projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/decode: Add static bounds checking utility
[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 "decode.h"
33 #include "util/macros.h"
34 #include "util/u_math.h"
35
36 #include "pan_pretty_print.h"
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 int pandecode_jc(mali_ptr jc_gpu_va, bool bifrost);
43
44 #define MEMORY_PROP(obj, p) {\
45 if (obj->p) { \
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
48 free(a); \
49 } \
50 }
51
52 #define MEMORY_PROP_DIR(obj, p) {\
53 if (obj.p) { \
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
56 free(a); \
57 } \
58 }
59
60 /* Semantic logging type.
61 *
62 * Raw: for raw messages to be printed as is.
63 * Message: for helpful information to be commented out in replays.
64 * Property: for properties of a struct
65 *
66 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
67 */
68
69 enum pandecode_log_type {
70 PANDECODE_RAW,
71 PANDECODE_MESSAGE,
72 PANDECODE_PROPERTY
73 };
74
75 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
76 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
77 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
78
79 unsigned pandecode_indent = 0;
80
81 static void
82 pandecode_make_indent(void)
83 {
84 for (unsigned i = 0; i < pandecode_indent; ++i)
85 printf(" ");
86 }
87
88 static void
89 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
90 {
91 va_list ap;
92
93 pandecode_make_indent();
94
95 if (type == PANDECODE_MESSAGE)
96 printf("// ");
97 else if (type == PANDECODE_PROPERTY)
98 printf(".");
99
100 va_start(ap, format);
101 vprintf(format, ap);
102 va_end(ap);
103
104 if (type == PANDECODE_PROPERTY)
105 printf(",\n");
106 }
107
108 static void
109 pandecode_log_cont(const char *format, ...)
110 {
111 va_list ap;
112
113 va_start(ap, format);
114 vprintf(format, ap);
115 va_end(ap);
116 }
117
118 /* To check for memory safety issues, validates that the given pointer in GPU
119 * memory is valid, containing at least sz bytes. The goal is to eliminate
120 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
121 * overruns) by statically validating pointers.
122 */
123
124 static void
125 pandecode_validate_buffer(mali_ptr addr, size_t sz)
126 {
127 if (!addr) {
128 pandecode_msg("XXX: null pointer deref");
129 return;
130 }
131
132 /* Find a BO */
133
134 struct pandecode_mapped_memory *bo =
135 pandecode_find_mapped_gpu_mem_containing(addr);
136
137 if (!bo) {
138 pandecode_msg("XXX: invalid memory dereference\n");
139 return;
140 }
141
142 /* Bounds check */
143
144 unsigned offset = addr - bo->gpu_va;
145 unsigned total = offset + sz;
146
147 if (total > bo->length) {
148 pandecode_msg("XXX: buffer overrun."
149 "Chunk of size %d at offset %d in buffer of size %d. "
150 "Overrun by %d bytes.",
151 sz, offset, bo->length, total - bo->length);
152 return;
153 }
154 }
155
156 struct pandecode_flag_info {
157 u64 flag;
158 const char *name;
159 };
160
161 static void
162 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
163 u64 flags)
164 {
165 bool decodable_flags_found = false;
166
167 for (int i = 0; flag_info[i].name; i++) {
168 if ((flags & flag_info[i].flag) != flag_info[i].flag)
169 continue;
170
171 if (!decodable_flags_found) {
172 decodable_flags_found = true;
173 } else {
174 pandecode_log_cont(" | ");
175 }
176
177 pandecode_log_cont("%s", flag_info[i].name);
178
179 flags &= ~flag_info[i].flag;
180 }
181
182 if (decodable_flags_found) {
183 if (flags)
184 pandecode_log_cont(" | 0x%" PRIx64, flags);
185 } else {
186 pandecode_log_cont("0x%" PRIx64, flags);
187 }
188 }
189
190 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
191 static const struct pandecode_flag_info gl_enable_flag_info[] = {
192 FLAG_INFO(OCCLUSION_QUERY),
193 FLAG_INFO(OCCLUSION_PRECISE),
194 FLAG_INFO(FRONT_CCW_TOP),
195 FLAG_INFO(CULL_FACE_FRONT),
196 FLAG_INFO(CULL_FACE_BACK),
197 {}
198 };
199 #undef FLAG_INFO
200
201 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
202 static const struct pandecode_flag_info clear_flag_info[] = {
203 FLAG_INFO(FAST),
204 FLAG_INFO(SLOW),
205 FLAG_INFO(SLOW_STENCIL),
206 {}
207 };
208 #undef FLAG_INFO
209
210 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
211 static const struct pandecode_flag_info mask_flag_info[] = {
212 FLAG_INFO(R),
213 FLAG_INFO(G),
214 FLAG_INFO(B),
215 FLAG_INFO(A),
216 {}
217 };
218 #undef FLAG_INFO
219
220 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
221 static const struct pandecode_flag_info u3_flag_info[] = {
222 FLAG_INFO(HAS_MSAA),
223 FLAG_INFO(CAN_DISCARD),
224 FLAG_INFO(HAS_BLEND_SHADER),
225 FLAG_INFO(DEPTH_TEST),
226 {}
227 };
228
229 static const struct pandecode_flag_info u4_flag_info[] = {
230 FLAG_INFO(NO_MSAA),
231 FLAG_INFO(NO_DITHER),
232 FLAG_INFO(DEPTH_RANGE_A),
233 FLAG_INFO(DEPTH_RANGE_B),
234 FLAG_INFO(STENCIL_TEST),
235 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
236 {}
237 };
238 #undef FLAG_INFO
239
240 #define FLAG_INFO(flag) { MALI_FRAMEBUFFER_##flag, "MALI_FRAMEBUFFER_" #flag }
241 static const struct pandecode_flag_info fb_fmt_flag_info[] = {
242 FLAG_INFO(MSAA_A),
243 FLAG_INFO(MSAA_B),
244 FLAG_INFO(MSAA_8),
245 {}
246 };
247 #undef FLAG_INFO
248
249 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
250 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
251 FLAG_INFO(MSAA),
252 FLAG_INFO(SRGB),
253 {}
254 };
255 #undef FLAG_INFO
256
257 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
258 static const struct pandecode_flag_info mfbd_extra_flag_info[] = {
259 FLAG_INFO(PRESENT),
260 FLAG_INFO(AFBC),
261 FLAG_INFO(ZS),
262 {}
263 };
264 #undef FLAG_INFO
265
266 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
267 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
268 FLAG_INFO(EARLY_Z),
269 FLAG_INFO(HELPER_INVOCATIONS),
270 FLAG_INFO(READS_TILEBUFFER),
271 FLAG_INFO(READS_ZS),
272 {}
273 };
274 #undef FLAG_INFO
275
276 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
277 static const struct pandecode_flag_info mfbd_flag_info [] = {
278 FLAG_INFO(DEPTH_WRITE),
279 FLAG_INFO(EXTRA),
280 {}
281 };
282 #undef FLAG_INFO
283
284 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
285 static const struct pandecode_flag_info sampler_flag_info [] = {
286 FLAG_INFO(MAG_NEAREST),
287 FLAG_INFO(MIN_NEAREST),
288 FLAG_INFO(MIP_LINEAR_1),
289 FLAG_INFO(MIP_LINEAR_2),
290 FLAG_INFO(NORM_COORDS),
291 {}
292 };
293 #undef FLAG_INFO
294
295 extern char *replace_fragment;
296 extern char *replace_vertex;
297
298 static char *
299 pandecode_job_type(enum mali_job_type type)
300 {
301 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
302
303 switch (type) {
304 DEFINE_CASE(NULL);
305 DEFINE_CASE(SET_VALUE);
306 DEFINE_CASE(CACHE_FLUSH);
307 DEFINE_CASE(COMPUTE);
308 DEFINE_CASE(VERTEX);
309 DEFINE_CASE(TILER);
310 DEFINE_CASE(FUSED);
311 DEFINE_CASE(FRAGMENT);
312
313 case JOB_NOT_STARTED:
314 return "NOT_STARTED";
315
316 default:
317 pandecode_log("Warning! Unknown job type %x\n", type);
318 return "!?!?!?";
319 }
320
321 #undef DEFINE_CASE
322 }
323
324 static char *
325 pandecode_draw_mode(enum mali_draw_mode mode)
326 {
327 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
328
329 switch (mode) {
330 DEFINE_CASE(DRAW_NONE);
331 DEFINE_CASE(POINTS);
332 DEFINE_CASE(LINES);
333 DEFINE_CASE(TRIANGLES);
334 DEFINE_CASE(TRIANGLE_STRIP);
335 DEFINE_CASE(TRIANGLE_FAN);
336 DEFINE_CASE(LINE_STRIP);
337 DEFINE_CASE(LINE_LOOP);
338 DEFINE_CASE(POLYGON);
339 DEFINE_CASE(QUADS);
340 DEFINE_CASE(QUAD_STRIP);
341
342 default:
343 return "MALI_TRIANGLES /* XXX: Unknown GL mode, check dump */";
