projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/midgard: Support disassembling to a file
[mesa.git] / src / panfrost / pandecode / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <panfrost-job.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "pan_pretty_print.h"
38 #include "midgard/disassemble.h"
39 #include "bifrost/disassemble.h"
40
41 #include "pan_encoder.h"
42
43 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
44
45 #define MEMORY_PROP(obj, p) {\
46 if (obj->p) { \
47 char *a = pointer_as_memory_reference(obj->p); \
48 pandecode_prop("%s = %s", #p, a); \
49 free(a); \
50 } \
51 }
52
53 #define MEMORY_PROP_DIR(obj, p) {\
54 if (obj.p) { \
55 char *a = pointer_as_memory_reference(obj.p); \
56 pandecode_prop("%s = %s", #p, a); \
57 free(a); \
58 } \
59 }
60
61 /* Semantic logging type.
62 *
63 * Raw: for raw messages to be printed as is.
64 * Message: for helpful information to be commented out in replays.
65 * Property: for properties of a struct
66 *
67 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
68 */
69
70 enum pandecode_log_type {
71 PANDECODE_RAW,
72 PANDECODE_MESSAGE,
73 PANDECODE_PROPERTY
74 };
75
76 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
77 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
78 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
79
80 unsigned pandecode_indent = 0;
81
82 static void
83 pandecode_make_indent(void)
84 {
85 for (unsigned i = 0; i < pandecode_indent; ++i)
86 printf(" ");
87 }
88
89 static void
90 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
91 {
92 va_list ap;
93
94 pandecode_make_indent();
95
96 if (type == PANDECODE_MESSAGE)
97 printf("// ");
98 else if (type == PANDECODE_PROPERTY)
99 printf(".");
100
101 va_start(ap, format);
102 vprintf(format, ap);
103 va_end(ap);
104
105 if (type == PANDECODE_PROPERTY)
106 printf(",\n");
107 }
108
109 static void
110 pandecode_log_cont(const char *format, ...)
111 {
112 va_list ap;
113
114 va_start(ap, format);
115 vprintf(format, ap);
116 va_end(ap);
117 }
118
119 /* To check for memory safety issues, validates that the given pointer in GPU
120 * memory is valid, containing at least sz bytes. The goal is to eliminate
121 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
122 * overruns) by statically validating pointers.
123 */
124
125 static void
126 pandecode_validate_buffer(mali_ptr addr, size_t sz)
127 {
128 if (!addr) {
129 pandecode_msg("XXX: null pointer deref");
130 return;
131 }
132
133 /* Find a BO */
134
135 struct pandecode_mapped_memory *bo =
136 pandecode_find_mapped_gpu_mem_containing(addr);
137
138 if (!bo) {
139 pandecode_msg("XXX: invalid memory dereference\n");
140 return;
141 }
142
143 /* Bounds check */
144
145 unsigned offset = addr - bo->gpu_va;
146 unsigned total = offset + sz;
147
148 if (total > bo->length) {
149 pandecode_msg("XXX: buffer overrun. "
150 "Chunk of size %zu at offset %d in buffer of size %zu. "
151 "Overrun by %zu bytes. \n",
152 sz, offset, bo->length, total - bo->length);
153 return;
154 }
155 }
156
157 struct pandecode_flag_info {
158 u64 flag;
159 const char *name;
160 };
161
162 static void
163 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
164 u64 flags)
165 {
166 bool decodable_flags_found = false;
167
168 for (int i = 0; flag_info[i].name; i++) {
169 if ((flags & flag_info[i].flag) != flag_info[i].flag)
170 continue;
171
172 if (!decodable_flags_found) {
173 decodable_flags_found = true;
174 } else {
175 pandecode_log_cont(" | ");
176 }
177
178 pandecode_log_cont("%s", flag_info[i].name);
179
180 flags &= ~flag_info[i].flag;
181 }
182
183 if (decodable_flags_found) {
184 if (flags)
185 pandecode_log_cont(" | 0x%" PRIx64, flags);
186 } else {
187 pandecode_log_cont("0x%" PRIx64, flags);
188 }
189 }
190
191 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
192 static const struct pandecode_flag_info gl_enable_flag_info[] = {
193 FLAG_INFO(OCCLUSION_QUERY),
194 FLAG_INFO(OCCLUSION_PRECISE),
195 FLAG_INFO(FRONT_CCW_TOP),
196 FLAG_INFO(CULL_FACE_FRONT),
197 FLAG_INFO(CULL_FACE_BACK),
198 {}
199 };
200 #undef FLAG_INFO
201
202 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
203 static const struct pandecode_flag_info clear_flag_info[] = {
204 FLAG_INFO(FAST),
205 FLAG_INFO(SLOW),
206 FLAG_INFO(SLOW_STENCIL),
207 {}
208 };
209 #undef FLAG_INFO
210
211 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
212 static const struct pandecode_flag_info mask_flag_info[] = {
213 FLAG_INFO(R),
214 FLAG_INFO(G),
215 FLAG_INFO(B),
216 FLAG_INFO(A),
217 {}
218 };
219 #undef FLAG_INFO
220
221 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
222 static const struct pandecode_flag_info u3_flag_info[] = {
223 FLAG_INFO(HAS_MSAA),
224 FLAG_INFO(CAN_DISCARD),
225 FLAG_INFO(HAS_BLEND_SHADER),
226 FLAG_INFO(DEPTH_WRITEMASK),
227 {}
228 };
229
230 static const struct pandecode_flag_info u4_flag_info[] = {
231 FLAG_INFO(NO_MSAA),
232 FLAG_INFO(NO_DITHER),
233 FLAG_INFO(DEPTH_RANGE_A),
234 FLAG_INFO(DEPTH_RANGE_B),
235 FLAG_INFO(STENCIL_TEST),
236 FLAG_INFO(SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
242 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
243 FLAG_INFO(MSAA),
244 FLAG_INFO(SRGB),
245 {}
246 };
247 #undef FLAG_INFO
248
249 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
250 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
251 FLAG_INFO(PRESENT),
252 {}
253 };
254 #undef FLAG_INFO
255
256 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
257 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
258 FLAG_INFO(ZS),
259 {}
260 };
261 #undef FLAG_INFO
262
263 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
264 static const struct pandecode_flag_info shader_midgard1_flag_info [] = {
265 FLAG_INFO(EARLY_Z),
266 FLAG_INFO(READS_TILEBUFFER),
267 FLAG_INFO(READS_ZS),
268 {}
269 };
270 #undef FLAG_INFO
271
272 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
273 static const struct pandecode_flag_info mfbd_flag_info [] = {
274 FLAG_INFO(DEPTH_WRITE),
275 FLAG_INFO(EXTRA),
276 {}
277 };
278 #undef FLAG_INFO
279
280 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
281 static const struct pandecode_flag_info sampler_flag_info [] = {
282 FLAG_INFO(MAG_NEAREST),
283 FLAG_INFO(MIN_NEAREST),
284 FLAG_INFO(MIP_LINEAR_1),
285 FLAG_INFO(MIP_LINEAR_2),
286 FLAG_INFO(NORM_COORDS),
287 {}
288 };
289 #undef FLAG_INFO
290
291 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
292 static const struct pandecode_flag_info sfbd_unk1_info [] = {
293 FLAG_INFO(MSAA_8),
294 FLAG_INFO(MSAA_A),
295 {}
296 };
297 #undef FLAG_INFO
298
299 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
300 static const struct pandecode_flag_info sfbd_unk2_info [] = {
301 FLAG_INFO(MSAA_B),
302 FLAG_INFO(SRGB),
303 {}
304 };
305 #undef FLAG_INFO
306
307 extern char *replace_fragment;
308 extern char *replace_vertex;
309
310 static char *
311 pandecode_job_type(enum mali_job_type type)
312 {
313 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
314
315 switch (type) {
316 DEFINE_CASE(NULL);
317 DEFINE_CASE(WRITE_VALUE);
318 DEFINE_CASE(CACHE_FLUSH);
319 DEFINE_CASE(COMPUTE);
320 DEFINE_CASE(VERTEX);
321 DEFINE_CASE(TILER);
322 DEFINE_CASE(FUSED);
323 DEFINE_CASE(FRAGMENT);
324
325 case JOB_NOT_STARTED:
326 return "NOT_STARTED";
327
328 default:
329 pandecode_log("Warning! Unknown job type %x\n", type);
330 return "!?!?!?";
331 }
332
333 #undef DEFINE_CASE
334 }
335
336 static char *
337 pandecode_draw_mode(enum mali_draw_mode mode)
338 {
339 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
340
341 switch (mode) {
342 DEFINE_CASE(DRAW_NONE);
343 DEFINE_CASE(POINTS);
344 DEFINE_CASE(LINES);
345 DEFINE_CASE(TRIANGLES);
346 DEFINE_CASE(TRIANGLE_STRIP);
347 DEFINE_CASE(TRIANGLE_FAN);
348 DEFINE_CASE(LINE_STRIP);
349 DEFINE_CASE(LINE_LOOP);
350 DEFINE_CASE(POLYGON);
351 DEFINE_CASE(QUADS);
352 DEFINE_CASE(QUAD_STRIP);
353
354 default:
355 pandecode_msg("XXX: invalid draw mode %X\n", mode);
356 return "";
357 }
358
359 #undef DEFINE_CASE
360 }
361
362 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
363 static char *
364 pandecode_func(enum mali_func mode)
365 {
366 switch (mode) {
367 DEFINE_CASE(NEVER);
368 DEFINE_CASE(LESS);
369 DEFINE_CASE(EQUAL);
370 DEFINE_CASE(LEQUAL);
371 DEFINE_CASE(GREATER);
372 DEFINE_CASE(NOTEQUAL);
373 DEFINE_CASE(GEQUAL);
374 DEFINE_CASE(ALWAYS);
375
376 default:
377 pandecode_msg("XXX: invalid func %X\n", mode);
378 return "";
379 }
380 }
381 #undef DEFINE_CASE
382
383 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
384 static char *
385 pandecode_stencil_op(enum mali_stencil_op op)
386 {
387 switch (op) {
388 DEFINE_CASE(KEEP);
389 DEFINE_CASE(REPLACE);
390 DEFINE_CASE(ZERO);
391 DEFINE_CASE(INVERT);
392 DEFINE_CASE(INCR_WRAP);
393 DEFINE_CASE(DECR_WRAP);
394 DEFINE_CASE(INCR);
395 DEFINE_CASE(DECR);
396
397 default:
398 pandecode_msg("XXX: invalid stencil op %X\n", op);
399 return "";
400 }
401 }
402
403 #undef DEFINE_CASE
404
405 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
406 {
407 switch(mode) {
408 /* TODO: Combine to just "instanced" once this can be done
409 * unambiguously in all known cases */
410 case MALI_ATTR_POT_DIVIDE:
411 return "instanced_pot";
412 case MALI_ATTR_MODULO:
413 return "instanced_mod";
414 case MALI_ATTR_NPOT_DIVIDE:
415 return "instanced_npot";
416 case MALI_ATTR_IMAGE:
417 return "image";
418 default:
419 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
420 return "";
421 }
422 }
423
424 static const char *
425 pandecode_special_record(uint64_t v, bool* attribute)
426 {
427 switch(v) {
428 case MALI_ATTR_VERTEXID:
429 *attribute = true;
430 return "gl_VertexID";
431 case MALI_ATTR_INSTANCEID:
432 *attribute = true;
433 return "gl_InstanceID";
434 case MALI_VARYING_FRAG_COORD:
435 return "gl_FragCoord";
436 case MALI_VARYING_FRONT_FACING:
437 return "gl_FrontFacing";
438 case MALI_VARYING_POINT_COORD:
439 return "gl_PointCoord";
440 default:
441 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
442 return "";
443 }
444 }
445
446 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
447 static char *
448 pandecode_wrap_mode(enum mali_wrap_mode op)
449 {
450 switch (op) {
451 DEFINE_CASE(REPEAT);
452 DEFINE_CASE(CLAMP_TO_EDGE);
453 DEFINE_CASE(CLAMP_TO_BORDER);
454 DEFINE_CASE(MIRRORED_REPEAT);
455
456 default:
457 pandecode_msg("XXX: invalid wrap mode %X\n", op);
458 return "";
459 }
460 }
461 #undef DEFINE_CASE
462
463 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
464 static char *
465 pandecode_block_format(enum mali_block_format fmt)
466 {
467 switch (fmt) {
468 DEFINE_CASE(TILED);
469 DEFINE_CASE(UNKNOWN);
470 DEFINE_CASE(LINEAR);
471 DEFINE_CASE(AFBC);
472
473 default:
