }
2608
2609 pandecode_indent--;
2610 pandecode_log("};\n");
2611
2612 return sizeof(*s);
2613 }
2614
2615 static int job_descriptor_number = 0;
2616
2617 int
2618 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2619 {
2620 struct mali_job_descriptor_header *h;
2621
2622 int start_number = 0;
2623
2624 bool first = true;
2625 bool last_size;
2626
2627 do {
2628 struct pandecode_mapped_memory *mem =
2629 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2630
2631 void *payload;
2632
2633 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2634
2635 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2636 * high 32-bits of the 64-bit pointer are reused to store
2637 * something else.
2638 */
2639 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2640 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2641 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2642
2643 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2644 MALI_PAYLOAD_SIZE);
2645
2646 int job_no = job_descriptor_number++;
2647
2648 if (first)
2649 start_number = job_no;
2650
2651 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2652 pandecode_indent++;
2653
2654 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2655
2656 /* Save for next job fixing */
2657 last_size = h->job_descriptor_size;
2658
2659 if (h->job_descriptor_size)
2660 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2661
2662 if (h->exception_status && h->exception_status != 0x1)
2663 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2664 h->exception_status,
2665 (h->exception_status >> 16) & 0xFFFF,
2666 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2667 h->exception_status & 0xFF);
2668
2669 if (h->first_incomplete_task)
2670 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2671
2672 if (h->fault_pointer)
2673 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2674
2675 if (h->job_barrier)
2676 pandecode_prop("job_barrier = %d", h->job_barrier);
2677
2678 pandecode_prop("job_index = %d", h->job_index);
2679
2680 if (h->unknown_flags)
2681 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2682
2683 if (h->job_dependency_index_1)
2684 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2685
2686 if (h->job_dependency_index_2)
2687 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2688
2689 pandecode_indent--;
2690 pandecode_log("};\n");
2691
2692 /* Do not touch the field yet -- decode the payload first, and
2693 * don't touch that either. This is essential for the uploads
2694 * to occur in sequence and therefore be dynamically allocated
2695 * correctly. Do note the size, however, for that related
2696 * reason. */
2697
2698 switch (h->job_type) {
2699 case JOB_TYPE_SET_VALUE: {
2700 struct mali_payload_set_value *s = payload;
2701 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2702 pandecode_indent++;
2703 MEMORY_PROP(s, out);
2704 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2705 pandecode_indent--;
2706 pandecode_log("};\n");
2707
2708 break;
2709 }
2710
2711 case JOB_TYPE_TILER:
2712 case JOB_TYPE_VERTEX:
2713 case JOB_TYPE_COMPUTE:
2714 if (bifrost) {
2715 if (h->job_type == JOB_TYPE_TILER)
2716 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2717 else
2718 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2719 } else
2720 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2721
2722 break;
2723
2724 case JOB_TYPE_FRAGMENT:
2725 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2726 break;
2727
2728 default:
2729 break;
2730 }
2731
2732 /* Handle linkage */
2733
2734 if (!first) {
2735 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2736
2737 if (last_size)
2738 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2739 else
2740 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2741 }
2742
2743 first = false;
2744
2745 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2746
2747 return start_number;
2748 }