344 }
345
346 #undef DEFINE_CASE
347 }
348
349 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
350 static char *
351 pandecode_func(enum mali_func mode)
352 {
353 switch (mode) {
354 DEFINE_CASE(NEVER);
355 DEFINE_CASE(LESS);
356 DEFINE_CASE(EQUAL);
357 DEFINE_CASE(LEQUAL);
358 DEFINE_CASE(GREATER);
359 DEFINE_CASE(NOTEQUAL);
360 DEFINE_CASE(GEQUAL);
361 DEFINE_CASE(ALWAYS);
362
363 default:
364 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
365 }
366 }
367 #undef DEFINE_CASE
368
369 /* Why is this duplicated? Who knows... */
370 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
371 static char *
372 pandecode_alt_func(enum mali_alt_func mode)
373 {
374 switch (mode) {
375 DEFINE_CASE(NEVER);
376 DEFINE_CASE(LESS);
377 DEFINE_CASE(EQUAL);
378 DEFINE_CASE(LEQUAL);
379 DEFINE_CASE(GREATER);
380 DEFINE_CASE(NOTEQUAL);
381 DEFINE_CASE(GEQUAL);
382 DEFINE_CASE(ALWAYS);
383
384 default:
385 return "MALI_FUNC_NEVER /* XXX: Unknown function, check dump */";
386 }
387 }
388 #undef DEFINE_CASE
389
390 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
391 static char *
392 pandecode_stencil_op(enum mali_stencil_op op)
393 {
394 switch (op) {
395 DEFINE_CASE(KEEP);
396 DEFINE_CASE(REPLACE);
397 DEFINE_CASE(ZERO);
398 DEFINE_CASE(INVERT);
399 DEFINE_CASE(INCR_WRAP);
400 DEFINE_CASE(DECR_WRAP);
401 DEFINE_CASE(INCR);
402 DEFINE_CASE(DECR);
403
404 default:
405 return "MALI_STENCIL_KEEP /* XXX: Unknown stencil op, check dump */";
406 }
407 }
408
409 #undef DEFINE_CASE
410
411 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
412 static char *pandecode_attr_mode(enum mali_attr_mode mode)
413 {
414 switch(mode) {
415 DEFINE_CASE(UNUSED);
416 DEFINE_CASE(LINEAR);
417 DEFINE_CASE(POT_DIVIDE);
418 DEFINE_CASE(MODULO);
419 DEFINE_CASE(NPOT_DIVIDE);
420 DEFINE_CASE(IMAGE);
421 DEFINE_CASE(INTERNAL);
422 default:
423 return "MALI_ATTR_UNUSED /* XXX: Unknown stencil op, check dump */";
424 }
425 }
426
427 #undef DEFINE_CASE
428
429 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
430 static char *
431 pandecode_channel(enum mali_channel channel)
432 {
433 switch (channel) {
434 DEFINE_CASE(RED);
435 DEFINE_CASE(GREEN);
436 DEFINE_CASE(BLUE);
437 DEFINE_CASE(ALPHA);
438 DEFINE_CASE(ZERO);
439 DEFINE_CASE(ONE);
440 DEFINE_CASE(RESERVED_0);
441 DEFINE_CASE(RESERVED_1);
442
443 default:
444 return "MALI_CHANNEL_ZERO /* XXX: Unknown channel, check dump */";
445 }
446 }
447 #undef DEFINE_CASE
448
449 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
450 static char *
451 pandecode_wrap_mode(enum mali_wrap_mode op)
452 {
453 switch (op) {
454 DEFINE_CASE(REPEAT);
455 DEFINE_CASE(CLAMP_TO_EDGE);
456 DEFINE_CASE(CLAMP_TO_BORDER);
457 DEFINE_CASE(MIRRORED_REPEAT);
458
459 default:
460 return "MALI_WRAP_REPEAT /* XXX: Unknown wrap mode, check dump */";
461 }
462 }
463 #undef DEFINE_CASE
464
465 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
466 static char *
467 pandecode_texture_type(enum mali_texture_type type)
468 {
469 switch (type) {
470 DEFINE_CASE(1D);
471 DEFINE_CASE(2D);
472 DEFINE_CASE(3D);
473 DEFINE_CASE(CUBE);
474
475 default:
476 unreachable("Unknown case");
477 }
478 }
479 #undef DEFINE_CASE
480
481 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
482 static char *
483 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
484 {
485 switch (fmt) {
486 DEFINE_CASE(TILED);
487 DEFINE_CASE(UNKNOWN);
488 DEFINE_CASE(LINEAR);
489 DEFINE_CASE(AFBC);
490
491 default:
492 unreachable("Invalid case");
493 }
494 }
495 #undef DEFINE_CASE
496
497 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
498 static char *
499 pandecode_exception_access(enum mali_exception_access fmt)
500 {
501 switch (fmt) {
502 DEFINE_CASE(NONE);
503 DEFINE_CASE(EXECUTE);
504 DEFINE_CASE(READ);
505 DEFINE_CASE(WRITE);
506
507 default:
508 unreachable("Invalid case");
509 }
510 }
511 #undef DEFINE_CASE
512
513 /* Midgard's tiler descriptor is embedded within the
514 * larger FBD */
515
516 static void
517 pandecode_midgard_tiler_descriptor(
518 const struct midgard_tiler_descriptor *t,
519 unsigned width,
520 unsigned height,
521 bool is_fragment)
522 {
523 pandecode_log(".tiler = {\n");
524 pandecode_indent++;
525
526 if (t->hierarchy_mask == MALI_TILER_DISABLED)
527 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
528 else
529 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
530
531 /* We know this name from the kernel, but we never see it nonzero */
532 if (t->flags)
533 pandecode_prop("flags = 0x%" PRIx16 " /* XXX: unexpected */", t->flags);
534
535 MEMORY_PROP(t, polygon_list);
536
537 /* The body is offset from the base of the polygon list */
538 assert(t->polygon_list_body > t->polygon_list);
539 unsigned body_offset = t->polygon_list_body - t->polygon_list;
540
541 /* It needs to fit inside the reported size */
542 assert(t->polygon_list_size >= body_offset);
543
544 /* Check that we fit */
545 struct pandecode_mapped_memory *plist =
546 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
547
548 assert(t->polygon_list_size <= plist->length);
549
550 /* Now that we've sanity checked, we'll try to calculate the sizes
551 * ourselves for comparison */
552
553 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
554 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
555
556 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
557 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
558 ref_header, ref_size);
559 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
560 pandecode_msg("body offset %d\n", body_offset);
561 }
562
563 /* The tiler heap has a start and end specified -- it should be
564 * identical to what we have in the BO. The exception is if tiling is
565 * disabled. */
566
567 MEMORY_PROP(t, heap_start);
568 assert(t->heap_end >= t->heap_start);
569
570 struct pandecode_mapped_memory *heap =
571 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
572
573 unsigned heap_size = t->heap_end - t->heap_start;
574
575 /* Tiling is enabled with a special flag */
576 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
577 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
578
579 bool tiling_enabled = hierarchy_mask;
580
581 if (tiling_enabled) {
582 /* When tiling is enabled, the heap should be a tight fit */
583 unsigned heap_offset = t->heap_start - heap->gpu_va;
584 if ((heap_offset + heap_size) != heap->length) {
585 pandecode_msg("XXX: heap size %d (expected %d)\n",
586 heap_size, heap->length - heap_offset);
587 }
588
589 /* We should also have no other flags */
590 if (tiler_flags)
591 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
592 } else {
593 /* When tiling is disabled, we should have that flag and no others */
594
595 if (tiler_flags != MALI_TILER_DISABLED) {
596 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
597 tiler_flags);
598 }
599
600 /* We should also have an empty heap */
601 if (heap_size) {
602 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
603 heap_size);
604 }
605
606 /* Disabled tiling is used only for clear-only jobs, which are
607 * purely FRAGMENT, so we should never see this for
608 * non-FRAGMENT descriptors. */
609
610 if (!is_fragment)
611 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
612 }
613
614 /* We've never seen weights used in practice, but we know from the
615 * kernel these fields is there */
616
617 bool nonzero_weights = false;
618
619 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
620 nonzero_weights |= t->weights[w] != 0x0;
621 }
622
623 if (nonzero_weights) {
624 pandecode_log(".weights = {");
625
626 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
