projects / mesa.git / blob
? search:


95f3d25fc01ef626a2832e9161d27344e009e093
[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 pandecode_msg("XXX: invalid draw mode %X\n", mode);
344 return "";
345 }
346
347 #undef DEFINE_CASE
348 }
349
350 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
351 static char *
352 pandecode_func(enum mali_func mode)
353 {
354 switch (mode) {
355 DEFINE_CASE(NEVER);
356 DEFINE_CASE(LESS);
357 DEFINE_CASE(EQUAL);
358 DEFINE_CASE(LEQUAL);
359 DEFINE_CASE(GREATER);
360 DEFINE_CASE(NOTEQUAL);
361 DEFINE_CASE(GEQUAL);
362 DEFINE_CASE(ALWAYS);
363
364 default:
365 pandecode_msg("XXX: invalid func %X\n", mode);
366 return "";
367 }
368 }
369 #undef DEFINE_CASE
370
371 /* Why is this duplicated? Who knows... */
372 #define DEFINE_CASE(name) case MALI_ALT_FUNC_ ## name: return "MALI_ALT_FUNC_" #name
373 static char *
374 pandecode_alt_func(enum mali_alt_func mode)
375 {
376 switch (mode) {
377 DEFINE_CASE(NEVER);
378 DEFINE_CASE(LESS);
379 DEFINE_CASE(EQUAL);
380 DEFINE_CASE(LEQUAL);
381 DEFINE_CASE(GREATER);
382 DEFINE_CASE(NOTEQUAL);
383 DEFINE_CASE(GEQUAL);
384 DEFINE_CASE(ALWAYS);
385
386 default:
387 pandecode_msg("XXX: invalid alt func %X\n", mode);
388 return "";
389 }
390 }
391 #undef DEFINE_CASE
392
393 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
394 static char *
395 pandecode_stencil_op(enum mali_stencil_op op)
396 {
397 switch (op) {
398 DEFINE_CASE(KEEP);
399 DEFINE_CASE(REPLACE);
400 DEFINE_CASE(ZERO);
401 DEFINE_CASE(INVERT);
402 DEFINE_CASE(INCR_WRAP);
403 DEFINE_CASE(DECR_WRAP);
404 DEFINE_CASE(INCR);
405 DEFINE_CASE(DECR);
406
407 default:
408 pandecode_msg("XXX: invalid stencil op %X\n", op);
409 return "";
410 }
411 }
412
413 #undef DEFINE_CASE
414
415 #define DEFINE_CASE(name) case MALI_ATTR_ ## name: return "MALI_ATTR_" #name
416 static char *pandecode_attr_mode(enum mali_attr_mode mode)
417 {
418 switch(mode) {
419 DEFINE_CASE(UNUSED);
420 DEFINE_CASE(LINEAR);
421 DEFINE_CASE(POT_DIVIDE);
422 DEFINE_CASE(MODULO);
423 DEFINE_CASE(NPOT_DIVIDE);
424 DEFINE_CASE(IMAGE);
425 DEFINE_CASE(INTERNAL);
426 default:
427 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
428 return "";
429 }
430 }
431
432 #undef DEFINE_CASE
433
434 #define DEFINE_CASE(name) case MALI_CHANNEL_## name: return "MALI_CHANNEL_" #name
435 static char *
436 pandecode_channel(enum mali_channel channel)
437 {
438 switch (channel) {
439 DEFINE_CASE(RED);
440 DEFINE_CASE(GREEN);
441 DEFINE_CASE(BLUE);
442 DEFINE_CASE(ALPHA);
443 DEFINE_CASE(ZERO);
444 DEFINE_CASE(ONE);
445
446 default:
447 pandecode_msg("XXX: invalid channel %X\n", channel);
448 return "";
449 }
450 }
451 #undef DEFINE_CASE
452
453 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
454 static char *
455 pandecode_wrap_mode(enum mali_wrap_mode op)
456 {
457 switch (op) {
458 DEFINE_CASE(REPEAT);
459 DEFINE_CASE(CLAMP_TO_EDGE);
460 DEFINE_CASE(CLAMP_TO_BORDER);
461 DEFINE_CASE(MIRRORED_REPEAT);
462
463 default:
464 pandecode_msg("XXX: invalid wrap mode %X\n", op);
465 return "";
466 }
467 }
468 #undef DEFINE_CASE
469
470 #define DEFINE_CASE(name) case MALI_TEX_## name: return "MALI_TEX_" #name
471 static char *
472 pandecode_texture_type(enum mali_texture_type type)
473 {
474 switch (type) {
475 DEFINE_CASE(1D);
476 DEFINE_CASE(2D);
477 DEFINE_CASE(3D);
478 DEFINE_CASE(CUBE);
479
480 default:
481 unreachable("Unknown case");
482 }
483 }
484 #undef DEFINE_CASE
485
486 #define DEFINE_CASE(name) case MALI_MFBD_BLOCK_## name: return "MALI_MFBD_BLOCK_" #name
487 static char *
488 pandecode_mfbd_block_format(enum mali_mfbd_block_format fmt)
489 {
490 switch (fmt) {
491 DEFINE_CASE(TILED);
492 DEFINE_CASE(UNKNOWN);
493 DEFINE_CASE(LINEAR);
494 DEFINE_CASE(AFBC);
495
496 default:
497 unreachable("Invalid case");
498 }
499 }
500 #undef DEFINE_CASE
501
502 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
503 static char *
504 pandecode_exception_access(enum mali_exception_access access)
505 {
506 switch (access) {
507 DEFINE_CASE(NONE);
508 DEFINE_CASE(EXECUTE);
509 DEFINE_CASE(READ);
510 DEFINE_CASE(WRITE);
511
512 default:
513 unreachable("Invalid case");
514 }
515 }
516 #undef DEFINE_CASE
517
518 /* Midgard's tiler descriptor is embedded within the
519 * larger FBD */
520
521 static void
522 pandecode_midgard_tiler_descriptor(
523 const struct midgard_tiler_descriptor *t,
524 unsigned width,
525 unsigned height,
526 bool is_fragment)
527 {
528 pandecode_log(".tiler = {\n");
529 pandecode_indent++;
530
531 if (t->hierarchy_mask == MALI_TILER_DISABLED)
532 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
533 else
534 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
535
536 /* We know this name from the kernel, but we never see it nonzero */
537
538 if (t->flags)
539 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
540
541 MEMORY_PROP(t, polygon_list);
542
543 /* The body is offset from the base of the polygon list */
544 assert(t->polygon_list_body > t->polygon_list);
545 unsigned body_offset = t->polygon_list_body - t->polygon_list;
546
547 /* It needs to fit inside the reported size */
548 assert(t->polygon_list_size >= body_offset);
549
550 /* Check that we fit */
551 struct pandecode_mapped_memory *plist =
552 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
553
554 assert(t->polygon_list_size <= plist->length);
555
556 /* Now that we've sanity checked, we'll try to calculate the sizes
557 * ourselves for comparison */
558
559 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask);
560 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask);
561
562 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
563 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
564 ref_header, ref_size);
565 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
566 pandecode_msg("body offset %d\n", body_offset);
567 }
568
569 /* The tiler heap has a start and end specified -- it should be
570 * identical to what we have in the BO. The exception is if tiling is
571 * disabled. */
572
573 MEMORY_PROP(t, heap_start);
574 assert(t->heap_end >= t->heap_start);
575
576 struct pandecode_mapped_memory *heap =
577 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
578
579 unsigned heap_size = t->heap_end - t->heap_start;
580
581 /* Tiling is enabled with a special flag */
582 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
583 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
584
585 bool tiling_enabled = hierarchy_mask;
586
587 if (tiling_enabled) {
588 /* When tiling is enabled, the heap should be a tight fit */
589 unsigned heap_offset = t->heap_start - heap->gpu_va;
590 if ((heap_offset + heap_size) != heap->length) {
591 pandecode_msg("XXX: heap size %d (expected %d)\n",
592 heap_size, heap->length - heap_offset);
593 }
594
595 /* We should also have no other flags */
596 if (tiler_flags)
597 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
598 } else {
599 /* When tiling is disabled, we should have that flag and no others */
600
601 if (tiler_flags != MALI_TILER_DISABLED) {
602 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
603 tiler_flags);
604 }
605
606 /* We should also have an empty heap */
607 if (heap_size) {
608 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
609 heap_size);
610 }
611
612 /* Disabled tiling is used only for clear-only jobs, which are
613 * purely FRAGMENT, so we should never see this for
614 * non-FRAGMENT descriptors. */
615
616 if (!is_fragment)
617 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
618 }
619
620 /* We've never seen weights used in practice, but we know from the
621 * kernel these fields is there */
622
623 bool nonzero_weights = false;
624
625 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
626 nonzero_weights |= t->weights[w] != 0x0;
627 }
628
629 if (nonzero_weights) {
630 pandecode_log(".weights = {");
631
632 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
633 pandecode_log("%d, ", t->weights[w]);
634 }
635