474 unreachable("Invalid case");
475 }
476 }
477 #undef DEFINE_CASE
478
479 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
480 char *
481 pandecode_exception_access(enum mali_exception_access access)
482 {
483 switch (access) {
484 DEFINE_CASE(NONE);
485 DEFINE_CASE(EXECUTE);
486 DEFINE_CASE(READ);
487 DEFINE_CASE(WRITE);
488
489 default:
490 unreachable("Invalid case");
491 }
492 }
493 #undef DEFINE_CASE
494
495 /* Midgard's tiler descriptor is embedded within the
496 * larger FBD */
497
498 static void
499 pandecode_midgard_tiler_descriptor(
500 const struct midgard_tiler_descriptor *t,
501 unsigned width,
502 unsigned height,
503 bool is_fragment,
504 bool has_hierarchy)
505 {
506 pandecode_log(".tiler = {\n");
507 pandecode_indent++;
508
509 if (t->hierarchy_mask == MALI_TILER_DISABLED)
510 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
511 else
512 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
513
514 /* We know this name from the kernel, but we never see it nonzero */
515
516 if (t->flags)
517 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
518
519 MEMORY_PROP(t, polygon_list);
520
521 /* The body is offset from the base of the polygon list */
522 assert(t->polygon_list_body > t->polygon_list);
523 unsigned body_offset = t->polygon_list_body - t->polygon_list;
524
525 /* It needs to fit inside the reported size */
526 assert(t->polygon_list_size >= body_offset);
527
528 /* Check that we fit */
529 struct pandecode_mapped_memory *plist =
530 pandecode_find_mapped_gpu_mem_containing(t->polygon_list);
531
532 assert(t->polygon_list_size <= plist->length);
533
534 /* Now that we've sanity checked, we'll try to calculate the sizes
535 * ourselves for comparison */
536
537 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
538 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
539
540 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
541 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
542 ref_header, ref_size);
543 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
544 pandecode_msg("body offset %d\n", body_offset);
545 }
546
547 /* The tiler heap has a start and end specified -- it should be
548 * identical to what we have in the BO. The exception is if tiling is
549 * disabled. */
550
551 MEMORY_PROP(t, heap_start);
552 assert(t->heap_end >= t->heap_start);
553
554 struct pandecode_mapped_memory *heap =
555 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
556
557 unsigned heap_size = t->heap_end - t->heap_start;
558
559 /* Tiling is enabled with a special flag */
560 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
561 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
562
563 bool tiling_enabled = hierarchy_mask;
564
565 if (tiling_enabled) {
566 /* When tiling is enabled, the heap should be a tight fit */
567 unsigned heap_offset = t->heap_start - heap->gpu_va;
568 if ((heap_offset + heap_size) != heap->length) {
569 pandecode_msg("XXX: heap size %u (expected %zu)\n",
570 heap_size, heap->length - heap_offset);
571 }
572
573 /* We should also have no other flags */
574 if (tiler_flags)
575 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
576 } else {
577 /* When tiling is disabled, we should have that flag and no others */
578
579 if (tiler_flags != MALI_TILER_DISABLED) {
580 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
581 tiler_flags);
582 }
583
584 /* We should also have an empty heap */
585 if (heap_size) {
586 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
587 heap_size);
588 }
589
590 /* Disabled tiling is used only for clear-only jobs, which are
591 * purely FRAGMENT, so we should never see this for
592 * non-FRAGMENT descriptors. */
593
594 if (!is_fragment)
595 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
596 }
597
598 /* We've never seen weights used in practice, but we know from the
599 * kernel these fields is there */
600
601 bool nonzero_weights = false;
602
603 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
604 nonzero_weights |= t->weights[w] != 0x0;
605 }
606
607 if (nonzero_weights) {
608 pandecode_log(".weights = {");
609
610 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
611 pandecode_log("%d, ", t->weights[w]);
612 }
613
614 pandecode_log("},");
615 }
616
617 pandecode_indent--;
618 pandecode_log("}\n");
619 }
620
621 /* Information about the framebuffer passed back for
622 * additional analysis */
623
624 struct pandecode_fbd {
625 unsigned width;
626 unsigned height;
627 unsigned rt_count;
628 bool has_extra;
629 };
630
631 static void
632 pandecode_sfbd_format(struct mali_sfbd_format format)
633 {
634 pandecode_log(".format = {\n");
635 pandecode_indent++;
636
637 pandecode_log(".unk1 = ");
638 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
639 pandecode_log_cont(",\n");
640
641 /* TODO: Map formats so we can check swizzles and print nicely */
642 pandecode_log("swizzle");
643 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
644 pandecode_log_cont(",\n");
645
646 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
647 (format.nr_channels + 1));
648
649 pandecode_log(".unk2 = ");
650 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
651 pandecode_log_cont(",\n");
652
653 pandecode_prop("block = %s", pandecode_block_format(format.block));
654
655 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
656
657 pandecode_indent--;
658 pandecode_log("},\n");
659 }
660
661 static struct pandecode_fbd
662 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
663 {
664 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
665 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
666
667 struct pandecode_fbd info = {
668 .has_extra = false,
669 .rt_count = 1
670 };
671
672 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
673 pandecode_indent++;
674
675 pandecode_prop("unknown1 = 0x%" PRIx32, s->unknown1);
676 pandecode_prop("unknown2 = 0x%" PRIx32, s->unknown2);
677
678 pandecode_sfbd_format(s->format);
679
680 info.width = s->width + 1;
681 info.height = s->height + 1;
682
683 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
684 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
685
686 MEMORY_PROP(s, checksum);
687
688 if (s->checksum_stride)
689 pandecode_prop("checksum_stride = %d", s->checksum_stride);
690
691 MEMORY_PROP(s, framebuffer);
692 pandecode_prop("stride = %d", s->stride);
693
694 /* Earlier in the actual commandstream -- right before width -- but we
695 * delay to flow nicer */
696
697 pandecode_log(".clear_flags = ");
698 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
699 pandecode_log_cont(",\n");
700
701 if (s->depth_buffer) {
702 MEMORY_PROP(s, depth_buffer);
703 pandecode_prop("depth_stride = %d", s->depth_stride);
704 }
705
706 if (s->stencil_buffer) {
707 MEMORY_PROP(s, stencil_buffer);
708 pandecode_prop("stencil_stride = %d", s->stencil_stride);
709 }
710
711 if (s->depth_stride_zero ||
712 s->stencil_stride_zero ||
713 s->zero7 || s->zero8) {
714 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
715 pandecode_prop("depth_stride_zero = 0x%x",
716 s->depth_stride_zero);
717 pandecode_prop("stencil_stride_zero = 0x%x",
718 s->stencil_stride_zero);
719 pandecode_prop("zero7 = 0x%" PRIx32,
720 s->zero7);
721 pandecode_prop("zero8 = 0x%" PRIx32,
722 s->zero8);
723 }
724
725 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
726 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
727 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
728 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
729 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
730 }
731
732 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
733 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
734 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
735 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
736 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
737 }
738
739 if (s->clear_stencil) {
740 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
741 }
742
743 MEMORY_PROP(s, scratchpad);
744 const struct midgard_tiler_descriptor t = s->tiler;
745
746 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
747 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
748
749 pandecode_indent--;
750 pandecode_log("};\n");
751
752 pandecode_prop("zero0 = 0x%" PRIx64, s->zero0);
753 pandecode_prop("zero1 = 0x%" PRIx64, s->zero1);
754 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
755 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
756 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
757
758 printf(".zero3 = {");
759
760 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
761 printf("%X, ", s->zero3[i]);
762
763 printf("},\n");
764
765 printf(".zero6 = {");
766
767 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
768 printf("%X, ", s->zero6[i]);
769
770 printf("},\n");
771
772 return info;
773 }
774
775 static void
776 pandecode_u32_slide(unsigned name, const u32 *slide, unsigned count)
777 {
778 pandecode_log(".unknown%d = {", name);
779
780 for (int i = 0; i < count; ++i)
781 printf("%X, ", slide[i]);
782
783 pandecode_log("},\n");
784 }
785
786 #define SHORT_SLIDE(num) \
787 pandecode_u32_slide(num, s->unknown ## num, ARRAY_SIZE(s->unknown ## num))
788
789 static void
790 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
791 {
792 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
793 const struct mali_compute_fbd *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
794
795 pandecode_log("struct mali_compute_fbd framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
796 pandecode_indent++;
797
798 SHORT_SLIDE(1);
799
800 pandecode_indent--;
801 printf("},\n");
802 }
803
804 /* Extracts the number of components associated with a Mali format */
805
806 static unsigned
807 pandecode_format_component_count(enum mali_format fmt)
808 {
809 /* Mask out the format class */
810 unsigned top = fmt & 0b11100000;
811
812 switch (top) {
813 case MALI_FORMAT_SNORM:
814 case MALI_FORMAT_UINT:
815 case MALI_FORMAT_UNORM:
816 case MALI_FORMAT_SINT:
817 return ((fmt >> 3) & 3) + 1;
818 default:
819 /* TODO: Validate */
820 return 4;
821 }
822 }
823
824 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
825
826 static unsigned
827 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
828 {
829 unsigned mask = 0;
830 assert(MALI_CHANNEL_RED == 0);
831
832 for (unsigned c = 0; c < 4; ++c) {
833 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
834
835 if (chan <= MALI_CHANNEL_ALPHA)
836 mask |= (1 << chan);
837 }
838
839 return mask;
840 }
841
842 /* Validates that a (format, swizzle) pair is valid, in the sense that the
843 * swizzle doesn't access any components that are undefined in the format.