627 pandecode_log("%d, ", t->weights[w]);
628 }
629
630 pandecode_log("},");
631 }
632
633 pandecode_indent--;
634 pandecode_log("}\n");
635 }
636
637 static void
638 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
639 {
640 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
641 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
642
643 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
644 pandecode_indent++;
645
646 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
647 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
648
649 pandecode_log(".format = ");
650 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
651 pandecode_log_cont(",\n");
652
653 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
654 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
655
656 MEMORY_PROP(s, framebuffer);
657 pandecode_prop("stride = %d", s->stride);
658
659 /* Earlier in the actual commandstream -- right before width -- but we
660 * delay to flow nicer */
661
662 pandecode_log(".clear_flags = ");
663 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
664 pandecode_log_cont(",\n");
665
666 if (s->depth_buffer | s->depth_buffer_enable) {
667 MEMORY_PROP(s, depth_buffer);
668 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
669 }
670
671 if (s->stencil_buffer | s->stencil_buffer_enable) {
672 MEMORY_PROP(s, stencil_buffer);
673 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
674 }
675
676 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
677 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
678 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
679 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
680 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
681 }
682
683 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
684 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
685 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
686 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
687 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
688 }
689
690 if (s->clear_stencil) {
691 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
692 }
693
694 MEMORY_PROP(s, unknown_address_0);
695 const struct midgard_tiler_descriptor t = s->tiler;
696 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
697
698 pandecode_indent--;
699 pandecode_log("};\n");
700
701 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
702 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
703 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
704 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
705
706 printf(".zero3 = {");
707
708 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
709 printf("%X, ", s->zero3[i]);
710
711 printf("},\n");
712
713 printf(".zero6 = {");
714
715 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
716 printf("%X, ", s->zero6[i]);
717
718 printf("},\n");
719 }
720
721 static void
722 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
723 {
724 pandecode_log(".unknown%d = {", name);
725
726 for (int i = 0; i < count; ++i)
727 printf("%X, ", slide[i]);
728
729 pandecode_log("},\n");
730 }
731
732 #define SHORT_SLIDE(num) \
733 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
734
735 static void
736 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
737 {
738 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
739 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
740
741 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
742 pandecode_indent++;
743
744 SHORT_SLIDE(1);
745
746 pandecode_indent--;
747 printf("},\n");
748 }
749
750 static void
751 pandecode_swizzle(unsigned swizzle)
752 {
753 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
754 pandecode_channel((swizzle >> 0) & 0x7),
755 pandecode_channel((swizzle >> 3) & 0x7),
756 pandecode_channel((swizzle >> 6) & 0x7),
757 pandecode_channel((swizzle >> 9) & 0x7));
758 }
759
760 static void
761 pandecode_rt_format(struct mali_rt_format format)
762 {
763 pandecode_log(".format = {\n");
764 pandecode_indent++;
765
766 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
767 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
768 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
769
770 pandecode_prop("block = %s",
771 pandecode_mfbd_block_format(format.block));
772
773 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
774 MALI_NEGATIVE(format.nr_channels));
775
776 pandecode_log(".flags = ");
777 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
778 pandecode_log_cont(",\n");
779
780 pandecode_swizzle(format.swizzle);
781
782 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
783
784 if (format.zero)
785 pandecode_prop("zero = 0x%" PRIx32, format.zero);
786
787 pandecode_indent--;
788 pandecode_log("},\n");
789 }
790
791 static void
792 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
793 {
794 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
795 pandecode_indent++;
796
797 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
798 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
799 struct pandecode_mapped_memory *mem =
800 pandecode_find_mapped_gpu_mem_containing(rt_va);
801 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
802
803 pandecode_log("{\n");
804 pandecode_indent++;
805
806 pandecode_rt_format(rt->format);
807
808 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
809 pandecode_log(".afbc = {\n");
810 pandecode_indent++;
811
812 char *a = pointer_as_memory_reference(rt->afbc.metadata);
813 pandecode_prop("metadata = %s", a);
814 free(a);
815
816 pandecode_prop("stride = %d", rt->afbc.stride);
817 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
818
819 pandecode_indent--;
820 pandecode_log("},\n");
821 } else {
822 pandecode_log(".chunknown = {\n");
823 pandecode_indent++;
824
825 pandecode_prop("unk = 0x%" PRIx64, rt->chunknown.unk);
826
827 char *a = pointer_as_memory_reference(rt->chunknown.pointer);
828 pandecode_prop("pointer = %s", a);
829 free(a);
830
831 pandecode_indent--;
832 pandecode_log("},\n");
833 }
834
835 MEMORY_PROP(rt, framebuffer);
836 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
837
838 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
839 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
840 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
841 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
842 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
843 }
844
845 if (rt->zero1 || rt->zero2 || rt->zero3) {
846 pandecode_msg("render target zeros tripped\n");
847 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
848 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
849 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
850 }
851
852 pandecode_indent--;
853 pandecode_log("},\n");
854 }
855
856 pandecode_indent--;
857 pandecode_log("};\n");
858 }
859
860 static unsigned
861 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
862 {
863 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
864 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
865
866 if (fb->sample_locations) {
867 /* The blob stores all possible sample locations in a single buffer
868 * allocated on startup, and just switches the pointer when switching
869 * MSAA state. For now, we just put the data into the cmdstream, but we
870 * should do something like what the blob does with a real driver.
871 *
872 * There seem to be 32 slots for sample locations, followed by another
873 * 16. The second 16 is just the center location followed by 15 zeros
874 * in all the cases I've identified (maybe shader vs. depth/color
875 * samples?).
