projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: XMLify mali_func
[mesa.git] / src / panfrost / lib / 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 <midgard_pack.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 "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
43
44 #define MEMORY_PROP(obj, p) {\
45 if (obj->p) { \
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
48 free(a); \
49 } \
50 }
51
52 #define MEMORY_PROP_DIR(obj, p) {\
53 if (obj.p) { \
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
56 free(a); \
57 } \
58 }
59
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, 0); \
65 }
66
67 #define DUMP_ADDR(title, T, addr, indent) {\
68 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
69 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
70 DUMP_CL(title, T, cl, indent); \
71 }
72
73 FILE *pandecode_dump_stream;
74
75 /* Semantic logging type.
76 *
77 * Raw: for raw messages to be printed as is.
78 * Message: for helpful information to be commented out in replays.
79 * Property: for properties of a struct
80 *
81 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
82 */
83
84 enum pandecode_log_type {
85 PANDECODE_RAW,
86 PANDECODE_MESSAGE,
87 PANDECODE_PROPERTY
88 };
89
90 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
91 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
92 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
93
94 unsigned pandecode_indent = 0;
95
96 static void
97 pandecode_make_indent(void)
98 {
99 for (unsigned i = 0; i < pandecode_indent; ++i)
100 fprintf(pandecode_dump_stream, " ");
101 }
102
103 static void
104 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
105 {
106 va_list ap;
107
108 pandecode_make_indent();
109
110 if (type == PANDECODE_MESSAGE)
111 fprintf(pandecode_dump_stream, "// ");
112 else if (type == PANDECODE_PROPERTY)
113 fprintf(pandecode_dump_stream, ".");
114
115 va_start(ap, format);
116 vfprintf(pandecode_dump_stream, format, ap);
117 va_end(ap);
118
119 if (type == PANDECODE_PROPERTY)
120 fprintf(pandecode_dump_stream, ",\n");
121 }
122
123 static void
124 pandecode_log_cont(const char *format, ...)
125 {
126 va_list ap;
127
128 va_start(ap, format);
129 vfprintf(pandecode_dump_stream, format, ap);
130 va_end(ap);
131 }
132
133 /* To check for memory safety issues, validates that the given pointer in GPU
134 * memory is valid, containing at least sz bytes. The goal is to eliminate
135 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
136 * overruns) by statically validating pointers.
137 */
138
139 static void
140 pandecode_validate_buffer(mali_ptr addr, size_t sz)
141 {
142 if (!addr) {
143 pandecode_msg("XXX: null pointer deref");
144 return;
145 }
146
147 /* Find a BO */
148
149 struct pandecode_mapped_memory *bo =
150 pandecode_find_mapped_gpu_mem_containing(addr);
151
152 if (!bo) {
153 pandecode_msg("XXX: invalid memory dereference\n");
154 return;
155 }
156
157 /* Bounds check */
158
159 unsigned offset = addr - bo->gpu_va;
160 unsigned total = offset + sz;
161
162 if (total > bo->length) {
163 pandecode_msg("XXX: buffer overrun. "
164 "Chunk of size %zu at offset %d in buffer of size %zu. "
165 "Overrun by %zu bytes. \n",
166 sz, offset, bo->length, total - bo->length);
167 return;
168 }
169 }
170
171 struct pandecode_flag_info {
172 u64 flag;
173 const char *name;
174 };
175
176 static void
177 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
178 u64 flags)
179 {
180 bool decodable_flags_found = false;
181
182 for (int i = 0; flag_info[i].name; i++) {
183 if ((flags & flag_info[i].flag) != flag_info[i].flag)
184 continue;
185
186 if (!decodable_flags_found) {
187 decodable_flags_found = true;