844 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
845 * omitted */
846
847 static bool
848 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
849 {
850 unsigned nr_comp = pandecode_format_component_count(fmt);
851 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
852 unsigned valid_mask = (1 << nr_comp) - 1;
853 unsigned invalid_mask = ~valid_mask;
854
855 if (access_mask & invalid_mask) {
856 pandecode_msg("XXX: invalid components accessed\n");
857 return false;
858 }
859
860 /* Check for the default non-swizzling swizzle so we can suppress
861 * useless printing for the defaults */
862
863 unsigned default_swizzles[4] = {
864 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
865 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
866 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
867 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
868 };
869
870 return (swizzle == default_swizzles[nr_comp - 1]);
871 }
872
873 /* Maps MALI_RGBA32F to rgba32f, etc */
874
875 static void
876 pandecode_format_short(enum mali_format fmt, bool srgb)
877 {
878 /* We want a type-like format, so cut off the initial MALI_ */
879 char *format = pandecode_format(fmt);
880 format += strlen("MALI_");
881
882 unsigned len = strlen(format);
883 char *lower_format = calloc(1, len + 1);
884
885 for (unsigned i = 0; i < len; ++i)
886 lower_format[i] = tolower(format[i]);
887
888 /* Sanity check sRGB flag is applied to RGB, per the name */
889 if (srgb && lower_format[0] != 'r')
890 pandecode_msg("XXX: sRGB applied to non-colour format\n");
891
892 /* Just prefix with an s, so you get formats like srgba8_unorm */
893 if (srgb)
894 pandecode_log_cont("s");
895
896 pandecode_log_cont("%s", lower_format);
897 free(lower_format);
898 }
899
900 static void
901 pandecode_swizzle(unsigned swizzle, enum mali_format format)
902 {
903 /* First, do some validation */
904 bool trivial_swizzle = pandecode_validate_format_swizzle(
905 format, swizzle);
906
907 if (trivial_swizzle)
908 return;
909
910 /* Next, print the swizzle */
911 pandecode_log_cont(".");
912
913 static const char components[] = "rgba01";
914
915 for (unsigned c = 0; c < 4; ++c) {
916 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
917
918 if (chan >= MALI_CHANNEL_RESERVED_0) {
919 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
920 continue;
921 }
922 pandecode_log_cont("%c", components[chan]);
923 }
924 }
925
926 static void
927 pandecode_rt_format(struct mali_rt_format format)
928 {
929 pandecode_log(".format = {\n");
930 pandecode_indent++;
931
932 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
933 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
934 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
935
936 pandecode_prop("block = %s", pandecode_block_format(format.block));
937
938 /* TODO: Map formats so we can check swizzles and print nicely */
939 pandecode_log("swizzle");
940 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
941 pandecode_log_cont(",\n");
942
943 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
944 (format.nr_channels + 1));
945
946 pandecode_log(".flags = ");
947 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
948 pandecode_log_cont(",\n");
949
950 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
951 * which is a faster hardware-based alternative to the wallpaper method
952 * to preserve framebuffer contents across frames. In practice, MFBD
953 * preload is buggy on Midgard, and so this is a chicken bit. If this
954 * bit isn't set, most likely something broke unrelated to preload */
955
956 if (!format.no_preload) {
957 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
958 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
959 }
960
961 if (format.zero)
962 pandecode_prop("zero = 0x%" PRIx32, format.zero);
963
964 pandecode_indent--;
965 pandecode_log("},\n");
966 }
967
968 static void
969 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct bifrost_framebuffer *fb)
970 {
971 pandecode_log("struct bifrost_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
972 pandecode_indent++;
973
974 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
975 mali_ptr rt_va = gpu_va + i * sizeof(struct bifrost_render_target);
976 struct pandecode_mapped_memory *mem =
977 pandecode_find_mapped_gpu_mem_containing(rt_va);
978 const struct bifrost_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
979
980 pandecode_log("{\n");
981 pandecode_indent++;
982
983 pandecode_rt_format(rt->format);
984
985 if (rt->format.block == MALI_BLOCK_AFBC) {
986 pandecode_log(".afbc = {\n");
987 pandecode_indent++;
988
989 char *a = pointer_as_memory_reference(rt->afbc.metadata);
990 pandecode_prop("metadata = %s", a);
991 free(a);
992
993 pandecode_prop("stride = %d", rt->afbc.stride);
994 pandecode_prop("unk = 0x%" PRIx32, rt->afbc.unk);
995
996 pandecode_indent--;
997 pandecode_log("},\n");
998 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.unk) {
999 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1000 rt->afbc.metadata,
1001 rt->afbc.stride,
1002 rt->afbc.unk);
1003 }
1004
1005 MEMORY_PROP(rt, framebuffer);
1006 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1007
1008 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1009 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1010 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1011 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1012 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1013 }
1014
1015 if (rt->zero1 || rt->zero2 || rt->zero3) {
1016 pandecode_msg("XXX: render target zeros tripped\n");
1017 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1018 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1019 pandecode_prop("zero3 = 0x%" PRIx32, rt->zero3);
1020 }
1021
1022 pandecode_indent--;
1023 pandecode_log("},\n");
1024 }
1025
1026 pandecode_indent--;
1027 pandecode_log("};\n");
1028 }
1029
1030 static struct pandecode_fbd
1031 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment)
1032 {
1033 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1034 const struct bifrost_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1035
1036 struct pandecode_fbd info;
1037
1038 if (fb->sample_locations) {
1039 /* The blob stores all possible sample locations in a single buffer
1040 * allocated on startup, and just switches the pointer when switching
1041 * MSAA state. For now, we just put the data into the cmdstream, but we
1042 * should do something like what the blob does with a real driver.
1043 *
1044 * There seem to be 32 slots for sample locations, followed by another
1045 * 16. The second 16 is just the center location followed by 15 zeros
1046 * in all the cases I've identified (maybe shader vs. depth/color
1047 * samples?).
1048 */
1049
1050 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->sample_locations);
1051
1052 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->sample_locations);
1053
1054 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1055 pandecode_indent++;
1056
1057 for (int i = 0; i < 32 + 16; i++) {
1058 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1059 }
1060
1061 pandecode_indent--;
1062 pandecode_log("};\n");
1063 }
1064
1065 pandecode_log("struct bifrost_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1066 pandecode_indent++;
1067
1068 pandecode_prop("stack_shift = 0x%x", fb->stack_shift);
1069 pandecode_prop("unk0 = 0x%x", fb->unk0);
1070
1071 if (fb->sample_locations)
1072 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1073
1074 /* Assume that unknown1 was emitted in the last job for
1075 * now */
1076 MEMORY_PROP(fb, unknown1);
1077
1078 info.width = fb->width1 + 1;
1079 info.height = fb->height1 + 1;
1080 info.rt_count = fb->rt_count_1 + 1;
1081
1082 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1083 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1084 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1085 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1086
1087 pandecode_prop("unk1 = 0x%x", fb->unk1);
1088 pandecode_prop("unk2 = 0x%x", fb->unk2);
1089 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1090 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1091
1092 pandecode_log(".mfbd_flags = ");
1093 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1094 pandecode_log_cont(",\n");
1095
1096 if (fb->clear_stencil)
1097 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1098
1099 if (fb->clear_depth)
1100 pandecode_prop("clear_depth = %f", fb->clear_depth);
1101
1102 /* TODO: What is this? Let's not blow up.. */
1103 if (fb->unknown2 != 0x1F)
1104 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1105
1106 pandecode_prop("unknown2 = 0x%x", fb->unknown2);
1107 MEMORY_PROP(fb, scratchpad);
1108 const struct midgard_tiler_descriptor t = fb->tiler;
1109 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1110
1111 if (fb->zero3 || fb->zero4) {
1112 pandecode_msg("XXX: framebuffer zeros tripped\n");
1113 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1114 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1115 }
1116
1117 pandecode_indent--;
1118 pandecode_log("};\n");
1119
1120 gpu_va += sizeof(struct bifrost_framebuffer);
1121
1122 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1123
1124 if (info.has_extra) {
1125 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1126 const struct bifrost_fb_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1127
1128 pandecode_log("struct bifrost_fb_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1129 pandecode_indent++;
1130
1131 MEMORY_PROP(fbx, checksum);
1132
1133 if (fbx->checksum_stride)
1134 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1135
1136 pandecode_log(".flags_hi = ");
1137 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_lo);
1138 pandecode_log_cont(",\n");
1139
1140 pandecode_log(".flags_lo = ");
1141 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1142 pandecode_log_cont(",\n");
1143
1144 pandecode_prop("zs_block = %s\n", pandecode_block_format(fbx->zs_block));
1145
1146 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1147 pandecode_log(".ds_afbc = {\n");
1148 pandecode_indent++;
1149
1150 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1151 pandecode_prop("depth_stencil_afbc_stride = %d",
1152 fbx->ds_afbc.depth_stencil_afbc_stride);
1153 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1154
1155 if (fbx->ds_afbc.zero1 || fbx->ds_afbc.padding) {
1156 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1157 pandecode_prop("zero1 = 0x%" PRIx32,
1158 fbx->ds_afbc.zero1);
1159 pandecode_prop("padding = 0x%" PRIx64,
1160 fbx->ds_afbc.padding);
1161 }
1162
1163 pandecode_indent--;
1164 pandecode_log("},\n");
1165 } else {
1166 pandecode_log(".ds_linear = {\n");
1167 pandecode_indent++;
1168
1169 if (fbx->ds_linear.depth) {
1170 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1171 pandecode_prop("depth_stride = %d",
1172 fbx->ds_linear.depth_stride);
1173 } else if (fbx->ds_linear.depth_stride) {
1174 pandecode_msg("XXX: depth stride zero tripped %d\n", fbx->ds_linear.depth_stride);
1175 }
1176
1177 if (fbx->ds_linear.stencil) {
1178 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1179 pandecode_prop("stencil_stride = %d",
1180 fbx->ds_linear.stencil_stride);
1181 } else if (fbx->ds_linear.stencil_stride) {
1182 pandecode_msg("XXX: stencil stride zero tripped %d\n", fbx->ds_linear.stencil_stride);
1183 }
1184
1185 if (fbx->ds_linear.depth_stride_zero ||
1186 fbx->ds_linear.stencil_stride_zero ||
1187 fbx->ds_linear.zero1 || fbx->ds_linear.zero2) {
1188 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1189 pandecode_prop("depth_stride_zero = 0x%x",
1190 fbx->ds_linear.depth_stride_zero);
1191 pandecode_prop("stencil_stride_zero = 0x%x",
1192 fbx->ds_linear.stencil_stride_zero);
1193 pandecode_prop("zero1 = 0x%" PRIx32,
1194 fbx->ds_linear.zero1);
1195 pandecode_prop("zero2 = 0x%" PRIx32,
1196 fbx->ds_linear.zero2);
1197 }
1198
1199 pandecode_indent--;
1200 pandecode_log("},\n");
1201 }
1202
1203 if (fbx->zero3 || fbx->zero4) {
1204 pandecode_msg("XXX: fb_extra zeros tripped\n");
1205 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1206 pandecode_prop("zero4 = 0x%" PRIx64, fbx->zero4);
1207 }
1208
1209 pandecode_indent--;
1210 pandecode_log("};\n");
1211
1212 gpu_va += sizeof(struct bifrost_fb_extra);
1213 }
1214
1215 if (is_fragment)
1216 pandecode_render_target(gpu_va, job_no, fb);
1217
1218 return info;
1219 }
1220
1221 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1222 * count */
1223
1224 static void
1225 pandecode_padded_vertices(unsigned shift, unsigned k)
1226 {
1227 unsigned odd = 2*k + 1;
1228 unsigned pot = 1 << shift;
1229 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1230 }
1231
1232 /* Given a magic divisor, recover what we were trying to divide by.