636 pandecode_log("},");
637 }
638
639 pandecode_indent--;
640 pandecode_log("}\n");
641 }
642
643 static void
644 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment)
645 {
646 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
647 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
648
649 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
650 pandecode_indent++;
651
652 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
653 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
654
655 pandecode_log(".format = ");
656 pandecode_log_decoded_flags(fb_fmt_flag_info, s->format);
657 pandecode_log_cont(",\n");
658
659 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", s->width + 1);
660 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", s->height + 1);
661
662 MEMORY_PROP(s, framebuffer);
663 pandecode_prop("stride = %d", s->stride);
664
665 /* Earlier in the actual commandstream -- right before width -- but we
666 * delay to flow nicer */
667
668 pandecode_log(".clear_flags = ");
669 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
670 pandecode_log_cont(",\n");
671
672 if (s->depth_buffer | s->depth_buffer_enable) {
673 MEMORY_PROP(s, depth_buffer);
674 pandecode_prop("depth_buffer_enable = %s", DS_ENABLE(s->depth_buffer_enable));
675 }
676
677 if (s->stencil_buffer | s->stencil_buffer_enable) {
678 MEMORY_PROP(s, stencil_buffer);
679 pandecode_prop("stencil_buffer_enable = %s", DS_ENABLE(s->stencil_buffer_enable));
680 }
681
682 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
683 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
684 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
685 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
686 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
687 }
688
689 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
690 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
691 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
692 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
693 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
694 }
695
696 if (s->clear_stencil) {
697 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
698 }
699
700 MEMORY_PROP(s, unknown_address_0);
701 const struct midgard_tiler_descriptor t = s->tiler;
702 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment);
703
704 pandecode_indent--;
705 pandecode_log("};\n");
706
707 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
708 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
709 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
710 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
711
712 printf(".zero3 = {");
713
714 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
715 printf("%X, ", s->zero3[i]);
716
717 printf("},\n");
718
719 printf(".zero6 = {");
720
721 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
722 printf("%X, ", s->zero6[i]);
723
724 printf("},\n");
725 }
726
727 static void
728 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
729 {
730 pandecode_log(".unknown%d = {", name);
731
732 for (int i = 0; i < count; ++i)
733 printf("%X, ", slide[i]);
734
735 pandecode_log("},\n");
736 }
737
738 #define SHORT_SLIDE(num) \
739 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
740
741 static void
742 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
743 {
744 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
745 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
746
747 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
748 pandecode_indent++;
749
750 SHORT_SLIDE(1);
751
752 pandecode_indent--;
753 printf("},\n");
754 }
755
756 static void
757 pandecode_swizzle(unsigned swizzle)
758 {
759 pandecode_prop("swizzle = %s | (%s << 3) | (%s << 6) | (%s << 9)",
760 pandecode_channel((swizzle >> 0) & 0x7),
761 pandecode_channel((swizzle >> 3) & 0x7),
762 pandecode_channel((swizzle >> 6) & 0x7),
763 pandecode_channel((swizzle >> 9) & 0x7));
764 }
765
766 static void
767 pandecode_rt_format(struct mali_rt_format format)
768 {
769 pandecode_log(".format = {\n");
770 pandecode_indent++;
771
772 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
773 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
774 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
775
776 pandecode_prop("block = %s",
777 pandecode_mfbd_block_format(format.block));
778
779 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
780 MALI_NEGATIVE(format.nr_channels));
781
782 pandecode_log(".flags = ");
783 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
784 pandecode_log_cont(",\n");
785
786 pandecode_swizzle(format.swizzle);
787
788 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
789 * which is a faster hardware-based alternative to the wallpaper method
790 * to preserve framebuffer contents across frames. In practice, MFBD
791 * preload is buggy on Midgard, and so this is a chicken bit. If this
792 * bit isn't set, most likely something broke unrelated to preload */
793
794 if (!format.no_preload) {
795 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
796 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
797 }
798
799 if (format.zero)
800 pandecode_prop("zero = 0x%" PRIx32, format.zero);
801
802 pandecode_indent--;
803 pandecode_log("},\n");
804 }
805
806 static void
807 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
808 {
809 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
810 pandecode_indent++;
811
812 for (int i = 0; i < MALI_NEGATIVE(fb->rt_count_1); i++) {
813 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
814 struct pandecode_mapped_memory *mem =
815 pandecode_find_mapped_gpu_mem_containing(rt_va);
816 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
817
818 pandecode_log("{\n");
819 pandecode_indent++;
820
821 pandecode_rt_format(rt->format);
822
823 if (rt->format.block == MALI_MFBD_BLOCK_AFBC) {
824 pandecode_log(".afbc = {\n");
825 pandecode_indent++;
826
827 char *a = pointer_as_memory_reference(rt->afbc.metadata);
828 pandecode_prop("metadata = %s", a);
829 free(a);
830
831 pandecode_prop("stride = %d", rt->afbc.stride);
832 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
833
834 pandecode_indent--;
835 pandecode_log("},\n");
836 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
837 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
838 rt->afbc.metadata,
839 rt->afbc.stride,
840 rt->afbc.unk);
841 }
842
843 MEMORY_PROP(rt, framebuffer);
844 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
845
846 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
847 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
848 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
849 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
850 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
851 }
852
853 if (rt->zero1 || rt->zero2 || rt->zero3) {
854 pandecode_msg("XXX: render target zeros tripped\n");
855 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
856 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
857 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
858 }
859
860 pandecode_indent--;
861 pandecode_log("},\n");
862 }
863
864 pandecode_indent--;
865 pandecode_log("};\n");
866 }
867
868 static unsigned
869 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
870 {
871 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
872 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
873
874 if (fb->sample_locations) {
875 /* The blob stores all possible sample locations in a single buffer
876 * allocated on startup, and just switches the pointer when switching
877 * MSAA state. For now, we just put the data into the cmdstream, but we
878 * should do something like what the blob does with a real driver.
879 *
880 * There seem to be 32 slots for sample locations, followed by another
881 * 16. The second 16 is just the center location followed by 15 zeros
882 * in all the cases I've identified (maybe shader vs. depth/color
883 * samples?).