188 } else {
189 pandecode_log_cont(" | ");
190 }
191
192 pandecode_log_cont("%s", flag_info[i].name);
193
194 flags &= ~flag_info[i].flag;
195 }
196
197 if (decodable_flags_found) {
198 if (flags)
199 pandecode_log_cont(" | 0x%" PRIx64, flags);
200 } else {
201 pandecode_log_cont("0x%" PRIx64, flags);
202 }
203 }
204
205 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
206 static const struct pandecode_flag_info gl_enable_flag_info[] = {
207 FLAG_INFO(OCCLUSION_QUERY),
208 FLAG_INFO(OCCLUSION_PRECISE),
209 FLAG_INFO(FRONT_CCW_TOP),
210 FLAG_INFO(CULL_FACE_FRONT),
211 FLAG_INFO(CULL_FACE_BACK),
212 {}
213 };
214 #undef FLAG_INFO
215
216 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
217 static const struct pandecode_flag_info clear_flag_info[] = {
218 FLAG_INFO(FAST),
219 FLAG_INFO(SLOW),
220 FLAG_INFO(SLOW_STENCIL),
221 {}
222 };
223 #undef FLAG_INFO
224
225 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
226 static const struct pandecode_flag_info mask_flag_info[] = {
227 FLAG_INFO(R),
228 FLAG_INFO(G),
229 FLAG_INFO(B),
230 FLAG_INFO(A),
231 {}
232 };
233 #undef FLAG_INFO
234
235 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
236 static const struct pandecode_flag_info u3_flag_info[] = {
237 FLAG_INFO(HAS_MSAA),
238 FLAG_INFO(PER_SAMPLE),
239 FLAG_INFO(CAN_DISCARD),
240 FLAG_INFO(HAS_BLEND_SHADER),
241 FLAG_INFO(DEPTH_WRITEMASK),
242 FLAG_INFO(DEPTH_CLIP_NEAR),
243 FLAG_INFO(DEPTH_CLIP_FAR),
244 {}
245 };
246
247 static const struct pandecode_flag_info u4_flag_info[] = {
248 FLAG_INFO(NO_MSAA),
249 FLAG_INFO(NO_DITHER),
250 FLAG_INFO(DEPTH_RANGE_A),
251 FLAG_INFO(DEPTH_RANGE_B),
252 FLAG_INFO(STENCIL_TEST),
253 FLAG_INFO(ALPHA_TO_COVERAGE),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
259 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
260 FLAG_INFO(SRGB),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
266 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
267 FLAG_INFO(YTR),
268 {}
269 };
270 #undef FLAG_INFO
271
272 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
273 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
274 FLAG_INFO(PRESENT),
275 {}
276 };
277 #undef FLAG_INFO
278
279 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
280 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
281 FLAG_INFO(ZS),
282 {}
283 };
284 #undef FLAG_INFO
285
286 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
287 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
288 FLAG_INFO(WRITES_Z),
289 FLAG_INFO(EARLY_Z),
290 FLAG_INFO(READS_TILEBUFFER),
291 FLAG_INFO(WRITES_GLOBAL),
292 FLAG_INFO(READS_ZS),
293 {}
294 };
295
296 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
297 FLAG_INFO(WRITES_S),
298 FLAG_INFO(SUPPRESS_INF_NAN),
299 {}
300 };
301 #undef FLAG_INFO
302
303 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
304 static const struct pandecode_flag_info shader_bifrost_info [] = {
305 FLAG_INFO(FULL_THREAD),
306 FLAG_INFO(EARLY_Z),
307 FLAG_INFO(FIRST_ATEST),
308 {}
309 };
310
311 #undef FLAG_INFO
312
313 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
314 static const struct pandecode_flag_info mfbd_flag_info [] = {
315 FLAG_INFO(DEPTH_WRITE),
316 FLAG_INFO(EXTRA),
317 {}
318 };
319 #undef FLAG_INFO
320
321 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
322 static const struct pandecode_flag_info sampler_flag_info [] = {
323 FLAG_INFO(MAG_NEAREST),
324 FLAG_INFO(MIN_NEAREST),
325 FLAG_INFO(MIP_LINEAR_1),
326 FLAG_INFO(MIP_LINEAR_2),
327 FLAG_INFO(NORM_COORDS),
328 {}
329 };
330 #undef FLAG_INFO
331
332 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
333 static const struct pandecode_flag_info sfbd_unk1_info [] = {