1233 *
1234 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1235 * 0 <= m < 2^N, for N bits in m.
1236 *
1237 * Let q represent the number we would like to divide by.
1238 *
1239 * By definition of a magic divisor for N-bit unsigned integers (a number you
1240 * multiply by to magically get division), m is a number such that:
1241 *
1242 * (m * x) & (2^N - 1) = floor(x/q).
1243 * for all x on Z where 0 <= x < 2^N
1244 *
1245 * Ignore the case where any of the above values equals zero; it is irrelevant
1246 * for our purposes (instanced arrays).
1247 *
1248 * Choose x = q. Then:
1249 *
1250 * (m * x) & (2^N - 1) = floor(x/q).
1251 * (m * q) & (2^N - 1) = floor(q/q).
1252 *
1253 * floor(q/q) = floor(1) = 1, therefore:
1254 *
1255 * (m * q) & (2^N - 1) = 1
1256 *
1257 * Recall the identity that the bitwise AND of one less than a power-of-two
1258 * equals the modulo with that power of two, i.e. for all x:
1259 *
1260 * x & (2^N - 1) = x % N
1261 *
1262 * Therefore:
1263 *
1264 * mq % (2^N) = 1
1265 *
1266 * By definition, a modular multiplicative inverse of a number m is the number
1267 * q such that with respect to a modulos M:
1268 *
1269 * mq % M = 1
1270 *
1271 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1272 *
1273 */
1274
1275 static void
1276 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1277 {
1278 #if 0
1279 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1280 * shift) the most lame way possible... just repeatedly add.
1281 * Asymptoptically slow but nobody cares in practice, unless you have
1282 * massive numbers of vertices or high divisors. */
1283
1284 unsigned inverse = 0;
1285
1286 /* Magic implicitly has the highest bit set */
1287 magic |= (1 << 31);
1288
1289 /* Depending on rounding direction */
1290 if (extra)
1291 magic++;
1292
1293 for (;;) {
1294 uint32_t product = magic * inverse;
1295
1296 if (shift) {
1297 product >>= shift;
1298 }
1299
1300 if (product == 1)
1301 break;
1302
1303 ++inverse;
1304 }
1305
1306 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1307
1308 /* Recall we're supposed to divide by (gl_level_divisor *
1309 * padded_num_vertices) */
1310
1311 unsigned padded_num_vertices = inverse / orig_divisor;
1312
1313 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1314 #endif
1315 }
1316
1317 static void
1318 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1319 mali_ptr addr, int job_no, char *suffix,
1320 int count, bool varying, enum mali_job_type job_type)
1321 {
1322 char *prefix = varying ? "varying" : "attribute";
1323 assert(addr);
1324
1325 if (!count) {
1326 pandecode_msg("warn: No %s records\n", prefix);
1327 return;
1328 }
1329
1330 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1331
1332 for (int i = 0; i < count; ++i) {
1333 /* First, check for special records */
1334 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1335 if (attr[i].size)
1336 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1337
1338 if (attr[i].stride) {
1339 /* gl_InstanceID passes a magic divisor in the
1340 * stride field to divide by the padded vertex
1341 * count. No other records should do so, so
1342 * stride should otherwise be zero. Note that
1343 * stride in the usual attribute sense doesn't
1344 * apply to special records. */
1345
1346 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1347
1348 pandecode_log_cont("/* %smagic divisor = %X */ ",
1349 has_divisor ? "" : "XXX: ", attr[i].stride);
1350 }
1351
1352 if (attr[i].shift || attr[i].extra_flags) {
1353 /* Attributes use these fields for
1354 * instancing/padding/etc type issues, but
1355 * varyings don't */
1356
1357 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1358 varying ? "XXX: " : "",
1359 attr[i].shift, attr[i].extra_flags);
1360 }
1361
1362 /* Print the special record name */
1363 bool attribute = false;
1364 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1365
1366 /* Sanity check */
1367 if (attribute == varying)
1368 pandecode_msg("XXX: mismatched special record\n");
1369
1370 continue;
1371 }
1372
1373 enum mali_attr_mode mode = attr[i].elements & 7;
1374
1375 if (mode == MALI_ATTR_UNUSED)
1376 pandecode_msg("XXX: unused attribute record\n");
1377
1378 /* For non-linear records, we need to print the type of record */
1379 if (mode != MALI_ATTR_LINEAR)
1380 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1381
1382 /* Print the name to link with attr_meta */
1383 pandecode_log_cont("%s_%d", prefix, i);
1384
1385 /* Print the stride and size */
1386 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1387
1388 /* TODO: Sanity check the quotient itself. It must be equal to
1389 * (or be greater than, if the driver added padding) the padded
1390 * vertex count. */
1391
1392 /* Finally, print the pointer */
1393 mali_ptr raw_elements = attr[i].elements & ~7;
1394 char *a = pointer_as_memory_reference(raw_elements);
1395 pandecode_log_cont(" = (%s);\n", a);
1396 free(a);
1397
1398 /* Check the pointer */
1399 pandecode_validate_buffer(raw_elements, attr[i].size);
1400
1401 /* shift/extra_flags exist only for instanced */
1402 if (attr[i].shift | attr[i].extra_flags) {
1403 /* These are set to random values by the blob for
1404 * varyings, most likely a symptom of uninitialized
1405 * memory where the hardware masked the bug. As such we
1406 * put this at a warning, not an error. */
1407
1408 if (mode == MALI_ATTR_LINEAR)
1409 pandecode_msg("warn: instancing fields set for linear\n");
1410
1411 pandecode_prop("shift = %d", attr[i].shift);
1412 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1413 }
1414
1415 /* Decode further where possible */
1416
1417 if (mode == MALI_ATTR_MODULO) {
1418 pandecode_padded_vertices(
1419 attr[i].shift,
1420 attr[i].extra_flags);
1421 }
1422
1423 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1424 i++;
1425 pandecode_log("{\n");
1426 pandecode_indent++;
1427 pandecode_prop("unk = 0x%x", attr[i].unk);
1428 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1429 if (attr[i].zero != 0)
1430 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1431 pandecode_prop("divisor = %d", attr[i].divisor);
1432 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1433 pandecode_indent--;
1434 pandecode_log("}, \n");
1435 }
1436
1437 }
1438
1439 pandecode_log("\n");
1440 }
1441
1442 static mali_ptr
1443 pandecode_shader_address(const char *name, mali_ptr ptr)
1444 {
1445 /* TODO: Decode flags */
1446 mali_ptr shader_ptr = ptr & ~15;
1447
1448 char *a = pointer_as_memory_reference(shader_ptr);
1449 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1450 free(a);
1451
1452 return shader_ptr;
1453 }
1454
1455 static void
1456 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1457 {
1458 unsigned any_nonzero =
1459 stencil->ref | stencil->mask | stencil->func |
1460 stencil->sfail | stencil->dpfail | stencil->dppass;
1461
1462 if (any_nonzero == 0)
1463 return;
1464
1465 const char *func = pandecode_func(stencil->func);
1466 const char *sfail = pandecode_stencil_op(stencil->sfail);
1467 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1468 const char *dppass = pandecode_stencil_op(stencil->dppass);
1469
1470 if (stencil->zero)
1471 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1472
1473 pandecode_log(".stencil_%s = {\n", name);
1474 pandecode_indent++;
1475 pandecode_prop("ref = %d", stencil->ref);
1476 pandecode_prop("mask = 0x%02X", stencil->mask);
1477 pandecode_prop("func = %s", func);
1478 pandecode_prop("sfail = %s", sfail);
1479 pandecode_prop("dpfail = %s", dpfail);
1480 pandecode_prop("dppass = %s", dppass);
1481 pandecode_indent--;
1482 pandecode_log("},\n");
1483 }
1484
1485 static void
1486 pandecode_blend_equation(const struct mali_blend_equation *blend)
1487 {
1488 if (blend->zero1)
1489 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1490
1491 pandecode_log(".equation = {\n");
1492 pandecode_indent++;
1493
1494 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1495 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1496
1497 pandecode_log(".color_mask = ");
1498 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1499 pandecode_log_cont(",\n");
1500
1501 pandecode_indent--;
1502 pandecode_log("},\n");
1503 }
1504
1505 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1506
1507 static unsigned
1508 decode_bifrost_constant(u16 constant)
1509 {
1510 float lo = (float) (constant & 0xFF);
1511 float hi = (float) (constant >> 8);
1512
1513 return (hi / 255.0) + (lo / 65535.0);
1514 }
1515
1516 static mali_ptr
1517 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1518 {
1519 struct bifrost_blend_rt *b =
1520 ((struct bifrost_blend_rt *) descs) + rt_no;
1521
1522 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1523 pandecode_indent++;
1524
1525 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1526 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1527 b->constant, decode_bifrost_constant(b->constant));
1528
1529 /* TODO figure out blend shader enable bit */
1530 pandecode_blend_equation(&b->equation);
1531 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1532 pandecode_prop("index = 0x%" PRIx16, b->index);
1533 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1534
1535 pandecode_indent--;
1536 pandecode_log("},\n");
1537
1538 return 0;
1539 }
1540
1541 static mali_ptr
1542 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1543 {
1544 /* constant/equation is in a union */
1545 if (!blend->shader)
1546 return 0;
1547
1548 pandecode_log(".blend = {\n");
1549 pandecode_indent++;
1550
1551 if (is_shader) {
1552 pandecode_shader_address("shader", blend->shader);
1553 } else {
1554 pandecode_blend_equation(&blend->equation);
1555 pandecode_prop("constant = %f", blend->constant);
1556 }
1557
1558 pandecode_indent--;
1559 pandecode_log("},\n");
1560
1561 /* Return blend shader to disassemble if present */
1562 return is_shader ? (blend->shader & ~0xF) : 0;
1563 }
1564
1565 static mali_ptr
1566 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1567 {
1568 struct midgard_blend_rt *b =
1569 ((struct midgard_blend_rt *) descs) + rt_no;
1570
1571 /* Flags determine presence of blend shader */
1572 bool is_shader = (b->flags & 0xF) >= 0x2;
1573
1574 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1575 pandecode_indent++;
1576
1577 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1578
1579 union midgard_blend blend = b->blend;
1580 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1581
1582 pandecode_indent--;
1583 pandecode_log("};\n");
1584
1585 return shader;
1586 }
1587
1588 /* Attributes and varyings have descriptor records, which contain information
1589 * about their format and ordering with the attribute/varying buffers. We'll
1590 * want to validate that the combinations specified are self-consistent.