884 */
885
886 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
887
888 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
889
890 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
891 pandecode_indent++;
892
893 for (int i = 0; i < 32 + 16; i++) {
894 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
895 }
896
897 pandecode_indent--;
898 pandecode_log("};\n");
899 }
900
901 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
902 pandecode_indent++;
903
904 pandecode_prop("unk0 = 0x%x", fb->unk0);
905
906 if (fb->sample_locations)
907 pandecode_prop("sample_locations = sample_locations_%d", job_no);
908
909 /* Assume that unknown1 was emitted in the last job for
910 * now */
911 MEMORY_PROP(fb, unknown1);
912
913 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
914 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
915 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
916 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
917
918 pandecode_prop("unk1 = 0x%x", fb->unk1);
919 pandecode_prop("unk2 = 0x%x", fb->unk2);
920 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
921 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
922
923 pandecode_log(".mfbd_flags = ");
924 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
925 pandecode_log_cont(",\n");
926
927 if (fb->clear_stencil)
928 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
929
930 if (fb->clear_depth)
931 pandecode_prop("clear_depth = %f", fb->clear_depth);
932
933 /* TODO: What is this? Let's not blow up.. */
934 if (fb->unknown2 != 0x1F)
935 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
936
937 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
938 MEMORY_PROP(fb, scratchpad);
939 const struct midgard_tiler_descriptor t = fb->tiler;
940 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment);
941
942 if (fb->zero3 || fb->zero4) {
943 pandecode_msg("XXX: framebuffer zeros tripped\n");
944 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
945 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
946 }
947
948 pandecode_indent--;
949 pandecode_log("};\n");
950
951 gpu_va += sizeof(struct bifrost_framebuffer);
952
953 if ((fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment) {
954 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
955 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
956
957 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
958 pandecode_indent++;
959
960 MEMORY_PROP(fbx, checksum);
961
962 if (fbx->checksum_stride)
963 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
964
965 pandecode_log(".flags = ");
966 pandecode_log_decoded_flags(mfbd_extra_flag_info, fbx->flags);
967 pandecode_log_cont(",\n");
968
969 if (fbx->flags & MALI_EXTRA_AFBC_ZS) {
970 pandecode_log(".ds_afbc = {\n");
971 pandecode_indent++;
972
973 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
974 pandecode_prop("depth_stencil_afbc_stride = %d",
975 fbx->ds_afbc.depth_stencil_afbc_stride);
976 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
977
978 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
979 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
980 pandecode_prop("zero1 = 0x%" PRIx32,
981 fbx->ds_afbc.zero1);
982 pandecode_prop("padding = 0x%" PRIx64,
983 fbx->ds_afbc.padding);
984 }
985
986 pandecode_indent--;
987 pandecode_log("},\n");
988 } else {
989 pandecode_log(".ds_linear = {\n");
990 pandecode_indent++;
991
992 if (fbx->ds_linear.depth) {
993 MEMORY_PROP_DIR(fbx->ds_linear, depth);
994 pandecode_prop("depth_stride = %d",
995 fbx->ds_linear.depth_stride);
996 }
997
998 if (fbx->ds_linear.stencil) {
999 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1000 pandecode_prop("stencil_stride = %d",
1001 fbx->ds_linear.stencil_stride);
1002 }
1003
1004 if (fbx->ds_linear.depth_stride_zero ||
1005 fbx->ds_linear.stencil_stride_zero ||
1006 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1007 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1008 pandecode_prop("depth_stride_zero = 0x%x",
1009 fbx->ds_linear.depth_stride_zero);
1010 pandecode_prop("stencil_stride_zero = 0x%x",
1011 fbx->ds_linear.stencil_stride_zero);
1012 pandecode_prop("zero1 = 0x%" PRIx32,
1013 fbx->ds_linear.zero1);
1014 pandecode_prop("zero2 = 0x%" PRIx32,
1015 fbx->ds_linear.zero2);
1016 }
1017
1018 pandecode_indent--;
1019 pandecode_log("},\n");
1020 }
1021
1022 if (fbx->zero3 || fbx->zero4) {
1023 pandecode_msg("XXX: fb_extra zeros tripped\n");
1024 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1025 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1026 }
1027
1028 pandecode_indent--;
1029 pandecode_log("};\n");
1030
1031 gpu_va += sizeof(struct bifrost_fb_extra);
1032 }
1033
1034 if (is_fragment)
1035 pandecode_render_target(gpu_va, job_no, fb);
1036
1037 /* Passback the render target count */
1038 return MALI_NEGATIVE(fb->rt_count_1);
1039 }
1040
1041 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1042 * count */
1043
1044 static void
1045 pandecode_padded_vertices(unsigned shift, unsigned k)
1046 {
1047 unsigned odd = 2*k + 1;
1048 unsigned pot = 1 << shift;
1049 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1050 }
1051
1052 /* Given a magic divisor, recover what we were trying to divide by.
1053 *
1054 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1055 * 0 <= m < 2^N, for N bits in m.
1056 *
1057 * Let q represent the number we would like to divide by.
1058 *
1059 * By definition of a magic divisor for N-bit unsigned integers (a number you
1060 * multiply by to magically get division), m is a number such that:
1061 *
1062 * (m * x) & (2^N - 1) = floor(x/q).
1063 * for all x on Z where 0 <= x < 2^N
1064 *
1065 * Ignore the case where any of the above values equals zero; it is irrelevant
1066 * for our purposes (instanced arrays).
1067 *
1068 * Choose x = q. Then:
1069 *
1070 * (m * x) & (2^N - 1) = floor(x/q).
1071 * (m * q) & (2^N - 1) = floor(q/q).
1072 *
1073 * floor(q/q) = floor(1) = 1, therefore:
1074 *
1075 * (m * q) & (2^N - 1) = 1
1076 *
1077 * Recall the identity that the bitwise AND of one less than a power-of-two
1078 * equals the modulo with that power of two, i.e. for all x:
1079 *
1080 * x & (2^N - 1) = x % N
1081 *
1082 * Therefore:
1083 *
1084 * mq % (2^N) = 1
1085 *
1086 * By definition, a modular multiplicative inverse of a number m is the number
1087 * q such that with respect to a modulos M:
1088 *
1089 * mq % M = 1
1090 *
1091 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1092 *
1093 */
1094
1095 static void
1096 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1097 {
1098 #if 0
1099 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1100 * shift) the most lame way possible... just repeatedly add.
1101 * Asymptoptically slow but nobody cares in practice, unless you have
1102 * massive numbers of vertices or high divisors. */
1103
1104 unsigned inverse = 0;
1105
1106 /* Magic implicitly has the highest bit set */
1107 magic |= (1 << 31);
1108
1109 /* Depending on rounding direction */
1110 if (extra)
1111 magic++;
1112
1113 for (;;) {
1114 uint32_t product = magic * inverse;
1115
1116 if (shift) {
1117 product >>= shift;
1118 }
1119
1120 if (product == 1)
1121 break;
1122
1123 ++inverse;
1124 }
1125
1126 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1127
1128 /* Recall we're supposed to divide by (gl_level_divisor *
1129 * padded_num_vertices) */
1130
1131 unsigned padded_num_vertices = inverse / orig_divisor;
1132
1133 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1134 #endif
1135 }
1136
1137 static void
1138 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1139 mali_ptr addr, int job_no, char *suffix,
1140 int count, bool varying)
1141 {
1142 char *prefix = varying ? "varyings" : "attributes";
1143
1144 if (!addr) {
1145 pandecode_msg("no %s\n", prefix);
1146 return;
1147 }
1148
1149 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1150
1151 char base[128];
1152 snprintf(base, sizeof(base), "%s_data_%d%s", prefix, job_no, suffix);
1153
1154 for (int i = 0; i < count; ++i) {
1155 enum mali_attr_mode mode = attr[i].elements & 7;
1156
1157 if (mode == MALI_ATTR_UNUSED)
1158 continue;
1159
1160 mali_ptr raw_elements = attr[i].elements & ~7;
1161
1162 /* TODO: Do we maybe want to dump the attribute values
1163 * themselves given the specified format? Or is that too hard?