334 FLAG_INFO(MSAA_8),
335 FLAG_INFO(MSAA_A),
336 {}
337 };
338 #undef FLAG_INFO
339
340 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
341 static const struct pandecode_flag_info sfbd_unk2_info [] = {
342 FLAG_INFO(MSAA_B),
343 FLAG_INFO(SRGB),
344 {}
345 };
346 #undef FLAG_INFO
347
348 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
349 static char *pandecode_format(enum mali_format format)
350 {
351 static char unk_format_str[10];
352
353 switch (format) {
354 DEFINE_CASE(ETC2_RGB8);
355 DEFINE_CASE(ETC2_R11_UNORM);
356 DEFINE_CASE(ETC2_RGBA8);
357 DEFINE_CASE(ETC2_RG11_UNORM);
358 DEFINE_CASE(ETC2_R11_SNORM);
359 DEFINE_CASE(ETC2_RG11_SNORM);
360 DEFINE_CASE(ETC2_RGB8A1);
361 DEFINE_CASE(NXR);
362 DEFINE_CASE(BC1_UNORM);
363 DEFINE_CASE(BC2_UNORM);
364 DEFINE_CASE(BC3_UNORM);
365 DEFINE_CASE(BC4_UNORM);
366 DEFINE_CASE(BC4_SNORM);
367 DEFINE_CASE(BC5_UNORM);
368 DEFINE_CASE(BC5_SNORM);
369 DEFINE_CASE(BC6H_UF16);
370 DEFINE_CASE(BC6H_SF16);
371 DEFINE_CASE(BC7_UNORM);
372 DEFINE_CASE(ASTC_3D_LDR);
373 DEFINE_CASE(ASTC_3D_HDR);
374 DEFINE_CASE(ASTC_2D_LDR);
375 DEFINE_CASE(ASTC_2D_HDR);
376 DEFINE_CASE(RGB565);
377 DEFINE_CASE(RGB5_X1_UNORM);
378 DEFINE_CASE(RGB5_A1_UNORM);
379 DEFINE_CASE(RGB10_A2_UNORM);
380 DEFINE_CASE(RGB10_A2_SNORM);
381 DEFINE_CASE(RGB10_A2UI);
382 DEFINE_CASE(RGB10_A2I);
383 DEFINE_CASE(RGB332_UNORM);
384 DEFINE_CASE(RGB233_UNORM);
385 DEFINE_CASE(Z24X8_UNORM);
386 DEFINE_CASE(R32_FIXED);
387 DEFINE_CASE(RG32_FIXED);
388 DEFINE_CASE(RGB32_FIXED);
389 DEFINE_CASE(RGBA32_FIXED);
390 DEFINE_CASE(R11F_G11F_B10F);
391 DEFINE_CASE(R9F_G9F_B9F_E5F);
392 DEFINE_CASE(VARYING_POS);
393 DEFINE_CASE(VARYING_DISCARD);
394
395 DEFINE_CASE(R8_SNORM);
396 DEFINE_CASE(R16_SNORM);
397 DEFINE_CASE(R32_SNORM);
398 DEFINE_CASE(RG8_SNORM);
399 DEFINE_CASE(RG16_SNORM);
400 DEFINE_CASE(RG32_SNORM);
401 DEFINE_CASE(RGB8_SNORM);
402 DEFINE_CASE(RGB16_SNORM);
403 DEFINE_CASE(RGB32_SNORM);
404 DEFINE_CASE(RGBA8_SNORM);
405 DEFINE_CASE(RGBA16_SNORM);
406 DEFINE_CASE(RGBA32_SNORM);
407
408 DEFINE_CASE(R8UI);
409 DEFINE_CASE(R16UI);
410 DEFINE_CASE(R32UI);
411 DEFINE_CASE(RG8UI);
412 DEFINE_CASE(RG16UI);
413 DEFINE_CASE(RG32UI);
414 DEFINE_CASE(RGB8UI);
415 DEFINE_CASE(RGB16UI);
416 DEFINE_CASE(RGB32UI);
417 DEFINE_CASE(RGBA8UI);
418 DEFINE_CASE(RGBA16UI);
419 DEFINE_CASE(RGBA32UI);
420
421 DEFINE_CASE(R8_UNORM);
422 DEFINE_CASE(R16_UNORM);
423 DEFINE_CASE(R32_UNORM);
424 DEFINE_CASE(R32F);
425 DEFINE_CASE(RG8_UNORM);
426 DEFINE_CASE(RG16_UNORM);
427 DEFINE_CASE(RG32_UNORM);
428 DEFINE_CASE(RG32F);
429 DEFINE_CASE(RGB8_UNORM);
430 DEFINE_CASE(RGB16_UNORM);
431 DEFINE_CASE(RGB32_UNORM);
432 DEFINE_CASE(RGB32F);
433 DEFINE_CASE(RGBA4_UNORM);
434 DEFINE_CASE(RGBA8_UNORM);
435 DEFINE_CASE(RGBA16_UNORM);
436 DEFINE_CASE(RGBA32_UNORM);
437 DEFINE_CASE(RGBA32F);
438
439 DEFINE_CASE(R8I);
440 DEFINE_CASE(R16I);
441 DEFINE_CASE(R32I);
442 DEFINE_CASE(RG8I);
443 DEFINE_CASE(R16F);
444 DEFINE_CASE(RG16I);
445 DEFINE_CASE(RG32I);
446 DEFINE_CASE(RG16F);
447 DEFINE_CASE(RGB8I);
448 DEFINE_CASE(RGB16I);
449 DEFINE_CASE(RGB32I);
450 DEFINE_CASE(RGB16F);
451 DEFINE_CASE(RGBA8I);
452 DEFINE_CASE(RGBA16I);
453 DEFINE_CASE(RGBA32I);
454 DEFINE_CASE(RGBA16F);
455
456 DEFINE_CASE(RGBA4);
457 DEFINE_CASE(RGBA8_2);
458 DEFINE_CASE(RGB10_A2_2);
459 default:
460 snprintf(unk_format_str, sizeof(unk_format_str), "MALI_0x%02x", format);
461 return unk_format_str;
462 }
463 }
464
465 #undef DEFINE_CASE
466
467 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
468 static char *
469 pandecode_msaa_mode(enum mali_msaa_mode mode)
470 {
471 switch (mode) {
472 DEFINE_CASE(SINGLE);
473 DEFINE_CASE(AVERAGE);