1591 */
1592
1593 static int
1594 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1595 {
1596 char base[128];
1597 char *prefix = varying ? "varying" : "attribute";
1598 unsigned max_index = 0;
1599 snprintf(base, sizeof(base), "%s_meta", prefix);
1600
1601 struct mali_attr_meta *attr_meta;
1602 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1603
1604 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1605
1606 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1607 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1608 sizeof(*attr_mem));
1609
1610 /* If the record is discard, it should be zero for everything else */
1611
1612 if (attr_meta->format == MALI_VARYING_DISCARD) {
1613 uint64_t zero =
1614 attr_meta->index |
1615 attr_meta->unknown1 |
1616 attr_meta->unknown3 |
1617 attr_meta->src_offset;
1618
1619 if (zero)
1620 pandecode_msg("XXX: expected empty record for varying discard\n");
1621
1622 /* We want to look for a literal 0000 swizzle -- this
1623 * is not encoded with all zeroes, however */
1624
1625 enum mali_channel z = MALI_CHANNEL_ZERO;
1626 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1627 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1628
1629 if (!good_swizzle)
1630 pandecode_msg("XXX: expected zero swizzle for discard\n");
1631
1632 if (!varying)
1633 pandecode_msg("XXX: cannot discard attribute\n");
1634
1635 /* If we're all good, omit the record */
1636 if (!zero && varying && good_swizzle) {
1637 pandecode_log("/* discarded varying */\n");
1638 continue;
1639 }
1640 }
1641
1642 if (attr_meta->index > max_index)
1643 max_index = attr_meta->index;
1644
1645 if (attr_meta->unknown1 != 0x2) {
1646 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1647 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1648 }
1649
1650 if (attr_meta->unknown3) {
1651 pandecode_msg("XXX: unexpected unknown3 set\n");
1652 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1653 }
1654
1655 pandecode_format_short(attr_meta->format, false);
1656 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1657
1658 if (attr_meta->src_offset)
1659 pandecode_log_cont("[%u]", attr_meta->src_offset);
1660
1661 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1662
1663 pandecode_log_cont(";\n");
1664 }
1665
1666 pandecode_log("\n");
1667
1668 return count ? (max_index + 1) : 0;
1669 }
1670
1671 /* return bits [lo, hi) of word */
1672 static u32
1673 bits(u32 word, u32 lo, u32 hi)
1674 {
1675 if (hi - lo >= 32)
1676 return word; // avoid undefined behavior with the shift
1677
1678 return (word >> lo) & ((1 << (hi - lo)) - 1);
1679 }
1680
1681 static void
1682 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1683 {
1684 pandecode_log_cont("{\n");
1685 pandecode_indent++;
1686
1687 /* Decode invocation_count. See the comment before the definition of
1688 * invocation_count for an explanation.
1689 */
1690
1691 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1692 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1693 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1694 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1695 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1696 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1697
1698 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1699 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1700 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1701
1702 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1703 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1704 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1705
1706 /* Even though we have this decoded, we want to ensure that the
1707 * representation is "unique" so we don't lose anything by printing only
1708 * the final result. More specifically, we need to check that we were
1709 * passed something in canonical form, since the definition per the
1710 * hardware is inherently not unique. How? Well, take the resulting
1711 * decode and pack it ourselves! If it is bit exact with what we
1712 * decoded, we're good to go. */
1713
1714 struct mali_vertex_tiler_prefix ref;
1715 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1716
1717 bool canonical =
1718 (p->invocation_count == ref.invocation_count) &&
1719 (p->invocation_shifts == ref.invocation_shifts);
1720
1721 if (!canonical) {
1722 pandecode_msg("XXX: non-canonical workgroups packing\n");
1723 pandecode_msg("expected: %X, %X",
1724 ref.invocation_count,
1725 ref.invocation_shifts);
1726
1727 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1728 pandecode_prop("size_y_shift = %d", size_y_shift);
1729 pandecode_prop("size_z_shift = %d", size_z_shift);
1730 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1731 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1732 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1733 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1734 }
1735
1736 /* Regardless, print the decode */
1737 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1738 size_x, size_y, size_z,
1739 groups_x, groups_y, groups_z);
1740
1741 /* TODO: Decode */
1742 if (p->unknown_draw)
1743 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1744
1745 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1746
1747 if (p->draw_mode != MALI_DRAW_NONE)
1748 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
1749
1750 /* Index count only exists for tiler jobs anyway */
1751
1752 if (p->index_count)
1753 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1754
1755
1756 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1757 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1758
1759 /* Validate an index buffer is present if we need one. TODO: verify
1760 * relationship between invocation_count and index_count */
1761
1762 if (p->indices) {
1763 unsigned count = p->index_count;
1764
1765 /* Grab the size */
1766 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1767
1768 /* Ensure we got a size, and if so, validate the index buffer
1769 * is large enough to hold a full set of indices of the given
1770 * size */
1771
1772 if (!index_raw_size)
1773 pandecode_msg("XXX: index size missing\n");
1774 else
1775 pandecode_validate_buffer(p->indices, count * size);
1776 } else if (index_raw_size)
1777 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
1778
1779 if (p->offset_bias_correction)
1780 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
1781
1782 /* TODO: Figure out what this is. It's not zero */
1783 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
1784
1785 pandecode_indent--;
1786 pandecode_log("},\n");
1787 }
1788
1789 static void
1790 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1791 {
1792 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1793 struct mali_uniform_buffer_meta *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1794
1795 for (int i = 0; i < ubufs_count; i++) {
1796 unsigned size = (ubufs[i].size + 1) * 16;
1797 mali_ptr addr = ubufs[i].ptr << 2;
1798
1799 pandecode_validate_buffer(addr, size);
1800
1801 char *ptr = pointer_as_memory_reference(ubufs[i].ptr << 2);
1802 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1803 free(ptr);
1804 }
1805
1806 pandecode_log("\n");
1807 }
1808
1809 static void
1810 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1811 {
1812 pandecode_validate_buffer(uniforms, uniform_count * 16);
1813
1814 char *ptr = pointer_as_memory_reference(uniforms);
1815 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1816 free(ptr);
1817 }
1818
1819 static void
1820 pandecode_scratchpad(uintptr_t pscratchpad, int job_no, char *suffix)
1821 {
1822
1823 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(pscratchpad);
1824
1825 struct bifrost_scratchpad *PANDECODE_PTR_VAR(scratchpad, mem, pscratchpad);
1826
1827 if (scratchpad->zero) {
1828 pandecode_msg("XXX: scratchpad zero tripped");
1829 pandecode_prop("zero = 0x%x\n", scratchpad->zero);
1830 }
1831
1832 pandecode_log("struct bifrost_scratchpad scratchpad_%"PRIx64"_%d%s = {\n", pscratchpad, job_no, suffix);
1833 pandecode_indent++;
1834
1835 pandecode_prop("flags = 0x%x", scratchpad->flags);
1836 MEMORY_PROP(scratchpad, gpu_scratchpad);
1837
1838 pandecode_indent--;
1839 pandecode_log("};\n");
1840 }
1841
1842 static const char *
1843 shader_type_for_job(unsigned type)
1844 {
1845 switch (type) {
1846 case JOB_TYPE_VERTEX: return "VERTEX";
1847 case JOB_TYPE_TILER: return "FRAGMENT";
1848 case JOB_TYPE_COMPUTE: return "COMPUTE";
1849 default:
1850 return "UNKNOWN";
1851 }
1852 }
1853
1854 static unsigned shader_id = 0;
1855
1856 static struct midgard_disasm_stats
1857 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1858 bool is_bifrost, unsigned gpu_id)
1859 {
1860 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1861 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1862
1863 /* Compute maximum possible size */
1864 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1865
1866 /* Print some boilerplate to clearly denote the assembly (which doesn't
1867 * obey indentation rules), and actually do the disassembly! */
1868
1869 printf("\n\n");
1870
1871 struct midgard_disasm_stats stats;
1872
1873 if (is_bifrost) {
1874 disassemble_bifrost(code, sz, false);
1875
1876 /* TODO: Extend stats to Bifrost */
1877 stats.texture_count = -128;
1878 stats.sampler_count = -128;
1879 stats.attribute_count = -128;
1880 stats.varying_count = -128;
1881 stats.uniform_count = -128;
1882 stats.uniform_buffer_count = -128;
1883 stats.work_count = -128;
1884
1885 stats.instruction_count = 0;
1886 stats.bundle_count = 0;
1887 stats.quadword_count = 0;
1888 stats.helper_invocations = false;
1889 } else {
1890 stats = disassemble_midgard(stdout, code, sz, gpu_id,
1891 type == JOB_TYPE_TILER ?