1164 * */
1165
1166 char *a = pointer_as_memory_reference(raw_elements);
1167 pandecode_log("mali_ptr %s_%d_p = %s;\n", base, i, a);
1168 free(a);
1169 }
1170
1171 pandecode_log("union mali_attr %s_%d[] = {\n", prefix, job_no);
1172 pandecode_indent++;
1173
1174 for (int i = 0; i < count; ++i) {
1175 pandecode_log("{\n");
1176 pandecode_indent++;
1177
1178 unsigned mode = attr[i].elements & 7;
1179 pandecode_prop("elements = (%s_%d_p) | %s", base, i, pandecode_attr_mode(mode));
1180 pandecode_prop("shift = %d", attr[i].shift);
1181 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1182 pandecode_prop("stride = 0x%" PRIx32, attr[i].stride);
1183 pandecode_prop("size = 0x%" PRIx32, attr[i].size);
1184
1185 /* Decode further where possible */
1186
1187 if (mode == MALI_ATTR_MODULO) {
1188 pandecode_padded_vertices(
1189 attr[i].shift,
1190 attr[i].extra_flags);
1191 }
1192
1193 pandecode_indent--;
1194 pandecode_log("}, \n");
1195
1196 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1197 i++;
1198 pandecode_log("{\n");
1199 pandecode_indent++;
1200 pandecode_prop("unk = 0x%x", attr[i].unk);
1201 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1202 if (attr[i].zero != 0)
1203 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1204 pandecode_prop("divisor = %d", attr[i].divisor);
1205 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1206 pandecode_indent--;
1207 pandecode_log("}, \n");
1208 }
1209
1210 }
1211
1212 pandecode_indent--;
1213 pandecode_log("};\n");
1214 }
1215
1216 static mali_ptr
1217 pandecode_shader_address(const char *name, mali_ptr ptr)
1218 {
1219 /* TODO: Decode flags */
1220 mali_ptr shader_ptr = ptr & ~15;
1221
1222 char *a = pointer_as_memory_reference(shader_ptr);
1223 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1224 free(a);
1225
1226 return shader_ptr;
1227 }
1228
1229 static bool
1230 all_zero(unsigned *buffer, unsigned count)
1231 {
1232 for (unsigned i = 0; i < count; ++i) {
1233 if (buffer[i])
1234 return false;
1235 }
1236
1237 return true;
1238 }
1239
1240 static void
1241 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1242 {
1243 if (all_zero((unsigned *) stencil, sizeof(stencil) / sizeof(unsigned)))
1244 return;
1245
1246 const char *func = pandecode_func(stencil->func);
1247 const char *sfail = pandecode_stencil_op(stencil->sfail);
1248 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1249 const char *dppass = pandecode_stencil_op(stencil->dppass);
1250
1251 if (stencil->zero)
1252 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1253
1254 pandecode_log(".stencil_%s = {\n", name);
1255 pandecode_indent++;
1256 pandecode_prop("ref = %d", stencil->ref);
1257 pandecode_prop("mask = 0x%02X", stencil->mask);
1258 pandecode_prop("func = %s", func);
1259 pandecode_prop("sfail = %s", sfail);
1260 pandecode_prop("dpfail = %s", dpfail);
1261 pandecode_prop("dppass = %s", dppass);
1262 pandecode_indent--;
1263 pandecode_log("},\n");
1264 }
1265
1266 static void
1267 pandecode_blend_equation(const struct mali_blend_equation *blend)
1268 {
1269 if (blend->zero1)
1270 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1271
1272 pandecode_log(".equation = {\n");
1273 pandecode_indent++;
1274
1275 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1276 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1277
1278 pandecode_log(".color_mask = ");
1279 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1280 pandecode_log_cont(",\n");
1281
1282 pandecode_indent--;
1283 pandecode_log("},\n");
1284 }
1285
1286 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1287
1288 static unsigned
1289 decode_bifrost_constant(u16 constant)
1290 {
1291 float lo = (float) (constant & 0xFF);
1292 float hi = (float) (constant >> 8);
1293
1294 return (hi / 255.0) + (lo / 65535.0);
1295 }
1296
1297 static mali_ptr
1298 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1299 {
1300 struct bifrost_blend_rt *b =
1301 ((struct bifrost_blend_rt *) descs) + rt_no;
1302
1303 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1304 pandecode_indent++;
1305
1306 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1307 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1308 b->constant, decode_bifrost_constant(b->constant));
1309
1310 /* TODO figure out blend shader enable bit */
1311 pandecode_blend_equation(&b->equation);
1312 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1313 pandecode_prop("index = 0x%" PRIx16, b->index);
1314 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1315
1316 pandecode_indent--;
1317 pandecode_log("},\n");
1318
1319 return 0;
1320 }
1321
1322 static mali_ptr
1323 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1324 {
1325 if (all_zero((unsigned *) blend, sizeof(blend) / sizeof(unsigned)))
1326 return 0;
1327
1328 pandecode_log(".blend = {\n");
1329 pandecode_indent++;
1330
1331 if (is_shader) {
1332 pandecode_shader_address("shader", blend->shader);
1333 } else {
1334 pandecode_blend_equation(&blend->equation);
1335 pandecode_prop("constant = %f", blend->constant);
1336 }
1337
1338 pandecode_indent--;
1339 pandecode_log("},\n");
1340
1341 /* Return blend shader to disassemble if present */
1342 return is_shader ? (blend->shader & ~0xF) : 0;
1343 }
1344
1345 static mali_ptr
1346 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1347 {
1348 struct midgard_blend_rt *b =
1349 ((struct midgard_blend_rt *) descs) + rt_no;
1350
1351 /* Flags determine presence of blend shader */
1352 bool is_shader = (b->flags & 0xF) >= 0x2;
1353
1354 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1355 pandecode_indent++;
1356
1357 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1358
1359 union midgard_blend blend = b->blend;
1360 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1361
1362 pandecode_indent--;
1363 pandecode_log("};\n");
1364
1365 return shader;
1366 }
1367
1368 /* Attributes and varyings have descriptor records, which contain information
1369 * about their format and ordering with the attribute/varying buffers. We'll
1370 * want to validate that the combinations specified are self-consistent.
1371 */
1372
1373 /* Extracts the number of components associated with a Mali format */
1374
1375 static unsigned
1376 pandecode_format_component_count(enum mali_format fmt)
1377 {
1378 /* Mask out the format class */
1379 unsigned top = fmt & 0b11100000;
1380
1381 switch (top) {
1382 case MALI_FORMAT_SNORM:
1383 case MALI_FORMAT_UINT:
1384 case MALI_FORMAT_UNORM:
1385 case MALI_FORMAT_SINT:
1386 return ((fmt >> 3) & 3) + 1;
1387 default:
1388 /* TODO: Validate */
1389 return 4;
1390 }
1391 }
1392
1393 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
1394
1395 static unsigned
1396 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
1397 {
1398 unsigned mask = 0;
1399 assert(MALI_CHANNEL_RED == 0);
1400
1401 for (unsigned c = 0; c < 4; ++c) {
1402 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
1403
1404 if (chan <= MALI_CHANNEL_ALPHA)
1405 mask |= (1 << chan);
1406 }
1407
1408 return mask;
1409 }
1410
1411 /* Validates that a (format, swizzle) pair is valid, in the sense that the
1412 * swizzle doesn't access any components that are undefined in the format.
1413 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
1414 * omitted */
1415
1416 static bool
1417 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
1418 {
1419 unsigned nr_comp = pandecode_format_component_count(fmt);
1420 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
1421 unsigned valid_mask = (1 << nr_comp) - 1;
1422 unsigned invalid_mask = ~valid_mask;
1423
1424 if (access_mask & invalid_mask) {
1425 pandecode_msg("XXX: invalid components accessed\n");
1426 return false;
1427 }
1428
1429 /* Check for the default non-swizzling swizzle so we can suppress
1430 * useless printing for the defaults */
1431
1432 unsigned default_swizzles[4] = {
1433 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1434 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1435 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
1436 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
1437 };
1438
1439 return (swizzle == default_swizzles[nr_comp - 1]);
1440 }
1441
1442 static int
1443 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1444 {
1445 char base[128];
1446 char *prefix = varying ? "varying" : "attribute";
1447 unsigned max_index = 0;
1448 snprintf(base, sizeof(base), "%s_meta", prefix);
1449
1450 pandecode_log("struct mali_attr_meta %s_%d%s[] = {\n", base, job_no, suffix);
1451 pandecode_indent++;
1452
1453 struct mali_attr_meta *attr_meta;
1454 mali_ptr p = varying ? (v->varying_meta & ~0xF) : v->attribute_meta;
1455
1456 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1457
1458 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1459 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1460 sizeof(*attr_mem));
1461
1462 /* If the record is discard, it should be zero for everything else */
1463
1464 if (attr_meta->format == MALI_VARYING_DISCARD) {
1465 uint64_t zero =
1466 attr_meta->index |
1467 attr_meta->unknown1 |
1468 attr_meta->unknown3 |
1469 attr_meta->src_offset;
1470
1471 if (zero)
1472 pandecode_msg("XXX: expected empty record for varying discard\n");
1473
1474 /* We want to look for a literal 0000 swizzle -- this
1475 * is not encoded with all zeroes, however */
1476
1477 enum mali_channel z = MALI_CHANNEL_ZERO;
1478 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1479 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1480
1481 if (!good_swizzle)
1482 pandecode_msg("XXX: expected zero swizzle for discard\n");
1483
1484 if (!varying)
1485 pandecode_msg("XXX: cannot discard attribute\n");
1486
1487 /* If we're all good, omit the record */
1488 if (!zero && varying && good_swizzle) {
1489 pandecode_log("/* discarded varying */\n");
1490 continue;
1491 }
1492 }
1493
1494 pandecode_log("{\n");
1495 pandecode_indent++;
1496 pandecode_prop("index = %d", attr_meta->index);
1497
1498 if (attr_meta->index > max_index)
1499 max_index = attr_meta->index;
1500
1501 /* Check the swizzle/format, and if they match and the swizzle
1502 * is simple enough, we can avoid printing the swizzle since
1503 * it's just noise */
1504
1505 bool trivial_swizzle = pandecode_validate_format_swizzle(
1506 attr_meta->format, attr_meta->swizzle);
1507
1508 if (!trivial_swizzle)
1509 pandecode_swizzle(attr_meta->swizzle);
1510
1511 pandecode_prop("format = %s", pandecode_format(attr_meta->format));
1512
1513 if (attr_meta->unknown1 != 0x2) {
1514 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1515 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1516 }
1517
1518 if (attr_meta->unknown3) {
1519 pandecode_msg("XXX: unexpected unknown3 set\n");
1520 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1521 }
1522
1523 pandecode_prop("src_offset = %d", attr_meta->src_offset);
1524 pandecode_indent--;
1525 pandecode_log("},\n");
1526
1527 }
1528
1529 pandecode_indent--;
1530 pandecode_log("};\n");
1531
1532 return count ? (max_index + 1) : 0;
1533 }
1534
1535 static void
1536 pandecode_indices(uintptr_t pindices, uint32_t index_count, int job_no)
1537 {
1538 struct pandecode_mapped_memory *imem = pandecode_find_mapped_gpu_mem_containing(pindices);
1539
1540 if (imem) {
1541 /* Indices are literally just a u32 array :) */
1542
1543 uint32_t *PANDECODE_PTR_VAR(indices, imem, pindices);
1544
1545 pandecode_log("uint32_t indices_%d[] = {\n", job_no);
1546 pandecode_indent++;
1547
1548 for (unsigned i = 0; i < (index_count + 1); i += 3)
1549 pandecode_log("%d, %d, %d,\n",
1550 indices[i],
1551 indices[i + 1],
1552 indices[i + 2]);
1553
1554 pandecode_indent--;
1555 pandecode_log("};\n");
1556 }
1557 }
1558
1559 /* return bits [lo, hi) of word */
1560 static u32
1561 bits(u32 word, u32 lo, u32 hi)
1562 {
1563 if (hi - lo >= 32)
1564 return word; // avoid undefined behavior with the shift
1565
1566 return (word >> lo) & ((1 << (hi - lo)) - 1);
1567 }
1568
1569 static void
1570 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool noninstanced)
1571 {
1572 pandecode_log_cont("{\n");
1573 pandecode_indent++;
1574
1575 /* Decode invocation_count. See the comment before the definition of
1576 * invocation_count for an explanation.