474 DEFINE_CASE(MULTIPLE);
475 DEFINE_CASE(LAYERED);
476 default:
477 unreachable("Impossible");
478 return "";
479 }
480 }
481 #undef DEFINE_CASE
482
483 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
484 static char *
485 pandecode_stencil_op(enum mali_stencil_op op)
486 {
487 switch (op) {
488 DEFINE_CASE(KEEP);
489 DEFINE_CASE(REPLACE);
490 DEFINE_CASE(ZERO);
491 DEFINE_CASE(INVERT);
492 DEFINE_CASE(INCR_WRAP);
493 DEFINE_CASE(DECR_WRAP);
494 DEFINE_CASE(INCR);
495 DEFINE_CASE(DECR);
496
497 default:
498 pandecode_msg("XXX: invalid stencil op %X\n", op);
499 return "";
500 }
501 }
502
503 #undef DEFINE_CASE
504
505 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
506 {
507 switch(mode) {
508 /* TODO: Combine to just "instanced" once this can be done
509 * unambiguously in all known cases */
510 case MALI_ATTR_POT_DIVIDE:
511 return "instanced_pot";
512 case MALI_ATTR_MODULO:
513 return "instanced_mod";
514 case MALI_ATTR_NPOT_DIVIDE:
515 return "instanced_npot";
516 case MALI_ATTR_IMAGE:
517 return "image";
518 default:
519 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
520 return "";
521 }
522 }
523
524 static const char *
525 pandecode_special_record(uint64_t v, bool* attribute)
526 {
527 switch(v) {
528 case MALI_ATTR_VERTEXID:
529 *attribute = true;
530 return "gl_VertexID";
531 case MALI_ATTR_INSTANCEID:
532 *attribute = true;
533 return "gl_InstanceID";
534 case MALI_VARYING_FRAG_COORD:
535 return "gl_FragCoord";
536 case MALI_VARYING_FRONT_FACING:
537 return "gl_FrontFacing";
538 case MALI_VARYING_POINT_COORD:
539 return "gl_PointCoord";
540 default:
541 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
542 return "";
543 }
544 }
545
546 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
547 static char *
548 pandecode_wrap_mode(enum mali_wrap_mode op)
549 {
550 switch (op) {
551 DEFINE_CASE(REPEAT);
552 DEFINE_CASE(CLAMP_TO_EDGE);
553 DEFINE_CASE(CLAMP);
554 DEFINE_CASE(CLAMP_TO_BORDER);
555 DEFINE_CASE(MIRRORED_REPEAT);
556 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE);
557 DEFINE_CASE(MIRRORED_CLAMP);
558 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER);
559
560 default:
561 pandecode_msg("XXX: invalid wrap mode %X\n", op);
562 return "";
563 }
564 }
565 #undef DEFINE_CASE
566
567 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
568 static char *
569 pandecode_block_format(enum mali_block_format fmt)
570 {
571 switch (fmt) {
572 DEFINE_CASE(TILED);
573 DEFINE_CASE(UNKNOWN);
574 DEFINE_CASE(LINEAR);
575 DEFINE_CASE(AFBC);
576
577 default:
578 unreachable("Invalid case");
579 }
580 }
581 #undef DEFINE_CASE
582
583 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
584 static char *
585 pandecode_exception_access(unsigned access)
586 {
587 switch (access) {
588 DEFINE_CASE(NONE);
589 DEFINE_CASE(EXECUTE);
590 DEFINE_CASE(READ);
591 DEFINE_CASE(WRITE);
592
593 default:
594 unreachable("Invalid case");
595 }
596 }
597 #undef DEFINE_CASE
598
599 /* Midgard's tiler descriptor is embedded within the
600 * larger FBD */
601
602 static void
603 pandecode_midgard_tiler_descriptor(
604 const struct midgard_tiler_descriptor *t,
605 unsigned width,
606 unsigned height,
607 bool is_fragment,
608 bool has_hierarchy)
609 {
610 pandecode_log(".tiler = {\n");
611 pandecode_indent++;
612
613 if (t->hierarchy_mask == MALI_TILER_DISABLED)
614 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
615 else
616 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
617
618 /* We know this name from the kernel, but we never see it nonzero */
619
620 if (t->flags)
621 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
622
623 MEMORY_PROP(t, polygon_list);
624