1892 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1893 }
1894
1895 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
1896 * driver-internal purposes with the blob and interfere */
1897
1898 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
1899
1900 if (should_shaderdb) {
1901 unsigned nr_threads =
1902 (stats.work_count <= 4) ? 4 :
1903 (stats.work_count <= 8) ? 2 :
1904 1;
1905
1906 printf("shader%d - MESA_SHADER_%s shader: "
1907 "%u inst, %u bundles, %u quadwords, "
1908 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1909 shader_id++,
1910 shader_type_for_job(type),
1911 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1912 stats.work_count, nr_threads);
1913 }
1914
1915
1916 return stats;
1917 }
1918
1919 static void
1920 pandecode_texture(mali_ptr u,
1921 struct pandecode_mapped_memory *tmem,
1922 unsigned job_no, unsigned tex)
1923 {
1924 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
1925
1926 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
1927 pandecode_indent++;
1928
1929 struct mali_texture_format f = t->format;
1930
1931 /* See the definiton of enum mali_texture_type */
1932
1933 bool is_cube = f.type == MALI_TEX_CUBE;
1934 unsigned dimension = is_cube ? 2 : f.type;
1935
1936 pandecode_make_indent();
1937
1938 /* TODO: Are there others? */
1939 bool is_zs = f.format == MALI_Z32_UNORM;
1940
1941 /* Recall Z/S switched the meaning of linear/tiled .. */
1942 if (is_zs && f.layout == MALI_TEXTURE_LINEAR)
1943 pandecode_msg("XXX: depth/stencil cannot be tiled\n");
1944
1945 /* Print the layout. Default is linear; a modifier can denote AFBC or
1946 * u-interleaved/tiled modes */
1947
1948 if (f.layout == MALI_TEXTURE_AFBC)
1949 pandecode_log_cont("afbc");
1950 else if (f.layout == MALI_TEXTURE_TILED)
1951 pandecode_log_cont("tiled");
1952 else if (f.layout == MALI_TEXTURE_LINEAR)
1953 pandecode_log_cont("linear");
1954 else
1955 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
1956
1957 pandecode_swizzle(t->swizzle, f.format);
1958 pandecode_log_cont(" ");
1959
1960 /* Distinguish cube/2D with modifier */
1961
1962 if (is_cube)
1963 pandecode_log_cont("cube ");
1964
1965 pandecode_format_short(f.format, f.srgb);
1966 pandecode_swizzle(f.swizzle, f.format);
1967
1968 /* All four width/height/depth/array_size dimensions are present
1969 * regardless of the type of texture, but it is an error to have
1970 * non-zero dimensions for unused dimensions. Verify this. array_size
1971 * can always be set, as can width. */
1972
1973 if (t->height && dimension < 2)
1974 pandecode_msg("XXX: nonzero height for <2D texture\n");
1975
1976 if (t->depth && dimension < 3)
1977 pandecode_msg("XXX: nonzero depth for <2D texture\n");
1978
1979 /* Print only the dimensions that are actually there */
1980
1981 pandecode_log_cont(": %d", t->width + 1);
1982
1983 if (dimension >= 2)
1984 pandecode_log_cont("x%u", t->height + 1);
1985
1986 if (dimension >= 3)
1987 pandecode_log_cont("x%u", t->depth + 1);
1988
1989 if (t->array_size)
1990 pandecode_log_cont("[%u]", t->array_size + 1);
1991
1992 if (t->levels)
1993 pandecode_log_cont(" mip %u", t->levels);
1994
1995 pandecode_log_cont("\n");
1996
1997 if (f.unknown1 | f.zero) {
1998 pandecode_msg("XXX: texture format zero tripped\n");
1999 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2000 pandecode_prop("zero = %" PRId32, f.zero);
2001 }
2002
2003 if (!f.unknown2) {
2004 pandecode_msg("XXX: expected unknown texture bit set\n");
2005 pandecode_prop("unknown2 = %" PRId32, f.unknown1);
2006 }
2007
2008 if (t->swizzle_zero) {
2009 pandecode_msg("XXX: swizzle zero tripped\n");
2010 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2011 }
2012
2013 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2014 pandecode_msg("XXX: texture zero tripped\n");
2015 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2016 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2017 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2018 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2019 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2020 }
2021
2022 pandecode_log(".payload = {\n");
2023 pandecode_indent++;
2024
2025 /* A bunch of bitmap pointers follow.
2026 * We work out the correct number,
2027 * based on the mipmap/cubemap
2028 * properties, but dump extra
2029 * possibilities to futureproof */
2030
2031 int bitmap_count = t->levels + 1;
2032
2033 /* Miptree for each face */
2034 if (f.type == MALI_TEX_CUBE)
2035 bitmap_count *= 6;
2036 else if (f.type == MALI_TEX_3D)
2037 bitmap_count *= t->depth;
2038
2039 /* Array of textures */
2040 bitmap_count *= (t->array_size + 1);
2041
2042 /* Stride for each element */
2043 if (f.manual_stride)
2044 bitmap_count *= 2;
2045
2046 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2047 u + sizeof(*t), sizeof(mali_ptr) * bitmap_count);
2048 for (int i = 0; i < bitmap_count; ++i) {
2049 /* How we dump depends if this is a stride or a pointer */
2050
2051 if (f.manual_stride && (i & 1)) {
2052 /* signed 32-bit snuck in as a 64-bit pointer */
2053 uint64_t stride_set = pointers_and_strides[i];
2054 uint32_t clamped_stride = stride_set;
2055 int32_t stride = clamped_stride;
2056 assert(stride_set == clamped_stride);
2057 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2058 } else {
2059 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2060 pandecode_log("%s, \n", a);
2061 free(a);
2062 }
2063 }
2064
2065 pandecode_indent--;
2066 pandecode_log("},\n");
2067
2068 pandecode_indent--;
2069 pandecode_log("};\n");
2070 }
2071
2072 /* For shader properties like texture_count, we have a claimed property in the shader_meta, and the actual Truth from static analysis (this may just be an upper limit). We validate accordingly */
2073
2074 static void
2075 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2076 {
2077 /* Nothing to do */
2078 if (claim == truth)
2079 return;
2080
2081 if (fuzzy)
2082 assert(truth >= 0);
2083
2084 if ((truth >= 0) && !fuzzy) {
2085 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2086 (truth < claim) ? "warn" : "XXX",
2087 name, truth, claim);
2088 } else if ((claim > -truth) && !fuzzy) {
2089 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2090 name, -truth, claim);
2091 } else if (fuzzy && (claim < truth))
2092 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2093 name, truth, claim);
2094
2095 pandecode_log(".%s = %" PRId16, name, claim);
2096
2097 if (fuzzy)
2098 pandecode_log_cont(" /* %u used */", truth);
2099
2100 pandecode_log_cont(",\n");
2101 }
2102
2103 static void
2104 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2105 bool is_bifrost, unsigned gpu_id)
2106 {
2107 struct midgard_disasm_stats stats =
2108 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2109
2110 bool has_texture = (stats.texture_count > 0);
2111 bool has_sampler = (stats.sampler_count > 0);
2112 bool has_attribute = (stats.attribute_count > 0);
2113 bool has_varying = (stats.varying_count > 0);
2114 bool has_uniform = (stats.uniform_count > 0);
2115 bool has_ubo = (stats.uniform_buffer_count > 0);
2116
2117 if (has_texture || has_sampler)
2118 pandecode_msg("XXX: blend shader accessing textures\n");
2119
2120 if (has_attribute || has_varying)
2121 pandecode_msg("XXX: blend shader accessing interstage\n");
2122
2123 if (has_uniform || has_ubo)
2124 pandecode_msg("XXX: blend shader accessing uniforms\n");
2125 }
2126
2127 static void
2128 pandecode_vertex_tiler_postfix_pre(
2129 const struct mali_vertex_tiler_postfix *p,
2130 int job_no, enum mali_job_type job_type,
2131 char *suffix, bool is_bifrost, unsigned gpu_id)
2132 {
2133 struct pandecode_mapped_memory *attr_mem;
2134
2135 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2136 * are the only things actually needed from the FBD, vertex/tiler jobs
2137 * no longer reference the FBD -- instead, this field points to some
2138 * info about the scratchpad.