1577 */
1578
1579 unsigned size_x = bits(p->invocation_count, 0, p->size_y_shift) + 1;
1580 unsigned size_y = bits(p->invocation_count, p->size_y_shift, p->size_z_shift) + 1;
1581 unsigned size_z = bits(p->invocation_count, p->size_z_shift, p->workgroups_x_shift) + 1;
1582
1583 unsigned groups_x = bits(p->invocation_count, p->workgroups_x_shift, p->workgroups_y_shift) + 1;
1584 unsigned groups_y = bits(p->invocation_count, p->workgroups_y_shift, p->workgroups_z_shift) + 1;
1585 unsigned groups_z = bits(p->invocation_count, p->workgroups_z_shift, 32) + 1;
1586
1587 /* Even though we have this decoded, we want to ensure that the
1588 * representation is "unique" so we don't lose anything by printing only
1589 * the final result. More specifically, we need to check that we were
1590 * passed something in canonical form, since the definition per the
1591 * hardware is inherently not unique. How? Well, take the resulting
1592 * decode and pack it ourselves! If it is bit exact with what we
1593 * decoded, we're good to go. */
1594
1595 struct mali_vertex_tiler_prefix ref;
1596 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, noninstanced);
1597
1598 bool canonical =
1599 (p->invocation_count == ref.invocation_count) &&
1600 (p->size_y_shift == ref.size_y_shift) &&
1601 (p->size_z_shift == ref.size_z_shift) &&
1602 (p->workgroups_x_shift == ref.workgroups_x_shift) &&
1603 (p->workgroups_y_shift == ref.workgroups_y_shift) &&
1604 (p->workgroups_z_shift == ref.workgroups_z_shift) &&
1605 (p->workgroups_x_shift_2 == ref.workgroups_x_shift_2);
1606
1607 if (!canonical) {
1608 pandecode_msg("XXX: non-canonical workgroups packing\n");
1609 pandecode_msg("expected: %X, %d, %d, %d, %d, %d\n",
1610 ref.invocation_count,
1611 ref.size_y_shift,
1612 ref.size_z_shift,
1613 ref.workgroups_x_shift,
1614 ref.workgroups_y_shift,
1615 ref.workgroups_z_shift,
1616 ref.workgroups_x_shift_2);
1617
1618 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1619 pandecode_prop("size_y_shift = %d", p->size_y_shift);
1620 pandecode_prop("size_z_shift = %d", p->size_z_shift);
1621 pandecode_prop("workgroups_x_shift = %d", p->workgroups_x_shift);
1622 pandecode_prop("workgroups_y_shift = %d", p->workgroups_y_shift);
1623 pandecode_prop("workgroups_z_shift = %d", p->workgroups_z_shift);
1624 pandecode_prop("workgroups_x_shift_2 = %d", p->workgroups_x_shift_2);
1625 }
1626
1627 /* Regardless, print the decode */
1628 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1629 size_x, size_y, size_z,
1630 groups_x, groups_y, groups_z);
1631
1632 /* TODO: Decode */
1633 if (p->unknown_draw)
1634 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1635
1636 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1637
1638 if (p->draw_mode != MALI_DRAW_NONE)
1639 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1640
1641 /* Index count only exists for tiler jobs anyway */
1642
1643 if (p->index_count)
1644 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1645
1646 if (p->offset_bias_correction)
1647 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1648
1649 /* TODO: Figure out what this is. It's not zero */
1650 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1651
1652 pandecode_indent--;
1653 pandecode_log("},\n");
1654 }
1655
1656 static void
1657 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1658 {
1659 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1660 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1661
1662 pandecode_log("struct mali_uniform_buffer_meta uniform_buffers_%"PRIx64"_%d[] = {\n",
1663 pubufs, job_no);
1664 pandecode_indent++;
1665
1666 for (int i = 0; i < ubufs_count; i++) {
1667 pandecode_log("{\n");
1668 pandecode_indent++;
1669
1670 unsigned size = (ubufs[i].size + 1) * 16;
1671 mali_ptr addr = ubufs[i].ptr << 2;
1672
1673 pandecode_validate_buffer(addr, size);
1674
1675 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1676 pandecode_prop("size = %u", size);
1677 pandecode_prop("ptr = (%s) >> 2", ptr);
1678 pandecode_indent--;
1679 pandecode_log("},\n");
1680 free(ptr);
1681 }
1682
1683 pandecode_indent--;
1684 pandecode_log("};\n");
1685 }
1686
1687 static void
1688 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1689 {
1690
1691 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1692
1693 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1694
1695 if (scratchpad->zero) {
1696 pandecode_msg("XXX: scratchpad zero tripped");
1697 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1698 }
1699
1700 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1701 pandecode_indent++;
1702
1703 pandecode_prop("flags = 0x%x", scratchpad->flags);
1704 MEMORY_PROP(scratchpad, gpu_scratchpad);
1705
1706 pandecode_indent--;
1707 pandecode_log("};\n");
1708 }
1709
1710 static unsigned shader_id = 0;
1711
1712 static void
1713 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1714 bool is_bifrost, unsigned nr_regs)
1715 {
1716 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1717 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1718
1719 /* Compute maximum possible size */
1720 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1721
1722 /* Print some boilerplate to clearly denote the assembly (which doesn't
1723 * obey indentation rules), and actually do the disassembly! */
1724
1725 printf("\n\n");
1726
1727 char prefix[512];
1728
1729 snprintf(prefix, sizeof(prefix) - 1, "shader%d - %s shader: ",
1730 shader_id++,
1731 (type == JOB_TYPE_TILER) ? "FRAGMENT" : "VERTEX");
1732
1733 if (is_bifrost) {
1734 disassemble_bifrost(code, sz, false);
1735 } else {
1736 disassemble_midgard(code, sz, true, nr_regs, prefix);
1737 }
1738
1739 printf("\n\n");
1740 }
1741
1742 static void
1743 pandecode_vertex_tiler_postfix_pre(const struct mali_vertex_tiler_postfix *p,
1744 int job_no, enum mali_job_type job_type,
1745 char *suffix, bool is_bifrost)
1746 {
1747 mali_ptr shader_meta_ptr = (u64) (uintptr_t) (p->_shader_upper << 4);
1748 struct pandecode_mapped_memory *attr_mem;
1749
1750 unsigned rt_count = 1;
1751
1752 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
1753 * are the only things actually needed from the FBD, vertex/tiler jobs
1754 * no longer reference the FBD -- instead, this field points to some
1755 * info about the scratchpad.