625 /* The body is offset from the base of the polygon list */
626 //assert(t->polygon_list_body > t->polygon_list);
627 unsigned body_offset = t->polygon_list_body - t->polygon_list;
628
629 /* It needs to fit inside the reported size */
630 //assert(t->polygon_list_size >= body_offset);
631
632 /* Now that we've sanity checked, we'll try to calculate the sizes
633 * ourselves for comparison */
634
635 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
636 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
637
638 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
639 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
640 ref_header, ref_size);
641 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
642 pandecode_msg("body offset %d\n", body_offset);
643 }
644
645 /* The tiler heap has a start and end specified -- it should be
646 * identical to what we have in the BO. The exception is if tiling is
647 * disabled. */
648
649 MEMORY_PROP(t, heap_start);
650 assert(t->heap_end >= t->heap_start);
651
652 struct pandecode_mapped_memory *heap =
653 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
654
655 unsigned heap_size = t->heap_end - t->heap_start;
656
657 /* Tiling is enabled with a special flag */
658 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
659 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
660
661 bool tiling_enabled = hierarchy_mask;
662
663 if (tiling_enabled) {
664 /* When tiling is enabled, the heap should be a tight fit */
665 unsigned heap_offset = t->heap_start - heap->gpu_va;
666 if ((heap_offset + heap_size) != heap->length) {
667 pandecode_msg("XXX: heap size %u (expected %zu)\n",
668 heap_size, heap->length - heap_offset);
669 }
670
671 /* We should also have no other flags */
672 if (tiler_flags)
673 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
674 } else {
675 /* When tiling is disabled, we should have that flag and no others */
676
677 if (tiler_flags != MALI_TILER_DISABLED) {
678 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
679 tiler_flags);
680 }
681
682 /* We should also have an empty heap */
683 if (heap_size) {
684 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
685 heap_size);
686 }
687
688 /* Disabled tiling is used only for clear-only jobs, which are
689 * purely FRAGMENT, so we should never see this for
690 * non-FRAGMENT descriptors. */
691
692 if (!is_fragment)
693 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
694 }
695
696 /* We've never seen weights used in practice, but we know from the
697 * kernel these fields is there */
698
699 bool nonzero_weights = false;
700
701 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
702 nonzero_weights |= t->weights[w] != 0x0;
703 }
704
705 if (nonzero_weights) {
706 pandecode_log(".weights = { ");
707
708 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
709 pandecode_log_cont("%d, ", t->weights[w]);
710 }
711
712 pandecode_log("},");
713 }
714
715 pandecode_indent--;
716 pandecode_log("}\n");
717 }
718
719 /* TODO: The Bifrost tiler is not understood at all yet */
720
721 static void
722 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
723 {
724 pandecode_log(".tiler = {\n");
725 pandecode_indent++;
726
727 MEMORY_PROP(fb, tiler_meta);
728
729 for (int i = 0; i < 16; i++) {
730 if (fb->zeros[i] != 0) {
731 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
732 i, fb->zeros[i]);
733 }
734 }
735
736 pandecode_log("},\n");
737
738 pandecode_indent--;
739 pandecode_log("}\n");
740
741 }
742
743 /* Information about the framebuffer passed back for
744 * additional analysis */
745
746 struct pandecode_fbd {
747 unsigned width;
748 unsigned height;
749 unsigned rt_count;