2139 */
2140
2141 struct pandecode_fbd fbd_info = {
2142 /* Default for Bifrost */
2143 .rt_count = 1
2144 };
2145
2146 if (is_bifrost)
2147 pandecode_scratchpad(p->framebuffer & ~1, job_no, suffix);
2148 else if (p->framebuffer & MALI_MFBD)
2149 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->framebuffer) & FBD_MASK, job_no, false);
2150 else if (job_type == JOB_TYPE_COMPUTE)
2151 pandecode_compute_fbd((u64) (uintptr_t) p->framebuffer, job_no);
2152 else
2153 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->framebuffer, job_no, false, gpu_id);
2154
2155 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2156 int texture_count = 0, sampler_count = 0;
2157
2158 if (p->shader) {
2159 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2160 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2161
2162 /* Disassemble ahead-of-time to get stats. Initialize with
2163 * stats for the missing-shader case so we get validation
2164 * there, too */
2165
2166 struct midgard_disasm_stats info = {
2167 .texture_count = 0,
2168 .sampler_count = 0,
2169 .attribute_count = 0,
2170 .varying_count = 0,
2171 .work_count = 1,
2172
2173 .uniform_count = -128,
2174 .uniform_buffer_count = 0
2175 };
2176
2177 if (s->shader & ~0xF)
2178 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2179
2180 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2181 pandecode_indent++;
2182
2183 /* Save for dumps */
2184 attribute_count = s->attribute_count;
2185 varying_count = s->varying_count;
2186 texture_count = s->texture_count;
2187 sampler_count = s->sampler_count;
2188
2189 if (is_bifrost) {
2190 uniform_count = s->bifrost2.uniform_count;
2191 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2192 } else {
2193 uniform_count = s->midgard1.uniform_count;
2194 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2195 }
2196
2197 pandecode_shader_address("shader", s->shader);
2198
2199 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2200 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2201 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2202 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2203 pandecode_shader_prop("uniform_buffer_count",
2204 uniform_buffer_count,
2205 info.uniform_buffer_count, true);
2206
2207 if (!is_bifrost) {
2208 pandecode_shader_prop("uniform_count",
2209 uniform_count,
2210 info.uniform_count, false);
2211
2212 pandecode_shader_prop("work_count",
2213 s->midgard1.work_count, info.work_count, false);
2214 }
2215
2216 if (is_bifrost) {
2217 pandecode_prop("bifrost1.unk1 = 0x%" PRIx32, s->bifrost1.unk1);
2218 } else {
2219 bool helpers = s->midgard1.flags & MALI_HELPER_INVOCATIONS;
2220 s->midgard1.flags &= ~MALI_HELPER_INVOCATIONS;
2221
2222 if (helpers != info.helper_invocations) {
2223 pandecode_msg("XXX: expected helpers %u but got %u\n",
2224 info.helper_invocations, helpers);
2225 }
2226
2227 pandecode_log(".midgard1.flags = ");
2228 pandecode_log_decoded_flags(shader_midgard1_flag_info, s->midgard1.flags);
2229 pandecode_log_cont(",\n");
2230
2231 pandecode_prop("midgard1.unknown2 = 0x%" PRIx32, s->midgard1.unknown2);
2232 }
2233
2234 if (s->depth_units || s->depth_factor) {
2235 pandecode_prop("depth_factor = %f", s->depth_factor);
2236 pandecode_prop("depth_units = %f", s->depth_units);
2237 }
2238
2239 if (s->alpha_coverage) {
2240 bool invert_alpha_coverage = s->alpha_coverage & 0xFFF0;
2241 uint16_t inverted_coverage = invert_alpha_coverage ? ~s->alpha_coverage : s->alpha_coverage;
2242
2243 pandecode_prop("alpha_coverage = %sMALI_ALPHA_COVERAGE(%f)",
2244 invert_alpha_coverage ? "~" : "",
2245 MALI_GET_ALPHA_COVERAGE(inverted_coverage));
2246 }
2247
2248 if (s->unknown2_3 || s->unknown2_4) {
2249 pandecode_log(".unknown2_3 = ");
2250
2251 int unknown2_3 = s->unknown2_3;
2252 int unknown2_4 = s->unknown2_4;
2253
2254 /* We're not quite sure what these flags mean without the depth test, if anything */
2255
2256 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2257 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2258 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2259
2260 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2261 }
2262
2263 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2264 pandecode_log_cont(",\n");
2265
2266 pandecode_log(".unknown2_4 = ");
2267 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2268 pandecode_log_cont(",\n");
2269 }
2270
2271 if (s->stencil_mask_front || s->stencil_mask_back) {
2272 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2273 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2274 }
2275
2276 pandecode_stencil("front", &s->stencil_front);
2277 pandecode_stencil("back", &s->stencil_back);
2278
2279 if (is_bifrost) {
2280 pandecode_log(".bifrost2 = {\n");
2281 pandecode_indent++;
2282
2283 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2284 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2285 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2286 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2287
2288 pandecode_indent--;
2289 pandecode_log("},\n");
2290 } else if (s->midgard2.unknown2_7) {
2291 pandecode_log(".midgard2 = {\n");
2292 pandecode_indent++;
2293
2294 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2295 pandecode_indent--;
2296 pandecode_log("},\n");
2297 }
2298
2299 if (s->unknown2_8)
2300 pandecode_prop("unknown2_8 = 0x%" PRIx32, s->unknown2_8);
2301
2302 if (!is_bifrost) {
2303 /* TODO: Blend shaders routing/disasm */
2304 union midgard_blend blend = s->blend;
2305 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2306 if (shader & ~0xF)
2307 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2308 }
2309
2310 pandecode_indent--;
2311 pandecode_log("};\n");
2312
2313 /* MRT blend fields are used whenever MFBD is used, with
2314 * per-RT descriptors */
2315
2316 if (job_type == JOB_TYPE_TILER && p->framebuffer & MALI_MFBD) {
2317 void* blend_base = (void *) (s + 1);
2318
2319 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2320 mali_ptr shader = 0;
2321
2322 if (is_bifrost)
2323 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2324 else
2325 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2326
2327 if (shader & ~0xF)
2328 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2329
2330 }
2331 }
2332 } else
2333 pandecode_msg("XXX: missing shader descriptor\n");
2334
2335 if (p->viewport) {
2336 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2337 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2338
2339 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2340 pandecode_indent++;
2341
2342 pandecode_prop("clip_minx = %f", f->clip_minx);
2343 pandecode_prop("clip_miny = %f", f->clip_miny);
2344 pandecode_prop("clip_minz = %f", f->clip_minz);
2345 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2346 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2347 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2348
2349 /* Only the higher coordinates are MALI_POSITIVE scaled */
2350
2351 pandecode_prop("viewport0 = { %d, %d }",
2352 f->viewport0[0], f->viewport0[1]);
2353
2354 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2355 f->viewport1[0] + 1, f->viewport1[1] + 1);
2356
2357 pandecode_indent--;
2358 pandecode_log("};\n");
2359 }
2360
2361 unsigned max_attr_index = 0;
2362
2363 if (p->attribute_meta)
2364 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2365
2366 if (p->attributes) {
2367 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2368 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2369 }
2370
2371 /* Varyings are encoded like attributes but not actually sent; we just
2372 * pass a zero buffer with the right stride/size set, (or whatever)
2373 * since the GPU will write to it itself */
2374
2375 if (p->varying_meta) {
2376 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2377 }
2378
2379 if (p->varyings) {
2380 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2381
2382 /* Number of descriptors depends on whether there are
2383 * non-internal varyings */
2384
2385 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2386 }
2387
2388 if (p->uniform_buffers) {
2389 if (uniform_buffer_count)
2390 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2391 else
2392 pandecode_msg("warn: UBOs specified but not referenced\n");
2393 } else if (uniform_buffer_count)
2394 pandecode_msg("XXX: UBOs referenced but not specified\n");
2395
2396 /* We don't want to actually dump uniforms, but we do need to validate
2397 * that the counts we were given are sane */
2398
2399 if (p->uniforms) {
2400 if (uniform_count)
2401 pandecode_uniforms(p->uniforms, uniform_count);
2402 else
2403 pandecode_msg("warn: Uniforms specified but not referenced\n");
2404 } else if (uniform_count)
2405 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2406
2407 if (p->texture_trampoline) {
2408 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(p->texture_trampoline);
2409
2410 if (mmem) {
2411 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline);
2412
2413 pandecode_log("uint64_t texture_trampoline_%"PRIx64"_%d[] = {\n", p->texture_trampoline, job_no);
2414 pandecode_indent++;
2415
2416 for (int tex = 0; tex < texture_count; ++tex) {
2417 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2418 char *a = pointer_as_memory_reference(*u);
2419 pandecode_log("%s,\n", a);
2420 free(a);
2421 }
2422
2423 pandecode_indent--;
2424 pandecode_log("};\n");
2425
2426 /* Now, finally, descend down into the texture descriptor */
2427 for (unsigned tex = 0; tex < texture_count; ++tex) {
2428 mali_ptr *PANDECODE_PTR_VAR(u, mmem, p->texture_trampoline + tex * sizeof(mali_ptr));
2429 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2430 if (tmem)
2431 pandecode_texture(*u, tmem, job_no, tex);
2432 }
2433 }
2434 }
2435
2436 if (p->sampler_descriptor) {
2437 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->sampler_descriptor);
2438
2439 if (smem) {
2440 struct mali_sampler_descriptor *s;
2441
2442 mali_ptr d = p->sampler_descriptor;
2443
2444 for (int i = 0; i < sampler_count; ++i) {
2445 s = pandecode_fetch_gpu_mem(smem, d + sizeof(*s) * i, sizeof(*s));
2446
2447 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", d + sizeof(*s) * i, job_no, i);
2448 pandecode_indent++;
2449
2450 pandecode_log(".filter_mode = ");
2451 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2452 pandecode_log_cont(",\n");
2453
2454 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2455 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2456
2457 if (s->lod_bias)
2458 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2459
2460 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2461 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2462 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2463
2464 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2465
2466 if (s->zero || s->zero2) {
2467 pandecode_msg("XXX: sampler zero tripped\n");
2468 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2469 }
2470
2471 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2472
2473 pandecode_prop("border_color = { %f, %f, %f, %f }",
2474 s->border_color[0],
2475 s->border_color[1],
2476 s->border_color[2],
2477 s->border_color[3]);
2478
2479 pandecode_indent--;
2480 pandecode_log("};\n");
2481 }
2482 }
2483 }
2484 }
2485
2486 static void
2487 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2488 {
2489 if (p->shader & 0xF)
2490 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2491
2492 if (!(p->position_varying || p->occlusion_counter))
2493 return;
2494
2495 pandecode_log(".postfix = {\n");
2496 pandecode_indent++;
2497
2498 MEMORY_PROP(p, position_varying);
2499 MEMORY_PROP(p, occlusion_counter);
2500
2501 pandecode_indent--;
2502 pandecode_log("},\n");
2503 }
2504
2505 static void
2506 pandecode_vertex_only_bfr(struct bifrost_vertex_only *v)
2507 {
2508 pandecode_log_cont("{\n");
2509 pandecode_indent++;
2510
2511 pandecode_prop("unk2 = 0x%x", v->unk2);
2512
2513 if (v->zero0 || v->zero1) {
2514 pandecode_msg("XXX: vertex only zero tripped");
2515 pandecode_prop("zero0 = 0x%" PRIx32, v->zero0);
2516 pandecode_prop("zero1 = 0x%" PRIx64, v->zero1);
2517 }
2518
2519 pandecode_indent--;
2520 pandecode_log("}\n");
2521 }
2522
2523 static void
2524 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2525 {
2526
2527 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2528 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2529
2530 pandecode_log("struct mali_tiler_heap_meta tiler_heap_meta_%d = {\n", job_no);
2531 pandecode_indent++;
2532
2533 if (h->zero) {
2534 pandecode_msg("XXX: tiler heap zero tripped\n");
2535 pandecode_prop("zero = 0x%x", h->zero);
2536 }
2537
2538 for (int i = 0; i < 12; i++) {
2539 if (h->zeros[i] != 0) {
2540 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2541 i, h->zeros[i]);
2542 }
2543 }
2544
2545 pandecode_prop("heap_size = 0x%x", h->heap_size);
2546 MEMORY_PROP(h, tiler_heap_start);
2547 MEMORY_PROP(h, tiler_heap_free);
2548
2549 /* this might point to the beginning of another buffer, when it's
2550 * really the end of the tiler heap buffer, so we have to be careful
2551 * here. but for zero length, we need the same pointer.