876 */
877
878 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
879
880 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
881
882 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
883 pandecode_indent++;
884
885 for (int i = 0; i < 32 + 16; i++) {
886 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
887 }
888
889 pandecode_indent--;
890 pandecode_log("};\n");
891 }
892
893 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
894 pandecode_indent++;
895
896 pandecode_prop("unk0 = 0x%x", fb->unk0);
897
898 if (fb->sample_locations)
899 pandecode_prop("sample_locations = sample_locations_%d", job_no);
900
901 /* Assume that unknown1 was emitted in the last job for
902 * now */
903 MEMORY_PROP(fb, unknown1);
904
905 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
906 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
907 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
908 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
909
910 pandecode_prop("unk1 = 0x%x", fb->unk1);
911 pandecode_prop("unk2 = 0x%x", fb->unk2);
912 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
913 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
914
915 pandecode_log(".mfbd_flags = ");
916 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
917 pandecode_log_cont(",\n");
918
919 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
920 pandecode_prop("clear_depth = %f", fb->clear_depth);
921
922 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
923 MEMORY_PROP(fb, scratchpad);
924 const struct midgard_tiler_descriptor t = fb->tiler;
925 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
926
927 if (fb->zero3 || fb->zero4) {
928 pandecode_msg("framebuffer zeros tripped\n");
929 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
930 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
931 }
932
933 pandecode_indent--;
934 pandecode_log("};\n");
935
936 gpu_va += sizeof(struct bifrost_framebuffer);
937
938 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
939 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
940 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
941
942 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
943 pandecode_indent++;
944
945 MEMORY_PROP(fbx, checksum);
946
947 if (fbx->checksum_stride)
948 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
949
950 pandecode_log(".flags = ");
951 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
952 pandecode_log_cont(",\n");
953
954 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
955 pandecode_log(".ds_afbc = {\n");
956 pandecode_indent++;
957
958 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
959 pandecode_prop("depth_stencil_afbc_stride = %d",
960 fbx->ds_afbc.depth_stencil_afbc_stride);
961 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
962
963 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
964 pandecode_msg("Depth/stencil AFBC zeros tripped\n");
965 pandecode_prop("zero1 = 0x%" PRIx32,
966 fbx->ds_afbc.zero1);
967 pandecode_prop("padding = 0x%" PRIx64,
968 fbx->ds_afbc.padding);
969 }
970
971 pandecode_indent--;
972 pandecode_log("},\n");
973 } else {
974 pandecode_log(".ds_linear = {\n");
975 pandecode_indent++;
976
977 if (fbx->ds_linear.depth) {
978 MEMORY_PROP_DIR(fbx->ds_linear, depth);
979 pandecode_prop("depth_stride = %d",
980 fbx->ds_linear.depth_stride);
981 }
982
983 if (fbx->ds_linear.stencil) {
984 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
985 pandecode_prop("stencil_stride = %d",
986 fbx->ds_linear.stencil_stride);
987 }
988
989 if (fbx->ds_linear.depth_stride_zero ||
990 fbx->ds_linear.stencil_stride_zero ||
991 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
992 pandecode_msg("Depth/stencil zeros tripped\n");
993 pandecode_prop("depth_stride_zero = 0x%x",
994 fbx->ds_linear.depth_stride_zero);
995 pandecode_prop("stencil_stride_zero = 0x%x",
996 fbx->ds_linear.stencil_stride_zero);
997 pandecode_prop("zero1 = 0x%" PRIx32,
998 fbx->ds_linear.zero1);
999 pandecode_prop("zero2 = 0x%" PRIx32,
1000 fbx->ds_linear.zero2);
1001 }
1002
1003 pandecode_indent--;
1004 pandecode_log("},\n");
1005 }
1006
1007 if (fbx->zero3 || fbx->zero4) {
1008 pandecode_msg("fb_extra zeros tripped\n");
1009 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1010 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1011 }
1012
1013 pandecode_indent--;
1014 pandecode_log("};\n");
1015
1016 gpu_va += sizeof(struct bifrost_fb_extra);
1017 }
1018
1019 if (is_fragment)
1020 pandecode_render_target(gpu_va, job_no, fb);
1021
1022 /* Passback the render target count */
1023 return MALI_NEGATIVE(fb->rt_count_1);
1024 }
1025
1026 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1027 * count */
1028
1029 static void
1030 pandecode_padded_vertices(unsigned shift, unsigned k)
1031 {
1032 unsigned odd = 2*k + 1;
1033 unsigned pot = 1 << shift;
1034 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1035 }
1036
1037 /* Given a magic divisor, recover what we were trying to divide by.
1038 *
1039 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1040 * 0 <= m < 2^N, for N bits in m.
1041 *
1042 * Let q represent the number we would like to divide by.
1043 *
1044 * By definition of a magic divisor for N-bit unsigned integers (a number you
1045 * multiply by to magically get division), m is a number such that:
1046 *
1047 * (m * x) & (2^N - 1) = floor(x/q).
1048 * for all x on Z where 0 <= x < 2^N
1049 *
1050 * Ignore the case where any of the above values equals zero; it is irrelevant
1051 * for our purposes (instanced arrays).
1052 *
1053 * Choose x = q. Then:
1054 *
1055 * (m * x) & (2^N - 1) = floor(x/q).
1056 * (m * q) & (2^N - 1) = floor(q/q).
1057 *
1058 * floor(q/q) = floor(1) = 1, therefore:
1059 *
1060 * (m * q) & (2^N - 1) = 1
1061 *
1062 * Recall the identity that the bitwise AND of one less than a power-of-two
1063 * equals the modulo with that power of two, i.e. for all x:
1064 *
1065 * x & (2^N - 1) = x % N
1066 *
1067 * Therefore:
1068 *
1069 * mq % (2^N) = 1
1070 *
1071 * By definition, a modular multiplicative inverse of a number m is the number
1072 * q such that with respect to a modulos M:
1073 *
1074 * mq % M = 1
1075 *
1076 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1077 *
1078 */
1079
1080 static void
1081 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1082 {
1083 #if 0
1084 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1085 * shift) the most lame way possible... just repeatedly add.
1086 * Asymptoptically slow but nobody cares in practice, unless you have
1087 * massive numbers of vertices or high divisors. */
1088
1089 unsigned inverse = 0;
1090
1091 /* Magic implicitly has the highest bit set */
1092 magic |= (1 << 31);
1093
1094 /* Depending on rounding direction */
1095 if (extra)
1096 magic++;
1097
1098 for (;;) {
1099 uint32_t product = magic * inverse;
1100
1101 if (shift) {
1102 product >>= shift;
1103 }
1104
1105 if (product == 1)
1106 break;
1107
1108 ++inverse;
1109 }
1110
1111 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1112
1113 /* Recall we're supposed to divide by (gl_level_divisor *
1114 * padded_num_vertices) */
1115
1116 unsigned padded_num_vertices = inverse / orig_divisor;
1117
1118 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1119 #endif
1120 }
1121
1122 static void
1123 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1124 mali_ptr addr, int job_no, char *suffix,
1125 int count, bool varying)
1126 {
1127 char *prefix = varying ? "varyings" : "attributes";
1128
1129 if (!addr) {
1130 pandecode_msg("no %s\n", prefix);
1131 return;
1132 }
1133
1134 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1135
1136 char base[128];
1137 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
1138
1139 for (int i = 0; i < count; ++i) {
1140 enum mali_attr_mode mode = attr[i].elements & 7;
1141
1142 if (mode == MALI_ATTR_UNUSED)
1143 continue;
1144
1145 mali_ptr raw_elements = attr[i].elements & ~7;
1146
1147 /* TODO: Do we maybe want to dump the attribute values
1148 * themselves given the specified format? Or is that too hard?
1149 * */
1150
1151 char *a = pointer_as_memory_reference(raw_elements);
1152 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
1153 free(a);
1154 }
1155
1156 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1157 pandecode_indent++;
1158
1159 for (int i = 0; i < count; ++i) {
1160 pandecode_log("{\n");
1161 pandecode_indent++;
1162
1163 unsigned mode = attr[i].elements & 7;
1164 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode));
1165 pandecode_prop("shift = %d", attr[i].shift);
1166 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1167 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1168 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1169
1170 /* Decode further where possible */
1171
1172 if (mode == MALI_ATTR_MODULO) {
1173 pandecode_padded_vertices(
1174 attr[i].shift,
1175 attr[i].extra_flags);
1176 }
1177
1178 pandecode_indent--;
1179 pandecode_log("}, \n");
1180
1181 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1182 i++;
1183 pandecode_log("{\n");
1184 pandecode_indent++;
1185 pandecode_prop("unk = 0x%x", attr[i].unk);
1186 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1187 if (attr[i].zero != 0)
1188 pandecode_prop("zero = 0x%x /* XXX zero tripped */", attr[i].zero);
1189 pandecode_prop("divisor = %d", attr[i].divisor);
1190 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1191 pandecode_indent--;
1192 pandecode_log("}, \n");
1193 }
1194
1195 }
1196
1197 pandecode_indent--;
1198 pandecode_log("};\n");
1199 }
1200
1201 static mali_ptr
1202 pandecode_shader_address(const char *name, mali_ptr ptr)
1203 {
1204 /* TODO: Decode flags */
1205 mali_ptr shader_ptr = ptr & ~15;
1206
1207 char *a = pointer_as_memory_reference(shader_ptr);
1208 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1209 free(a);
1210
1211 return shader_ptr;
1212 }
1213
1214 static bool
1215 all_zero(unsigned *buffer, unsigned count)
1216 {