1756 */
1757 if (is_bifrost)
1758 pandecode_scratchpad(p->framebuffer & ~FBD_TYPE, job_no, suffix);
1759 else if (p->framebuffer & MALI_MFBD)
1760 rt_count = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
1761 else if (job_type == JOB_TYPE_COMPUTE)
1762 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
1763 else
1764 pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false);
1765
1766 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1767 int texture_count = 0, sampler_count = 0;
1768
1769 if (shader_meta_ptr) {
1770 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(shader_meta_ptr);
1771 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, shader_meta_ptr);
1772
1773 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", shader_meta_ptr, job_no, suffix);
1774 pandecode_indent++;
1775
1776 /* Save for dumps */
1777 attribute_count = s->attribute_count;
1778 varying_count = s->varying_count;
1779 texture_count = s->texture_count;
1780 sampler_count = s->sampler_count;
1781
1782 if (is_bifrost) {
1783 uniform_count = s->bifrost2.uniform_count;
1784 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
1785 } else {
1786 uniform_count = s->midgard1.uniform_buffer_count;
1787 uniform_buffer_count = s->midgard1.uniform_buffer_count;
1788 }
1789
1790 mali_ptr shader_ptr = pandecode_shader_address("shader", s->shader);
1791
1792 pandecode_prop("texture_count = %" PRId16, s->texture_count);
1793 pandecode_prop("sampler_count = %" PRId16, s->sampler_count);
1794 pandecode_prop("attribute_count = %" PRId16, s->attribute_count);
1795 pandecode_prop("varying_count = %" PRId16, s->varying_count);
1796
1797 unsigned nr_registers = 0;
1798
1799 if (is_bifrost) {
1800 pandecode_log(".bifrost1 = {\n");
1801 pandecode_indent++;
1802
1803 pandecode_prop("uniform_buffer_count = %" PRId32, s->bifrost1.uniform_buffer_count);
1804 pandecode_prop("unk1 = 0x%" PRIx32, s->bifrost1.unk1);
1805
1806 pandecode_indent--;
1807 pandecode_log("},\n");
1808 } else {
1809 pandecode_log(".midgard1 = {\n");
1810 pandecode_indent++;
1811
1812 pandecode_prop("uniform_count = %" PRId16, s->midgard1.uniform_count);
1813 pandecode_prop("uniform_buffer_count = %" PRId16, s->midgard1.uniform_buffer_count);
1814 pandecode_prop("work_count = %" PRId16, s->midgard1.work_count);
1815 nr_registers = s->midgard1.work_count;
1816
1817 pandecode_log(".flags = ");
1818 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
1819 pandecode_log_cont(",\n");
1820
1821 pandecode_prop("unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
1822
1823 pandecode_indent--;
1824 pandecode_log("},\n");
1825 }
1826
1827 if (s->depth_units || s->depth_factor) {
1828 pandecode_prop("depth_factor = %f", s->depth_factor);
1829 pandecode_prop("depth_units = %f", s->depth_units);
1830 }
1831
1832 if (s->alpha_coverage) {
1833 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
1834 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
1835
1836 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
1837 invert_alpha_coverage ? "~" : "",
1838 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
1839 }
1840
1841 if (s->unknown2_3 || s->unknown2_4) {
1842 pandecode_log(".unknown2_3 = ");
1843
1844 int unknown2_3 = s->unknown2_3;
1845 int unknown2_4 = s->unknown2_4;
1846
1847 /* We're not quite sure what these flags mean without the depth test, if anything */
1848
1849 if (unknown2_3 & (MALI_DEPTH_TEST | MALI_DEPTH_FUNC_MASK)) {
1850 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
1851 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
1852
1853 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
1854 }
1855
1856 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
1857 pandecode_log_cont(",\n");
1858
1859 pandecode_log(".unknown2_4 = ");
1860 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
1861 pandecode_log_cont(",\n");
1862 }
1863
1864 if (s->stencil_mask_front || s->stencil_mask_back) {
1865 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
1866 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
1867 }
1868
1869 pandecode_stencil("front", &s->stencil_front);
1870 pandecode_stencil("back", &s->stencil_back);
1871
1872 if (is_bifrost) {
1873 pandecode_log(".bifrost2 = {\n");
1874 pandecode_indent++;
1875
1876 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
1877 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
1878 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
1879 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
1880
1881 pandecode_indent--;
1882 pandecode_log("},\n");
1883 } else if (s->midgard2.unknown2_7) {
1884 pandecode_log(".midgard2 = {\n");
1885 pandecode_indent++;
1886
1887 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
1888 pandecode_indent--;
1889 pandecode_log("},\n");
1890 }
1891
1892 if (s->unknown2_8)
1893 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
1894
1895 if (!is_bifrost) {
1896 /* TODO: Blend shaders routing/disasm */
1897
1898 union midgard_blend blend = s->blend;
1899 pandecode_midgard_blend(&blend, false);
1900 }
1901
1902 pandecode_indent--;
1903 pandecode_log("};\n");
1904
1905 /* MRT blend fields are used whenever MFBD is used, with
1906 * per-RT descriptors */
1907
1908 if (job_type == JOB_TYPE_TILER) {
1909 void* blend_base = (void *) (s + 1);
1910
1911 for (unsigned i = 0; i < rt_count; i++) {
1912 mali_ptr shader = 0;
1913
1914 if (is_bifrost)
1915 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1916 else
1917 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1918
1919 if (shader & ~0xF)
1920 pandecode_shader_disassemble(shader, job_no, job_type, false, 0);
1921 }
1922 }
1923
1924 if (shader_ptr & ~0xF)
1925 pandecode_shader_disassemble(shader_ptr, job_no, job_type, is_bifrost, nr_registers);
1926 } else
1927 pandecode_msg("<no shader>\n");
1928
1929 if (p->viewport) {
1930 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
1931 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
1932
1933 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
1934 pandecode_indent++;
1935
1936 pandecode_prop("clip_minx = %f", f->clip_minx);
1937 pandecode_prop("clip_miny = %f", f->clip_miny);
1938 pandecode_prop("clip_minz = %f", f->clip_minz);
1939 pandecode_prop("clip_maxx = %f", f->clip_maxx);
1940 pandecode_prop("clip_maxy = %f", f->clip_maxy);
1941 pandecode_prop("clip_maxz = %f", f->clip_maxz);
1942
1943 /* Only the higher coordinates are MALI_POSITIVE scaled */
1944
1945 pandecode_prop("viewport0 = { %d, %d }",
1946 f->viewport0[0], f->viewport0[1]);
1947
1948 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
1949 f->viewport1[0] + 1, f->viewport1[1] + 1);
1950
1951 pandecode_indent--;
1952 pandecode_log("};\n");
1953 }
1954
1955 if (p->attribute_meta) {
1956 unsigned max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
1957
1958 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
1959 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false);
1960 }
1961
1962 /* Varyings are encoded like attributes but not actually sent; we just
1963 * pass a zero buffer with the right stride/size set, (or whatever)
1964 * since the GPU will write to it itself */
1965
1966 if (p->varying_meta) {
1967 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
1968 }
1969
1970 if (p->varyings) {
1971 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
1972
1973 /* Number of descriptors depends on whether there are
1974 * non-internal varyings */
1975
1976 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true);
1977 }
1978
1979 if (p->uniform_buffers) {
1980 if (uniform_buffer_count)
1981 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1982 else
1983 pandecode_msg("XXX: UBOs specified but not referenced\n");
1984 } else if (uniform_buffer_count)
1985 pandecode_msg("XXX: UBOs referenced but not specified\n");
1986
1987 /* We don't want to actually dump uniforms, but we do need to validate
1988 * that the counts we were given are sane */
1989
1990 if (p->uniforms) {
1991 if (uniform_count)
1992 pandecode_validate_buffer(p->uniforms, uniform_count * 16);
1993 else
1994 pandecode_msg("XXX: Uniforms specified but not referenced");
1995 } else if (uniform_count)
1996 pandecode_msg("XXX: UBOs referenced but not specified\n");
1997
1998 if (p->texture_trampoline) {
1999 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2000
2001 if (mmem) {
2002 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2003
2004 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2005 pandecode_indent++;
2006
2007 for (int tex = 0; tex < texture_count; ++tex) {
2008 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2009 char *a = pointer_as_memory_reference(*u);
2010 pandecode_log("%s,\n", a);
2011 free(a);
2012 }
2013
2014 pandecode_indent--;
2015 pandecode_log("};\n");
2016
2017 /* Now, finally, descend down into the texture descriptor */
2018 for (int tex = 0; tex < texture_count; ++tex) {
2019 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2020 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2021
2022 if (tmem) {
2023 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, *u);
2024
2025 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", *u, job_no, tex);
2026 pandecode_indent++;
2027
2028 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", t->width + 1);
2029 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", t->height + 1);
2030 pandecode_prop("depth = MALI_POSITIVE(%" PRId16 ")", t->depth + 1);
2031 pandecode_prop("array_size = MALI_POSITIVE(%" PRId16 ")", t->array_size + 1);
2032 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2033 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2034 pandecode_prop("nr_mipmap_levels = %" PRId8, t->nr_mipmap_levels);
2035
2036 struct mali_texture_format f = t->format;
2037
2038 pandecode_log(".format = {\n");
2039 pandecode_indent++;
2040
2041 pandecode_swizzle(f.swizzle);
2042 pandecode_prop("format = %s", pandecode_format(f.format));
2043 pandecode_prop("type = %s", pandecode_texture_type(f.type));
2044 pandecode_prop("srgb = %" PRId32, f.srgb);
2045 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2046 pandecode_prop("usage2 = 0x%" PRIx32, f.usage2);
2047
2048 pandecode_indent--;
2049 pandecode_log("},\n");
2050
2051 pandecode_swizzle(t->swizzle);
2052
2053 if (t->swizzle_zero) {
2054 /* Shouldn't happen */
2055 pandecode_msg("XXX: swizzle zero tripped\n");
2056 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2057 }
2058
2059 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2060
2061 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2062 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2063 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2064
2065 pandecode_log(".payload = {\n");
2066 pandecode_indent++;
2067
2068 /* A bunch of bitmap pointers follow.