2552 */
2553
2554 if (h->tiler_heap_end == h->tiler_heap_start) {
2555 MEMORY_PROP(h, tiler_heap_start);
2556 } else {
2557 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2558 pandecode_prop("tiler_heap_end = %s + 1", a);
2559 free(a);
2560 }
2561
2562 pandecode_indent--;
2563 pandecode_log("};\n");
2564 }
2565
2566 static void
2567 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2568 {
2569 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2570 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2571
2572 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2573
2574 pandecode_log("struct bifrost_tiler_meta tiler_meta_%d = {\n", job_no);
2575 pandecode_indent++;
2576
2577 if (t->zero0 || t->zero1) {
2578 pandecode_msg("XXX: tiler meta zero tripped\n");
2579 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
2580 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2581 }
2582
2583 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
2584 pandecode_prop("flags = 0x%" PRIx16, t->flags);
2585
2586 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
2587 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
2588
2589 for (int i = 0; i < 12; i++) {
2590 if (t->zeros[i] != 0) {
2591 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
2592 i, t->zeros[i]);
2593 }
2594 }
2595
2596 pandecode_indent--;
2597 pandecode_log("};\n");
2598 }
2599
2600 static void
2601 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2602 {
2603 pandecode_log(".gl_enables = ");
2604
2605 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2606
2607 pandecode_log_cont(",\n");
2608 }
2609
2610 static void
2611 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
2612 {
2613 if (u.pointer == 0x0)
2614 return;
2615
2616 pandecode_log(".primitive_size = {\n");
2617 pandecode_indent++;
2618
2619 if (constant) {
2620 pandecode_prop("constant = %f", u.constant);
2621 } else {
2622 MEMORY_PROP((&u), pointer);
2623 }
2624
2625 pandecode_indent--;
2626 pandecode_log("},\n");
2627 }
2628
2629 static void
2630 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
2631 {
2632 pandecode_log_cont("{\n");
2633 pandecode_indent++;
2634
2635 /* TODO: gl_PointSize on Bifrost */
2636 pandecode_primitive_size(t->primitive_size, true);
2637
2638 pandecode_gl_enables(t->gl_enables, JOB_TYPE_TILER);
2639
2640 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
2641 || t->zero6 || t->zero7 || t->zero8) {
2642 pandecode_msg("XXX: tiler only zero tripped\n");
2643 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
2644 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
2645 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
2646 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
2647 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
2648 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
2649 pandecode_prop("zero7 = 0x%" PRIx32, t->zero7);
2650 pandecode_prop("zero8 = 0x%" PRIx64, t->zero8);
2651 }
2652
2653 pandecode_indent--;
2654 pandecode_log("},\n");
2655 }
2656
2657 static int
2658 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
2659 const struct pandecode_mapped_memory *mem,
2660 mali_ptr payload, int job_no, unsigned gpu_id)
2661 {
2662 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
2663
2664 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
2665
2666 pandecode_log("struct bifrost_payload_vertex payload_%d = {\n", job_no);
2667 pandecode_indent++;
2668
2669 pandecode_log(".prefix = ");
2670 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
2671
2672 pandecode_log(".vertex = ");
2673 pandecode_vertex_only_bfr(&v->vertex);
2674
2675 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
2676
2677 pandecode_indent--;
2678 pandecode_log("};\n");
2679
2680 return sizeof(*v);
2681 }
2682
2683 static int
2684 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
2685 const struct pandecode_mapped_memory *mem,
2686 mali_ptr payload, int job_no, unsigned gpu_id)
2687 {
2688 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
2689
2690 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
2691 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
2692
2693 pandecode_log("struct bifrost_payload_tiler payload_%d = {\n", job_no);
2694 pandecode_indent++;
2695
2696 pandecode_log(".prefix = ");
2697 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
2698
2699 pandecode_log(".tiler = ");
2700 pandecode_tiler_only_bfr(&t->tiler, job_no);
2701
2702 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
2703
2704 pandecode_indent--;
2705 pandecode_log("};\n");
2706
2707 return sizeof(*t);
2708 }
2709
2710 static int
2711 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
2712 const struct pandecode_mapped_memory *mem,
2713 mali_ptr payload, int job_no, unsigned gpu_id)
2714 {
2715 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
2716
2717 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
2718
2719 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
2720 pandecode_indent++;
2721
2722 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
2723 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
2724
2725 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
2726
2727 pandecode_log(".prefix = ");
2728 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
2729
2730 pandecode_gl_enables(v->gl_enables, h->job_type);
2731
2732 if (v->instance_shift || v->instance_odd) {
2733 pandecode_prop("instance_shift = 0x%d /* %d */",
2734 v->instance_shift, 1 << v->instance_shift);
2735 pandecode_prop("instance_odd = 0x%X /* %d */",
2736 v->instance_odd, (2 * v->instance_odd) + 1);
2737
2738 pandecode_padded_vertices(v->instance_shift, v->instance_odd);
2739 }
2740
2741 if (v->offset_start)
2742 pandecode_prop("offset_start = %d", v->offset_start);
2743
2744 if (v->zero5) {
2745 pandecode_msg("XXX: midgard payload zero tripped\n");
2746 pandecode_prop("zero5 = 0x%" PRIx64, v->zero5);
2747 }
2748
2749 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
2750
2751 pandecode_indent--;
2752 pandecode_log("};\n");
2753
2754 return sizeof(*v);
2755 }
2756
2757 static int
2758 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
2759 mali_ptr payload, int job_no,
2760 bool is_bifrost, unsigned gpu_id)
2761 {
2762 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
2763
2764 bool is_mfbd = s->framebuffer & MALI_MFBD;
2765
2766 /* Bifrost theoretically may retain support for SFBD on compute jobs,
2767 * but for graphics workloads with a FRAGMENT payload, use MFBD */
2768
2769 if (!is_mfbd && is_bifrost)
2770 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
2771
2772 struct pandecode_fbd info;
2773
2774 if (is_mfbd)
2775 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true);
2776 else
2777 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
2778
2779 /* Compute the tag for the tagged pointer. This contains the type of
2780 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
2781 * additional structures follow the MFBD header (an extra payload or
2782 * not, as well as a count of render targets) */
2783
2784 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
2785
2786 if (is_mfbd) {
2787 if (info.has_extra)
2788 expected_tag |= MALI_MFBD_TAG_EXTRA;
2789
2790 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
2791 }
2792
2793 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2794 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2795 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2796 pandecode_prop("max_tile_coord = 0x%X\n", s->min_tile_coord);
2797 }
2798
2799 /* Extract tile coordinates */
2800
2801 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2802 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2803
2804 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2805 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2806
2807 /* For the max, we also want the floored (rather than ceiled) version for checking */
2808
2809 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2810 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2811
2812 /* Validate the coordinates are well-ordered */
2813
2814 if (min_x == max_x)
2815 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2816 else if (min_x > max_x)
2817 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2818
2819 if (min_y == max_y)
2820 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2821 else if (min_y > max_y)
2822 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2823
2824 /* Validate the coordinates fit inside the framebuffer. We use floor,
2825 * rather than ceil, for the max coordinates, since the tile
2826 * coordinates for something like an 800x600 framebuffer will actually
2827 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2828
2829 if ((min_x > info.width) || (max_x_f > info.width))
2830 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2831
2832 if ((min_y > info.height) || (max_y_f > info.height))
2833 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2834
2835 /* After validation, we print */
2836
2837 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2838
2839 /* The FBD is a tagged pointer */
2840
2841 unsigned tag = (s->framebuffer & ~FBD_MASK);
2842
2843 if (tag != expected_tag)
2844 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2845
2846 return sizeof(*s);
2847 }
2848
2849 static int job_descriptor_number = 0;
2850
2851 int
2852 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id)
2853 {
2854 struct mali_job_descriptor_header *h;
2855
2856 int start_number = 0;
2857
2858 bool first = true;
2859 bool last_size;
2860
2861 do {
2862 struct pandecode_mapped_memory *mem =
2863 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2864
2865 void *payload;
2866
2867 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2868
2869 /* On Midgard, for 32-bit jobs except for fragment jobs, the
2870 * high 32-bits of the 64-bit pointer are reused to store
2871 * something else.
2872 */
2873 int offset = h->job_descriptor_size == MALI_JOB_32 &&
2874 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
2875 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
2876
2877 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
2878
2879 int job_no = job_descriptor_number++;
2880
2881 if (first)
2882 start_number = job_no;
2883
2884 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2885 pandecode_indent++;
2886
2887 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
2888
2889 /* Save for next job fixing */
2890 last_size = h->job_descriptor_size;
2891
2892 if (h->job_descriptor_size)
2893 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2894
2895 if (h->exception_status && h->exception_status != 0x1)
2896 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2897 h->exception_status,
2898 (h->exception_status >> 16) & 0xFFFF,
2899 pandecode_exception_access((h->exception_status >> 8) & 0x3),
2900 h->exception_status & 0xFF);
2901
2902 if (h->first_incomplete_task)
2903 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2904
2905 if (h->fault_pointer)
2906 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2907
2908 if (h->job_barrier)
2909 pandecode_prop("job_barrier = %d", h->job_barrier);
2910
2911 pandecode_prop("job_index = %d", h->job_index);
2912
2913 if (h->unknown_flags)
2914 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2915
2916 if (h->job_dependency_index_1)
2917 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2918
2919 if (h->job_dependency_index_2)
2920 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2921
2922 pandecode_indent--;
2923 pandecode_log("};\n");
2924
2925 /* Do not touch the field yet -- decode the payload first, and
2926 * don't touch that either. This is essential for the uploads
2927 * to occur in sequence and therefore be dynamically allocated
2928 * correctly. Do note the size, however, for that related
2929 * reason. */
2930
2931 switch (h->job_type) {
2932 case JOB_TYPE_WRITE_VALUE: {
2933 struct mali_payload_write_value *s = payload;
2934 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2935 pandecode_indent++;
2936 MEMORY_PROP(s, address);
2937
2938 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2939 pandecode_msg("XXX: unknown value descriptor\n");
2940 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2941 }
2942
2943 if (s->reserved) {
2944 pandecode_msg("XXX: set value tripped\n");
2945 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2946 }
2947
2948 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2949 pandecode_indent--;
2950 pandecode_log("};\n");
2951
2952 break;
2953 }
2954
2955 case JOB_TYPE_TILER:
2956 case JOB_TYPE_VERTEX:
2957 case JOB_TYPE_COMPUTE:
2958 if (bifrost) {
2959 if (h->job_type == JOB_TYPE_TILER)
2960 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2961 else
2962 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2963 } else
2964 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2965
2966 break;
2967
2968 case JOB_TYPE_FRAGMENT:
2969 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2970 break;
2971
2972 default:
2973 break;
2974 }
2975
2976 /* Handle linkage */
2977
2978 if (!first) {
2979 pandecode_log("((struct mali_job_descriptor_header *) (uintptr_t) job_%d_p)->", job_no - 1);
2980 pandecode_log_cont("next_job = job_%d_p;\n\n", job_no);
2981 }
2982
2983 first = false;
2984
2985 } while ((jc_gpu_va = h->next_job));
2986
2987 return start_number;
2988 }