1217 for (unsigned i = 0; i < count; ++i) {
1218 if (buffer[i])
1219 return false;
1220 }
1221
1222 return true;
1223 }
1224
1225 static void
1226 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1227 {
1228 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1229 return;
1230
1231 const char *func = pandecode_func(stencil->func);
1232 const char *sfail = pandecode_stencil_op(stencil->sfail);
1233 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1234 const char *dppass = pandecode_stencil_op(stencil->dppass);
1235
1236 if (stencil->zero)
1237 pandecode_msg("Stencil zero tripped: %X\n", stencil->zero);
1238
1239 pandecode_log(".stencil_%s = {\n", name);
1240 pandecode_indent++;
1241 pandecode_prop("ref = %d", stencil->ref);
1242 pandecode_prop("mask = 0x%02X", stencil->mask);
1243 pandecode_prop("func = %s", func);
1244 pandecode_prop("sfail = %s", sfail);
1245 pandecode_prop("dpfail = %s", dpfail);
1246 pandecode_prop("dppass = %s", dppass);
1247 pandecode_indent--;
1248 pandecode_log("},\n");
1249 }
1250
1251 static void
1252 pandecode_blend_equation(const struct mali_blend_equation *blend)
1253 {
1254 if (blend->zero1)
1255 pandecode_msg("Blend zero tripped: %X\n", blend->zero1);
1256
1257 pandecode_log(".equation = {\n");
1258 pandecode_indent++;
1259
1260 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1261 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1262
1263 pandecode_log(".color_mask = ");
1264 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1265 pandecode_log_cont(",\n");
1266
1267 pandecode_indent--;
1268 pandecode_log("},\n");
1269 }
1270
1271 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1272
1273 static unsigned
1274 decode_bifrost_constant(u16 constant)
1275 {
1276 float lo = (float) (constant & 0xFF);
1277 float hi = (float) (constant >> 8);
1278
1279 return (hi / 255.0) + (lo / 65535.0);
1280 }
1281
1282 static mali_ptr
1283 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1284 {
1285 struct bifrost_blend_rt *b =
1286 ((struct bifrost_blend_rt *) descs) + rt_no;
1287
1288 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1289 pandecode_indent++;
1290
1291 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1292 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1293 b->constant, decode_bifrost_constant(b->constant));
1294
1295 /* TODO figure out blend shader enable bit */
1296 pandecode_blend_equation(&b->equation);
1297 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1298 pandecode_prop("index = 0x%" PRIx16, b->index);
1299 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1300
1301 pandecode_indent--;
1302 pandecode_log("},\n");
1303
1304 return 0;
1305 }
1306
1307 static mali_ptr
1308 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1309 {
1310 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1311 return 0;
1312
1313 pandecode_log(".blend = {\n");
1314 pandecode_indent++;
1315
1316 if (is_shader) {
1317 pandecode_shader_address("shader", blend->shader);
1318 } else {
1319 pandecode_blend_equation(&blend->equation);
1320 pandecode_prop("constant = %f", blend->constant);
1321 }
1322
1323 pandecode_indent--;
1324 pandecode_log("},\n");
1325
1326 /* Return blend shader to disassemble if present */
1327 return is_shader ? (blend->shader & ~0xF) : 0;
1328 }
1329
1330 static mali_ptr
1331 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1332 {
1333 struct midgard_blend_rt *b =
1334 ((struct midgard_blend_rt *) descs) + rt_no;
1335
1336 /* Flags determine presence of blend shader */
1337 bool is_shader = (b->flags & 0xF) >= 0x2;
1338
1339 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1340 pandecode_indent++;
1341
1342 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1343
1344 union midgard_blend blend = b->blend;
1345 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1346
1347 pandecode_indent--;
1348 pandecode_log("};\n");
1349
1350 return shader;
1351 }
1352
1353 static int
1354 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1355 {
1356 char base[128];
1357 char *prefix = varying ? "varying" : "attribute";
1358 unsigned max_index = 0;
1359 snprintf(base, sizeof(base), "%s_meta", prefix);
1360
1361 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1362 pandecode_indent++;
1363
1364 struct mali_attr_meta *attr_meta;
1365 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1366
1367 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1368
1369 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1370 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1371 sizeof(*attr_mem));
1372
1373 pandecode_log("{\n");
1374 pandecode_indent++;
1375 pandecode_prop("index = %d", attr_meta->index);
1376
1377 if (attr_meta->index > max_index)
1378 max_index = attr_meta->index;
1379 pandecode_swizzle(attr_meta->swizzle);
1380 pandecode_prop("format = %s", pandecode_format(attr_meta->format));
1381
1382 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1383 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1384 pandecode_prop("src_offset = %d", attr_meta->src_offset);
1385 pandecode_indent--;
1386 pandecode_log("},\n");
1387
1388 }
1389
1390 pandecode_indent--;
1391 pandecode_log("};\n");
1392
1393 return count ? (max_index + 1) : 0;
1394 }
1395
1396 static void
1397 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1398 {
1399 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1400
1401 if (imem) {
1402 /* Indices are literally just a u32 array :) */
1403
1404 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1405
1406 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1407 pandecode_indent++;
1408
1409 for (unsigned i = 0; i < (index_count + 1); i += 3)
1410 pandecode_log("%d, %d, %d,\n",
1411 indices[i],
1412 indices[i + 1],
1413 indices[i + 2]);
1414
1415 pandecode_indent--;
1416 pandecode_log("};\n");
1417 }
1418 }
1419
1420 /* return bits [lo, hi) of word */
1421 static u32
1422 bits(u32 word, u32 lo, u32 hi)
1423 {
1424 if (hi - lo >= 32)
1425 return word; // avoid undefined behavior with the shift
1426
1427 return (word >> lo) & ((1 << (hi - lo)) - 1);
1428 }
1429
1430 static void
1431 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1432 {
1433 pandecode_log_cont("{\n");
1434 pandecode_indent++;
1435
1436 /* Decode invocation_count. See the comment before the definition of
1437 * invocation_count for an explanation.
1438 */
1439
1440 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1441 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1442 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1443
1444 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1445 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1446 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1447
1448 /* Even though we have this decoded, we want to ensure that the
1449 * representation is "unique" so we don't lose anything by printing only
1450 * the final result. More specifically, we need to check that we were
1451 * passed something in canonical form, since the definition per the
1452 * hardware is inherently not unique. How? Well, take the resulting
1453 * decode and pack it ourselves! If it is bit exact with what we
1454 * decoded, we're good to go. */
1455
1456 struct mali_vertex_tiler_prefix ref;
1457 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1458
1459 bool canonical =
1460 (p->invocation_count == ref.invocation_count) &&
1461 (p->size_y_shift == ref.size_y_shift) &&
1462 (p->size_z_shift == ref.size_z_shift) &&
1463 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1464 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1465 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1466 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1467
1468 if (!canonical) {
1469 pandecode_msg("XXX: non-canonical workgroups packing\n");
1470 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1471 ref.invocation_count,
1472 ref.size_y_shift,
1473 ref.size_z_shift,
1474 ref.workgroups_x_shift,
1475 ref.workgroups_y_shift,
1476 ref.workgroups_z_shift,
1477 ref.workgroups_x_shift_2);
1478
1479 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1480 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1481 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1482 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1483 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1484 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1485 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1486 }
1487
1488 /* Regardless, print the decode */
1489 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1490 size_x, size_y, size_z,
1491 groups_x, groups_y, groups_z);
1492
1493 /* TODO: Decode */
1494 if (p->unknown_draw)
1495 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1496
1497 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1498
1499 if (p->draw_mode != MALI_DRAW_NONE)
1500 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1501
1502 /* Index count only exists for tiler jobs anyway */
1503
1504 if (p->index_count)
1505 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1506
1507 if (p->offset_bias_correction)
1508 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1509
1510 if (p->zero1) {
1511 pandecode_msg("Zero tripped\n");
1512 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1513 }
1514
1515 pandecode_indent--;
1516 pandecode_log("},\n");
1517 }
1518
1519 static void
1520 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1521 {
1522 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1523
1524 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1525
1526 for (int i = 0; i < ubufs_count; i++) {
1527 mali_ptr ptr = ubufs[i].ptr << 2;
1528 struct pandecode_mapped_memory *umem2 = pandecode_find_mapped_gpu_mem_containing(ptr);
1529 uint32_t *PANDECODE_PTR_VAR(ubuf, umem2, ptr);
1530 char name[50];
1531 snprintf(name, sizeof(name), "ubuf_%d", i);
1532 /* The blob uses ubuf 0 to upload internal stuff and
1533 * uniforms that won't fit/are accessed indirectly, so
1534 * it puts it in the batchbuffer.