2069 * We work out the correct number,
2070 * based on the mipmap/cubemap
2071 * properties, but dump extra
2072 * possibilities to futureproof */
2073
2074 int bitmap_count = MALI_NEGATIVE(t->nr_mipmap_levels);
2075 bool manual_stride = f.usage2 & MALI_TEX_MANUAL_STRIDE;
2076
2077 /* Miptree for each face */
2078 if (f.type == MALI_TEX_CUBE)
2079 bitmap_count *= 6;
2080
2081 /* Array of textures */
2082 bitmap_count *= MALI_NEGATIVE(t->array_size);
2083
2084 /* Stride for each element */
2085 if (manual_stride)
2086 bitmap_count *= 2;
2087
2088 /* Sanity check the size */
2089 int max_count = sizeof(t->payload) / sizeof(t->payload[0]);
2090 assert (bitmap_count <= max_count);
2091
2092 for (int i = 0; i < bitmap_count; ++i) {
2093 /* How we dump depends if this is a stride or a pointer */
2094
2095 if ((f.usage2 & MALI_TEX_MANUAL_STRIDE) && (i & 1)) {
2096 /* signed 32-bit snuck in as a 64-bit pointer */
2097 uint64_t stride_set = t->payload[i];
2098 uint32_t clamped_stride = stride_set;
2099 int32_t stride = clamped_stride;
2100 assert(stride_set == clamped_stride);
2101 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2102 } else {
2103 char *a = pointer_as_memory_reference(t->payload[i]);
2104 pandecode_log("%s, \n", a);
2105 free(a);
2106 }
2107 }
2108
2109 pandecode_indent--;
2110 pandecode_log("},\n");
2111
2112 pandecode_indent--;
2113 pandecode_log("};\n");
2114 }
2115 }
2116 }
2117 }
2118
2119 if (p->sampler_descriptor) {
2120 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2121
2122 if (smem) {
2123 struct mali_sampler_descriptor *s;
2124
2125 mali_ptr d = p->sampler_descriptor;
2126
2127 for (int i = 0; i < sampler_count; ++i) {
2128 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2129
2130 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2131 pandecode_indent++;
2132
2133 pandecode_log(".filter_mode = ");
2134 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2135 pandecode_log_cont(",\n");
2136
2137 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2138 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2139
2140 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2141 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2142 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2143
2144 pandecode_prop("compare_func = %s", pandecode_alt_func(s->compare_func));
2145
2146 if (s->zero || s->zero2) {
2147 pandecode_msg("XXX: sampler zero tripped\n");
2148 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2149 }
2150
2151 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2152
2153 pandecode_prop("border_color = { %f, %f, %f, %f }",
2154 s->border_color[0],
2155 s->border_color[1],
2156 s->border_color[2],
2157 s->border_color[3]);
2158
2159 pandecode_indent--;
2160 pandecode_log("};\n");
2161 }
2162 }
2163 }
2164 }
2165
2166 static void
2167 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2168 {
2169 if (!(p->position_varying || p->occlusion_counter || p->flags))
2170 return;
2171
2172 pandecode_log(".postfix = {\n");
2173 pandecode_indent++;
2174
2175 MEMORY_PROP(p, position_varying);
2176 MEMORY_PROP(p, occlusion_counter);
2177
2178 if (p->flags)
2179 pandecode_prop("flags = %d", p->flags);
2180
2181 pandecode_indent--;
2182 pandecode_log("},\n");
2183 }
2184
2185 static void
2186 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2187 {
2188 pandecode_log_cont("{\n");
2189 pandecode_indent++;
2190
2191 pandecode_prop("unk2 = 0x%x", v->unk2);
2192
2193 if (v->zero0 || v->zero1) {
2194 pandecode_msg("XXX: vertex only zero tripped");
2195 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2196 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2197 }
2198
2199 pandecode_indent--;
2200 pandecode_log("}\n");
2201 }
2202
2203 static void
2204 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2205 {
2206
2207 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2208 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2209
2210 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2211 pandecode_indent++;
2212
2213 if (h->zero) {
2214 pandecode_msg("XXX: tiler heap zero tripped\n");
2215 pandecode_prop("zero = 0x%x", h->zero);
2216 }
2217
2218 for (int i = 0; i < 12; i++) {
2219 if (h->zeros[i] != 0) {
2220 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2221 i, h->zeros[i]);
2222 }
2223 }
2224
2225 pandecode_prop("heap_size = 0x%x", h->heap_size);
2226 MEMORY_PROP(h, tiler_heap_start);
2227 MEMORY_PROP(h, tiler_heap_free);
2228
2229 /* this might point to the beginning of another buffer, when it's
2230 * really the end of the tiler heap buffer, so we have to be careful
2231 * here. but for zero length, we need the same pointer.
2232 */
2233
2234 if (h->tiler_heap_end == h->tiler_heap_start) {
2235 MEMORY_PROP(h, tiler_heap_start);
2236 } else {
2237 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2238 pandecode_prop("tiler_heap_end = %s + 1", a);
2239 free(a);
2240 }
2241
2242 pandecode_indent--;
2243 pandecode_log("};\n");
2244 }
2245
2246 static void
2247 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2248 {
2249 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2250 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2251
2252 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2253
2254 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2255 pandecode_indent++;
2256
2257 if (t->zero0 || t->zero1) {
2258 pandecode_msg("XXX: tiler meta zero tripped\n");
2259 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2260 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2261 }
2262
2263 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2264 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2265
2266 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2267 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2268
2269 for (int i = 0; i < 12; i++) {
2270 if (t->zeros[i] != 0) {
2271 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2272 i, t->zeros[i]);
2273 }
2274 }
2275
2276 pandecode_indent--;
2277 pandecode_log("};\n");
2278 }
2279
2280 static void
2281 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2282 {
2283 pandecode_log(".gl_enables = ");
2284
2285 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2286
2287 pandecode_log_cont(",\n");
2288 }
2289
2290 static void
2291 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2292 {
2293 if (u.pointer == 0x0)
2294 return;
2295
2296 pandecode_log(".primitive_size = {\n");
2297 pandecode_indent++;
2298
2299 if (constant) {
2300 pandecode_prop("constant = %f", u.constant);
2301 } else {
2302 MEMORY_PROP((&u), pointer);
2303 }
2304
2305 pandecode_indent--;
2306 pandecode_log("},\n");
2307 }
2308
2309 static void
2310 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2311 {
2312 pandecode_log_cont("{\n");
2313 pandecode_indent++;
2314
2315 /* TODO: gl_PointSize on Bifrost */
2316 pandecode_primitive_size(t->primitive_size, true);
2317
2318 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2319
2320 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2321 || t->zero6 || t->zero7 || t->zero8) {
2322 pandecode_msg("XXX: tiler only zero tripped\n");
2323 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2324 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2325 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2326 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2327 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2328 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2329 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2330 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2331 }
2332
2333 pandecode_indent--;
2334 pandecode_log("},\n");
2335 }
2336
2337 static int
2338 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2339 const struct pandecode_mapped_memory *mem,
2340 mali_ptr payload, int job_no)
2341 {
2342 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2343
2344 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true);
2345
2346 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2347 pandecode_indent++;
2348
2349 pandecode_log(".prefix = ");
2350 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2351
2352 pandecode_log(".vertex = ");
2353 pandecode_vertex_only_bfr(&v->vertex);
2354
2355 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2356
2357 pandecode_indent--;
2358 pandecode_log("};\n");
2359
2360 return sizeof(*v);
2361 }
2362
2363 static int
2364 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2365 const struct pandecode_mapped_memory *mem,
2366 mali_ptr payload, int job_no)
2367 {
2368 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2369
2370 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true);
2371
2372 pandecode_indices(t->prefix.indices, t->prefix.index_count, job_no);
2373 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2374