1535 */
1536 pandecode_log("uint32_t %s_%d[] = {\n", name, job_no);
1537 pandecode_indent++;
1538
1539 for (int j = 0; j <= ubufs[i].size; j++) {
1540 for (int k = 0; k < 4; k++) {
1541 if (k == 0)
1542 pandecode_log("0x%"PRIx32", ", ubuf[4 * j + k]);
1543 else
1544 pandecode_log_cont("0x%"PRIx32", ", ubuf[4 * j + k]);
1545
1546 }
1547
1548 pandecode_log_cont("\n");
1549 }
1550
1551 pandecode_indent--;
1552 pandecode_log("};\n");
1553 }
1554
1555 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1556 pubufs, job_no);
1557 pandecode_indent++;
1558
1559 for (int i = 0; i < ubufs_count; i++) {
1560 pandecode_log("{\n");
1561 pandecode_indent++;
1562 pandecode_prop("size = MALI_POSITIVE(%d)", ubufs[i].size + 1);
1563 pandecode_prop("ptr = ubuf_%d_%d_p >> 2", i, job_no);
1564 pandecode_indent--;
1565 pandecode_log("},\n");
1566 }
1567
1568 pandecode_indent--;
1569 pandecode_log("};\n");
1570 }
1571
1572 static void
1573 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1574 {
1575
1576 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1577
1578 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1579
1580 if (scratchpad->zero)
1581 pandecode_msg("XXX scratchpad zero tripped");
1582
1583 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1584 pandecode_indent++;
1585
1586 pandecode_prop("flags = 0x%x", scratchpad->flags);
1587 MEMORY_PROP(scratchpad, gpu_scratchpad);
1588
1589 pandecode_indent--;
1590 pandecode_log("};\n");
1591 }
1592
1593 static unsigned shader_id = 0;
1594
1595 static void
1596 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1597 bool is_bifrost, unsigned nr_regs)
1598 {
1599 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1600 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1601
1602 /* Compute maximum possible size */
1603 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1604
1605 /* Print some boilerplate to clearly denote the assembly (which doesn't
1606 * obey indentation rules), and actually do the disassembly! */
1607
1608 printf("\n\n");
1609
1610 char prefix[512];
1611
1612 snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ",
1613 shader_id++,
1614 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX");
1615
1616 if (is_bifrost) {
1617 disassemble_bifrost(code, sz, false);
1618 } else {
1619 disassemble_midgard(code, sz, true, nr_regs, prefix);
1620 }
1621
1622 printf("\n\n");
1623 }
1624
1625 static void
1626 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1627 int job_no, enum mali_job_type job_type,
1628 char *suffix, bool is_bifrost)
1629 {
1630 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1631 struct pandecode_mapped_memory *attr_mem;
1632
1633 unsigned rt_count = 1;
1634
1635 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1636 * are the only things actually needed from the FBD, vertex/tiler jobs
1637 * no longer reference the FBD -- instead, this field points to some
1638 * info about the scratchpad.
1639 */
1640 if (is_bifrost)
1641 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1642 else if (p->framebuffer & MALI_MFBD)
1643 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1644 else if (job_type == JOB_TYPE_COMPUTE)
1645 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1646 else
1647 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1648
1649 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1650 int texture_count = 0, sampler_count = 0;
1651
1652 if (shader_meta_ptr) {
1653 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1654 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1655
1656 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1657 pandecode_indent++;
1658
1659 /* Save for dumps */
1660 attribute_count = s->attribute_count;
1661 varying_count = s->varying_count;
1662 texture_count = s->texture_count;
1663 sampler_count = s->sampler_count;
1664
1665 if (is_bifrost) {
1666 uniform_count = s->bifrost2.uniform_count;
1667 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1668 } else {
1669 uniform_count = s->midgard1.uniform_count;
1670 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1671 }
1672
1673 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1674
1675 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1676 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1677 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1678 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1679
1680 unsigned nr_registers = 0;
1681
1682 if (is_bifrost) {
1683 pandecode_log(".bifrost1 = {\n");
1684 pandecode_indent++;
1685
1686 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1687 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1688
1689 pandecode_indent--;
1690 pandecode_log("},\n");
1691 } else {
1692 pandecode_log(".midgard1 = {\n");
1693 pandecode_indent++;
1694
1695 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1696 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1697 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1698 nr_registers = s->midgard1.work_count;
1699
1700 pandecode_log(".flags = ");
1701 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1702 pandecode_log_cont(",\n");
1703
1704 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1705
1706 pandecode_indent--;
1707 pandecode_log("},\n");
1708 }
1709
1710 if (s->depth_units || s->depth_factor) {
1711 pandecode_prop("depth_factor = %f", s->depth_factor);
1712 pandecode_prop("depth_units = %f", s->depth_units);
1713 }
1714
1715 if (s->alpha_coverage) {
1716 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1717 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1718
1719 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1720 invert_alpha_coverage ? "~" : "",
1721 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1722 }
1723
1724 if (s->unknown2_3 || s->unknown2_4) {
1725 pandecode_log(".unknown2_3 = ");
1726
1727 int unknown2_3 = s->unknown2_3;
1728 int unknown2_4 = s->unknown2_4;
1729
1730 /* We're not quite sure what these flags mean without the depth test, if anything */
1731
1732 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1733 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1734 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1735
1736 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1737 }
1738
1739 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1740 pandecode_log_cont(",\n");
1741
1742 pandecode_log(".unknown2_4 = ");
1743 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1744 pandecode_log_cont(",\n");
1745 }
1746
1747 if (s->stencil_mask_front || s->stencil_mask_back) {
1748 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1749 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1750 }
1751
1752 pandecode_stencil("front", &s->stencil_front);
1753 pandecode_stencil("back", &s->stencil_back);
1754
1755 if (is_bifrost) {
1756 pandecode_log(".bifrost2 = {\n");
1757 pandecode_indent++;
1758
1759 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1760 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1761 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1762 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1763
1764 pandecode_indent--;
1765 pandecode_log("},\n");
1766 } else if (s->midgard2.unknown2_7) {
1767 pandecode_log(".midgard2 = {\n");
1768 pandecode_indent++;
1769
1770 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1771 pandecode_indent--;
1772 pandecode_log("},\n");
1773 }
1774
1775 if (s->unknown2_8)
1776 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1777
1778 if (!is_bifrost) {
1779 /* TODO: Blend shaders routing/disasm */
1780
1781 union midgard_blend blend = s->blend;
1782 pandecode_midgard_blend(&blend, false);
1783 }
1784
1785 pandecode_indent--;
1786 pandecode_log("};\n");
1787
1788 /* MRT blend fields are used whenever MFBD is used, with
1789 * per-RT descriptors */
1790
1791 if (job_type == JOB_TYPE_TILER) {
1792 void* blend_base = (void *) (s + 1);
1793
1794 for (unsigned i = 0; i < rt_count; i++) {
1795 mali_ptr shader = 0;
1796
1797 if (is_bifrost)
1798 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1799 else
1800 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1801
1802 if (shader & ~0xF)
1803 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
1804 }
1805 }
1806
1807 if (shader_ptr & ~0xF)
1808 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
1809 } else
1810 pandecode_msg("<no shader>\n");
1811
1812 if (p->viewport) {
1813 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1814 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1815
1816 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
1817 pandecode_indent++;
1818
1819 pandecode_prop("clip_minx = %f", f->clip_minx);
1820 pandecode_prop("clip_miny = %f", f->clip_miny);
1821 pandecode_prop("clip_minz = %f", f->clip_minz);
1822 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1823 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1824 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1825
1826 /* Only the higher coordinates are MALI_POSITIVE scaled */
1827
1828 pandecode_prop("viewport0 = { %d, %d }",
1829 f->viewport0[0], f->viewport0[1]);
1830
1831 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1832 f->viewport1[0] + 1, f->viewport1[1] + 1);
1833
1834 pandecode_indent--;
1835 pandecode_log("};\n");
1836 }
1837
1838 if (p->attribute_meta) {
1839 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1840