2375 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2376 pandecode_indent++;
2377
2378 pandecode_log(".prefix = ");
2379 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2380
2381 pandecode_log(".tiler = ");
2382 pandecode_tiler_only_bfr(&t->tiler, job_no);
2383
2384 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2385
2386 pandecode_indent--;
2387 pandecode_log("};\n");
2388
2389 return sizeof(*t);
2390 }
2391
2392 static int
2393 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2394 const struct pandecode_mapped_memory *mem,
2395 mali_ptr payload, int job_no)
2396 {
2397 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2398
2399 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false);
2400
2401 pandecode_indices(v->prefix.indices, v->prefix.index_count, job_no);
2402
2403 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2404 pandecode_indent++;
2405
2406 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2407 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2408
2409 bool instanced = v->instance_shift || v->instance_odd;
2410 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2411
2412 pandecode_log(".prefix = ");
2413 pandecode_vertex_tiler_prefix(&v->prefix, job_no, !instanced && is_graphics);
2414
2415 pandecode_gl_enables(v->gl_enables, h->job_type);
2416
2417 if (v->instance_shift || v->instance_odd) {
2418 pandecode_prop("instance_shift = 0x%d /* %d */",
2419 v->instance_shift, 1 << v->instance_shift);
2420 pandecode_prop("instance_odd = 0x%X /* %d */",
2421 v->instance_odd, (2 * v->instance_odd) + 1);
2422
2423 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2424 }
2425
2426 if (v->offset_start)
2427 pandecode_prop("offset_start = %d", v->offset_start);
2428
2429 if (v->zero5) {
2430 pandecode_msg("XXX: midgard payload zero tripped\n");
2431 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2432 }
2433
2434 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2435
2436 pandecode_indent--;
2437 pandecode_log("};\n");
2438
2439 return sizeof(*v);
2440 }
2441
2442 static int
2443 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2444 mali_ptr payload, int job_no,
2445 bool is_bifrost)
2446 {
2447 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2448
2449 bool fbd_dumped = false;
2450
2451 if (!is_bifrost && (s->framebuffer & FBD_TYPE) == MALI_SFBD) {
2452 /* Only SFBDs are understood, not MFBDs. We're speculating,
2453 * based on the versioning, kernel code, etc, that the
2454 * difference is between Single FrameBuffer Descriptor and
2455 * Multiple FrmaeBuffer Descriptor; the change apparently lines
2456 * up with multi-framebuffer support being added (T7xx onwards,
2457 * including Gxx). In any event, there's some field shuffling
2458 * that we haven't looked into yet. */
2459
2460 pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true);
2461 fbd_dumped = true;
2462 } else if ((s->framebuffer & FBD_TYPE) == MALI_MFBD) {
2463 /* We don't know if Bifrost supports SFBD's at all, since the
2464 * driver never uses them. And the format is different from
2465 * Midgard anyways, due to the tiler heap and scratchpad being
2466 * moved out into separate structures, so it's not clear what a
2467 * Bifrost SFBD would even look like without getting an actual
2468 * trace, which appears impossible.
2469 */
2470
2471 pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2472 fbd_dumped = true;
2473 }
2474
2475 uintptr_t p = (uintptr_t) s->framebuffer & FBD_MASK;
2476 pandecode_log("struct mali_payload_fragment payload_%"PRIx64"_%d = {\n", payload, job_no);
2477 pandecode_indent++;
2478
2479 /* See the comments by the macro definitions for mathematical context
2480 * on why this is so weird */
2481
2482 if (MALI_TILE_COORD_FLAGS(s->max_tile_coord) || MALI_TILE_COORD_FLAGS(s->min_tile_coord))
2483 pandecode_msg("Tile coordinate flag missed, replay wrong\n");
2484
2485 pandecode_prop("min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(%d, %d)",
2486 MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT,
2487 MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT);
2488
2489 pandecode_prop("max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(%d, %d)",
2490 (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT,
2491 (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT);
2492
2493 /* If the FBD was just decoded, we can refer to it by pointer. If not,
2494 * we have to fallback on offsets. */
2495
2496 const char *fbd_type = s->framebuffer & MALI_MFBD ? "MALI_MFBD" : "MALI_SFBD";
2497
2498 /* TODO: Decode */
2499 unsigned extra_flags = (s->framebuffer & ~FBD_MASK) & ~MALI_MFBD;
2500
2501 if (fbd_dumped)
2502 pandecode_prop("framebuffer = framebuffer_%d_p | %s | 0x%X", job_no,
2503 fbd_type, extra_flags);
2504 else
2505 pandecode_prop("framebuffer = %s | %s | 0x%X", pointer_as_memory_reference(p),
2506 fbd_type, extra_flags);
2507
2508 pandecode_indent--;
2509 pandecode_log("};\n");
2510
2511 return sizeof(*s);
2512 }
2513
2514 static int job_descriptor_number = 0;
2515
2516 int
2517 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost)
2518 {
2519 struct mali_job_descriptor_header *h;
2520
2521 int start_number = 0;
2522
2523 bool first = true;
2524 bool last_size;
2525
2526 do {
2527 struct pandecode_mapped_memory *mem =
2528 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2529
2530 void *payload;
2531
2532 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2533
2534 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2535 * high 32-bits of the 64-bit pointer are reused to store
2536 * something else.
2537 */
2538 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2539 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2540 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2541
2542 payload = pandecode_fetch_gpu_mem(mem, payload_ptr,
2543 MALI_PAYLOAD_SIZE);
2544
2545 int job_no = job_descriptor_number++;
2546
2547 if (first)
2548 start_number = job_no;
2549
2550 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2551 pandecode_indent++;
2552
2553 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2554
2555 /* Save for next job fixing */
2556 last_size = h->job_descriptor_size;
2557
2558 if (h->job_descriptor_size)
2559 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2560
2561 if (h->exception_status && h->exception_status != 0x1)
2562 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2563 h->exception_status,
2564 (h->exception_status >> 16) & 0xFFFF,
2565 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2566 h->exception_status & 0xFF);
2567
2568 if (h->first_incomplete_task)
2569 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2570
2571 if (h->fault_pointer)
2572 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2573
2574 if (h->job_barrier)
2575 pandecode_prop("job_barrier = %d", h->job_barrier);
2576
2577 pandecode_prop("job_index = %d", h->job_index);
2578
2579 if (h->unknown_flags)
2580 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2581
2582 if (h->job_dependency_index_1)
2583 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2584
2585 if (h->job_dependency_index_2)
2586 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2587
2588 pandecode_indent--;
2589 pandecode_log("};\n");
2590
2591 /* Do not touch the field yet -- decode the payload first, and
2592 * don't touch that either. This is essential for the uploads
2593 * to occur in sequence and therefore be dynamically allocated
2594 * correctly. Do note the size, however, for that related
2595 * reason. */
2596
2597 switch (h->job_type) {
2598 case JOB_TYPE_SET_VALUE: {
2599 struct mali_payload_set_value *s = payload;
2600 pandecode_log("struct mali_payload_set_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2601 pandecode_indent++;
2602 MEMORY_PROP(s, out);
2603 pandecode_prop("unknown = 0x%" PRIX64, s->unknown);
2604 pandecode_indent--;
2605 pandecode_log("};\n");
2606
2607 break;
2608 }
2609
2610 case JOB_TYPE_TILER:
2611 case JOB_TYPE_VERTEX:
2612 case JOB_TYPE_COMPUTE:
2613 if (bifrost) {
2614 if (h->job_type == JOB_TYPE_TILER)
2615 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no);
2616 else
2617 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no);
2618 } else
2619 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no);
2620
2621 break;
2622
2623 case JOB_TYPE_FRAGMENT:
2624 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost);
2625 break;
2626
2627 default:
2628 break;
2629 }
2630
2631 /* Handle linkage */
2632
2633 if (!first) {
2634 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2635
2636 if (last_size)
2637 pandecode_log_cont("next_job_64 = job_%d_p;\n\n", job_no);
2638 else
2639 pandecode_log_cont("next_job_32 = (u32) (uintptr_t) job_%d_p;\n\n", job_no);
2640 }
2641
2642 first = false;
2643
2644 } while ((jc_gpu_va = h->job_descriptor_size ? h->next_job_64 : h->next_job_32));
2645
2646 return start_number;
2647 }