1841 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1842 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
1843 }
1844
1845 /* Varyings are encoded like attributes but not actually sent; we just
1846 * pass a zero buffer with the right stride/size set, (or whatever)
1847 * since the GPU will write to it itself */
1848
1849 if (p->varying_meta) {
1850 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1851 }
1852
1853 if (p->varyings) {
1854 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1855
1856 /* Number of descriptors depends on whether there are
1857 * non-internal varyings */
1858
1859 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
1860 }
1861
1862 bool is_compute = job_type == JOB_TYPE_COMPUTE;
1863
1864 if (p->uniforms && !is_compute) {
1865 int rows = uniform_count, width = 4;
1866 size_t sz = rows * width * sizeof(float);
1867
1868 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1869 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1870 u32 *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1871
1872 pandecode_log("u32 uniforms_%d%s[] = {\n", job_no, suffix);
1873
1874 pandecode_indent++;
1875
1876 for (int row = 0; row < rows; row++) {
1877 for (int i = 0; i < width; i++) {
1878 u32 v = uniforms[i];
1879 float f;
1880 memcpy(&f, &v, sizeof(v));
1881 pandecode_log_cont("%X /* %f */, ", v, f);
1882 }
1883
1884 pandecode_log_cont("\n");
1885
1886 uniforms += width;
1887 }
1888
1889 pandecode_indent--;
1890 pandecode_log("};\n");
1891 } else if (p->uniforms) {
1892 int rows = uniform_count * 2;
1893 size_t sz = rows * sizeof(mali_ptr);
1894
1895 struct pandecode_mapped_memory *uniform_mem = pandecode_find_mapped_gpu_mem_containing(p->uniforms);
1896 pandecode_fetch_gpu_mem(uniform_mem, p->uniforms, sz);
1897 mali_ptr *PANDECODE_PTR_VAR(uniforms, uniform_mem, p->uniforms);
1898
1899 pandecode_log("mali_ptr uniforms_%d%s[] = {\n", job_no, suffix);
1900
1901 pandecode_indent++;
1902
1903 for (int row = 0; row < rows; row++) {
1904 char *a = pointer_as_memory_reference(uniforms[row]);
1905 pandecode_log("%s,\n", a);
1906 free(a);
1907 }
1908
1909 pandecode_indent--;
1910 pandecode_log("};\n");
1911
1912 }
1913
1914 if (p->uniform_buffers) {
1915 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1916 }
1917
1918 if (p->texture_trampoline) {
1919 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
1920
1921 if (mmem) {
1922 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
1923
1924 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
1925 pandecode_indent++;
1926
1927 for (int tex = 0; tex < texture_count; ++tex) {
1928 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1929 char *a = pointer_as_memory_reference(*u);
1930 pandecode_log("%s,\n", a);
1931 free(a);
1932 }
1933
1934 pandecode_indent--;
1935 pandecode_log("};\n");
1936
1937 /* Now, finally, descend down into the texture descriptor */
1938 for (int tex = 0; tex < texture_count; ++tex) {
1939 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
1940 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1941
1942 if (tmem) {
1943 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
1944
1945 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
1946 pandecode_indent++;
1947
1948 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
1949 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
1950 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
1951 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
1952 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
1953 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
1954 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
1955
1956 struct mali_texture_format f = t->format;
1957
1958 pandecode_log(".format = {\n");
1959 pandecode_indent++;
1960
1961 pandecode_swizzle(f.swizzle);
1962 pandecode_prop("format = %s", pandecode_format(f.format));
1963 pandecode_prop("type = %s", pandecode_texture_type(f.type));
1964 pandecode_prop("srgb = %" PRId32, f.srgb);
1965 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
1966 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
1967
1968 pandecode_indent--;
1969 pandecode_log("},\n");
1970
1971 pandecode_swizzle(t->swizzle);
1972
1973 if (t->swizzle_zero) {
1974 /* Shouldn't happen */
1975 pandecode_msg("Swizzle zero tripped but replay will be fine anyway");
1976 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
1977 }
1978
1979 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
1980
1981 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
1982 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
1983 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
1984
1985 pandecode_log(".payload = {\n");
1986 pandecode_indent++;
1987
1988 /* A bunch of bitmap pointers follow.
1989 * We work out the correct number,
1990 * based on the mipmap/cubemap
1991 * properties, but dump extra
1992 * possibilities to futureproof */
1993
1994 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
1995 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
1996
1997 /* Miptree for each face */
1998 if (f.type == MALI_TEX_CUBE)
1999 bitmap_count *= 6;
2000
2001 /* Array of textures */
2002 bitmap_count *= MALI_NEGATIVE(t->array_size);
2003
2004 /* Stride for each element */
2005 if (manual_stride)
2006 bitmap_count *= 2;
2007
2008 /* Sanity check the size */
2009 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2010 assert (bitmap_count <= max_count);
2011
2012 for (int i = 0; i < bitmap_count; ++i) {
2013 /* How we dump depends if this is a stride or a pointer */
2014
2015 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
2016 /* signed 32-bit snuck in as a 64-bit pointer */
2017 uint64_t stride_set = t->payload[i];
2018 uint32_t clamped_stride = stride_set;
2019 int32_t stride = clamped_stride;
2020 assert(stride_set == clamped_stride);
2021 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2022 } else {
2023 char *a = pointer_as_memory_reference(t->payload[i]);
2024 pandecode_log("%s, \n", a);
2025 free(a);
2026 }
2027 }
2028
2029 pandecode_indent--;
2030 pandecode_log("},\n");
2031
2032 pandecode_indent--;
2033 pandecode_log("};\n");
2034 }
2035 }
2036 }
2037 }
2038
2039 if (p->sampler_descriptor) {
2040 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2041
2042 if (smem) {
2043 struct mali_sampler_descriptor *s;
2044
2045 mali_ptr d = p->sampler_descriptor;
2046
2047 for (int i = 0; i < sampler_count; ++i) {
2048 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2049
2050 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2051 pandecode_indent++;
2052
2053 pandecode_log(".filter_mode = ");
2054 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2055 pandecode_log_cont(",\n");
2056
2057 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2058 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2059
2060 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2061 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2062 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2063
2064 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2065
2066 if (s->zero || s->zero2) {
2067 pandecode_msg("Zero tripped\n");
2068 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2069 }
2070
2071 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2072
2073 pandecode_prop("border_color = { %f, %f, %f, %f }",
2074 s->border_color[0],
2075 s->border_color[1],
2076 s->border_color[2],
2077 s->border_color[3]);
2078
2079 pandecode_indent--;
2080 pandecode_log("};\n");
2081 }
2082 }
2083 }
2084 }
2085
2086 static void
2087 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2088 {
2089 if (!(p->position_varying || p->occlusion_counter || p->flags))
2090 return;
2091
2092 pandecode_log(".postfix = {\n");
2093 pandecode_indent++;
2094
2095 MEMORY_PROP(p, position_varying);
2096 MEMORY_PROP(p, occlusion_counter);
2097
2098 if (p->flags)
2099 pandecode_prop("flags = %d", p->flags);
2100
2101 pandecode_indent--;
2102 pandecode_log("},\n");
2103 }
2104
2105 static void
2106 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2107 {
2108 pandecode_log_cont("{\n");
2109 pandecode_indent++;
2110
2111 pandecode_prop("unk2 = 0x%x", v->unk2);
2112
2113 if (v->zero0 || v->zero1) {
2114 pandecode_msg("vertex only zero tripped");
2115 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2116 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2117 }
2118
2119 pandecode_indent--;
2120 pandecode_log("}\n");
2121 }
2122
2123 static void
2124 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2125 {
2126
2127 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2128 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2129
2130 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2131 pandecode_indent++;
2132
2133 if (h->zero) {
2134 pandecode_msg("tiler heap zero tripped\n");
2135 pandecode_prop("zero = 0x%x", h->zero);
2136 }
2137
2138 for (int i = 0; i < 12; i++) {
2139 if (h->zeros[i] != 0) {
2140 pandecode_msg("tiler heap zero %d tripped, value %x\n",
2141 i, h->zeros[i]);
2142 }
2143 }
2144
2145 pandecode_prop("heap_size = 0x%x", h->heap_size);
2146 MEMORY_PROP(h, tiler_heap_start);
2147 MEMORY_PROP(h, tiler_heap_free);
2148
2149 /* this might point to the beginning of another buffer, when it's
2150 * really the end of the tiler heap buffer, so we have to be careful
2151 * here. but for zero length, we need the same pointer.
2152 */
2153
2154 if (h->tiler_heap_end == h->tiler_heap_start) {
2155 MEMORY_PROP(h, tiler_heap_start);
2156 } else {
2157 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2158 pandecode_prop("tiler_heap_end = %s + 1", a);
2159 free(a);
2160 }
2161
2162 pandecode_indent--;
2163 pandecode_log("};\n");
2164 }
2165
2166 static void
2167 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2168 {
2169 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2170 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2171
2172 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2173
2174 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2175 pandecode_indent++;
2176
2177 if (t->zero0 || t->zero1) {
2178 pandecode_msg("tiler meta zero tripped");
2179 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2180 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2181 }
2182
2183 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2184 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2185
2186 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2187 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2188
2189 for (int i = 0; i < 12; i++) {
2190 if (t->zeros[i] != 0) {
2191 pandecode_msg("tiler heap zero %d tripped, value %" PRIx64 "\n",
2192 i, t->zeros[i]);
2193 }
2194 }
2195
2196 pandecode_indent--;
2197 pandecode_log("};\n");
2198 }
2199
2200 static void
2201 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2202 {
2203 pandecode_log(".gl_enables = ");
2204
2205 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2206
2207 pandecode_log_cont(",\n");
2208 }
2209
2210 static void
2211 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2212 {
2213 if (u.pointer == 0x0)
2214 return;
2215
2216 pandecode_log(".primitive_size = {\n");
2217 pandecode_indent++;
2218
2219 if (constant) {
2220 pandecode_prop("constant = %f", u.constant);
2221 } else {
2222 MEMORY_PROP((&u), pointer);
2223 }
2224
2225 pandecode_indent--;
2226 pandecode_log("},\n");
2227 }
2228
2229 static void
2230 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2231 {
2232 pandecode_log_cont("{\n");
2233 pandecode_indent++;
2234
2235 /* TODO: gl_PointSize on Bifrost */
2236 pandecode_primitive_size(t->primitive_size, true);
2237
2238 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2239
2240 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2241 || t->zero6 || t->zero7 || t->zero8) {
2242 pandecode_msg("tiler only zero tripped");
2243 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2244 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2245 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2246 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2247 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2248 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2249 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2250 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2251 }
2252
2253 pandecode_indent--;
2254 pandecode_log("},\n");
2255 }
2256
2257 static int
2258 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2259 const struct pandecode_mapped_memory *mem,
2260 mali_ptr payload, int job_no)
2261 {
2262 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2263
2264 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2265
2266 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2267 pandecode_indent++;
2268
2269 pandecode_log(".prefix = ");
2270 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2271
2272 pandecode_log(".vertex = ");
2273 pandecode_vertex_only_bfr(&v->vertex);
2274
2275 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2276
2277 pandecode_indent--;
2278 pandecode_log("};\n");
2279
2280 return sizeof(*v);
2281 }
2282
2283 static int
2284 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2285 const struct pandecode_mapped_memory *mem,
2286 mali_ptr payload, int job_no)
2287 {
2288 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2289
2290 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2291
2292 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2293 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2294
2295 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2296 pandecode_indent++;
2297
2298 pandecode_log(".prefix = ");
2299 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2300
2301 pandecode_log(".tiler = ");
2302 pandecode_tiler_only_bfr(&t->tiler, job_no);
2303
2304 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2305
2306 pandecode_indent--;
2307 pandecode_log("};\n");
2308
2309 return sizeof(*t);
2310 }
2311
2312 static int
2313 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2314 const struct pandecode_mapped_memory *mem,
2315 mali_ptr payload, int job_no)
2316 {
2317 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2318
2319 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2320
2321 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2322
2323 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2324 pandecode_indent++;
2325
2326 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2327 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2328
2329 bool instanced = v->instance_shift || v->instance_odd;
2330 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2331
2332 pandecode_log(".prefix = ");
2333 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2334
2335 pandecode_gl_enables(v->gl_enables, h->job_type);
2336
2337 if (v->instance_shift || v->instance_odd) {
2338 pandecode_prop("instance_shift = 0x%d /* %d */",
2339 v->instance_shift, 1 << v->instance_shift);
2340 pandecode_prop("instance_odd = 0x%X /* %d */",
2341 v->instance_odd, (2 * v->instance_odd) + 1);
2342
2343 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2344 }
2345
2346 if (v->offset_start)
2347 pandecode_prop("offset_start = %d", v->offset_start);
2348
2349 if (v->zero5) {
2350 pandecode_msg("Zero tripped\n");
2351 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2352 }
2353
2354 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2355
2356 pandecode_indent--;
2357 pandecode_log("};\n");
2358
2359 return sizeof(*v);
2360 }
2361
2362 static int
2363 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2364 mali_ptr payload, int job_no,
2365 bool is_bifrost)
2366 {
2367 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2368
2369 bool fbd_dumped = false;
2370
2371 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2372 /* Only SFBDs are understood, not MFBDs. We're speculating,
2373 * based on the versioning, kernel code, etc, that the
2374 * difference is between Single FrameBuffer Descriptor and
2375 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2376 * up with multi-framebuffer support being added (T7xx onwards,
2377 * including Gxx). In any event, there's some field shuffling
2378 * that we haven't looked into yet. */
2379
2380 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2381 fbd_dumped = true;
2382 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2383 /* We don't know if Bifrost supports SFBD's at all, since the
2384 * driver never uses them. And the format is different from
2385 * Midgard anyways, due to the tiler heap and scratchpad being
2386 * moved out into separate structures, so it's not clear what a
2387 * Bifrost SFBD would even look like without getting an actual
2388 * trace, which appears impossible.
2389 */
2390
2391 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2392 fbd_dumped = true;
2393 }
2394
2395 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2396 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2397 pandecode_indent++;
2398
2399 /* See the comments by the macro definitions for mathematical context
2400 * on why this is so weird */
2401
2402 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2403 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2404
2405 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2406 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2407 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2408
2409 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2410 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2411 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2412
2413 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2414 * we have to fallback on offsets. */
2415
2416 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2417
2418 /* TODO: Decode */
2419 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2420
2421 if (fbd_dumped)
2422 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2423 fbd_type, extra_flags);
2424 else
2425 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p),
2426 fbd_type, extra_flags);
2427
2428 pandecode_indent--;
2429 pandecode_log("};\n");
2430
2431 return sizeof(*s);
2432 }
2433
2434 static int job_descriptor_number = 0;
2435
2436 int
2437 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2438 {
2439 struct mali_job_descriptor_header *h;
2440
2441 int start_number = 0;
2442
2443 bool first = true;
2444 bool last_size;
2445
2446 do {
2447 struct pandecode_mapped_memory *mem =
2448 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2449
2450 void *payload;
2451
2452 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2453
2454 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2455 * high 32-bits of the 64-bit pointer are reused to store
2456 * something else.
2457 */
2458 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2459 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2460 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2461
2462 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2463 MALI_PAYLOAD_SIZE);
2464
2465 int job_no = job_descriptor_number++;
2466
2467 if (first)
2468 start_number = job_no;
2469
2470 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2471 pandecode_indent++;
2472
2473 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2474
2475 /* Save for next job fixing */
2476 last_size = h->job_descriptor_size;
2477
2478 if (h->job_descriptor_size)
2479 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2480
2481 if (h->exception_status && h->exception_status != 0x1)
2482 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2483 h->exception_status,
2484 (h->exception_status >> 16) & 0xFFFF,
2485 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2486 h->exception_status & 0xFF);
2487
2488 if (h->first_incomplete_task)
2489 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2490
2491 if (h->fault_pointer)
2492 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2493
2494 if (h->job_barrier)
2495 pandecode_prop("job_barrier = %d", h->job_barrier);
2496
2497 pandecode_prop("job_index = %d", h->job_index);
2498
2499 if (h->unknown_flags)
2500 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2501
2502 if (h->job_dependency_index_1)
2503 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2504
2505 if (h->job_dependency_index_2)
2506 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2507
2508 pandecode_indent--;
2509 pandecode_log("};\n");
2510
2511 /* Do not touch the field yet -- decode the payload first, and
2512 * don't touch that either. This is essential for the uploads
2513 * to occur in sequence and therefore be dynamically allocated
2514 * correctly. Do note the size, however, for that related
2515 * reason. */
2516
2517 switch (h->job_type) {
2518 case JOB_TYPE_SET_VALUE: {
2519 struct mali_payload_set_value *s = payload;
2520 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2521 pandecode_indent++;
2522 MEMORY_PROP(s, out);
2523 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2524 pandecode_indent--;
2525 pandecode_log("};\n");
2526
2527 break;
2528 }
2529
2530 case JOB_TYPE_TILER:
2531 case JOB_TYPE_VERTEX:
2532 case JOB_TYPE_COMPUTE:
2533 if (bifrost) {
2534 if (h->job_type == JOB_TYPE_TILER)
2535 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2536 else
2537 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2538 } else
2539 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2540
2541 break;
2542
2543 case JOB_TYPE_FRAGMENT:
2544 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2545 break;
2546
2547 default:
2548 break;
2549 }
2550
2551 /* Handle linkage */
2552
2553 if (!first) {
2554 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2555
2556 if (last_size)
2557 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2558 else
2559 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2560 }
2561
2562 first = false;
2563
2564 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2565
2566 return start_number;
2567 }