750 bool has_extra;
751 };
752
753 static void
754 pandecode_sfbd_format(struct mali_sfbd_format format)
755 {
756 pandecode_log(".format = {\n");
757 pandecode_indent++;
758
759 pandecode_log(".unk1 = ");
760 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
761 pandecode_log_cont(",\n");
762
763 /* TODO: Map formats so we can check swizzles and print nicely */
764 pandecode_log("swizzle");
765 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
766 pandecode_log_cont(",\n");
767
768 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
769 (format.nr_channels + 1));
770
771 pandecode_log(".unk2 = ");
772 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
773 pandecode_log_cont(",\n");
774
775 pandecode_prop("block = %s", pandecode_block_format(format.block));
776
777 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
778
779 pandecode_indent--;
780 pandecode_log("},\n");
781 }
782
783 static void
784 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
785 {
786 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
787
788 if (desc->unk0)
789 pandecode_prop("unk0 = 0x%x", desc->unk0);
790
791 if (desc->shared_workgroup_count != 0x1F) {
792 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
793 if (!is_compute)
794 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
795 }
796
797 if (desc->shared_unk1 || desc->shared_shift) {
798 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
799 pandecode_prop("shared_shift = %X", desc->shared_shift);
800
801 if (!is_compute)
802 pandecode_msg("XXX: shared memory configured in noncompute shader");
803 }
804
805 if (desc->shared_zero) {
806 pandecode_msg("XXX: shared memory zero tripped\n");
807 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
808 }
809
810 if (desc->shared_memory && !is_compute)
811 pandecode_msg("XXX: shared memory used in noncompute shader\n");
812
813 MEMORY_PROP(desc, scratchpad);
814 MEMORY_PROP(desc, shared_memory);
815 MEMORY_PROP(desc, unknown1);
816
817 if (desc->scratchpad) {
818 struct pandecode_mapped_memory *smem =
819 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
820
821 pandecode_msg("scratchpad size %u\n", smem->length);
822 }
823
824 }
825
826 static struct pandecode_fbd
827 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
828 {
829 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
830 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
831
832 struct pandecode_fbd info = {
833 .has_extra = false,
834 .rt_count = 1
835 };
836
837 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
838 pandecode_indent++;
839
840 pandecode_log(".shared_memory = {\n");
841 pandecode_indent++;
842 pandecode_shared_memory(&s->shared_memory, false);
843 pandecode_indent--;
844 pandecode_log("},\n");
845
846 pandecode_sfbd_format(s->format);
847
848 info.width = s->width + 1;
849 info.height = s->height + 1;
850
851 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
852 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
853
854 MEMORY_PROP(s, checksum);
855
856 if (s->checksum_stride)
857 pandecode_prop("checksum_stride = %d", s->checksum_stride);
858
859 MEMORY_PROP(s, framebuffer);
860 pandecode_prop("stride = %d", s->stride);
861
862 /* Earlier in the actual commandstream -- right before width -- but we
863 * delay to flow nicer */
864
865 pandecode_log(".clear_flags = ");
866 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
867 pandecode_log_cont(",\n");
868
869 if (s->depth_buffer) {
870 MEMORY_PROP(s, depth_buffer);
871 pandecode_prop("depth_stride = %d", s->depth_stride);
872 }
873
874 if (s->stencil_buffer) {
875 MEMORY_PROP(s, stencil_buffer);
876 pandecode_prop("stencil_stride = %d", s->stencil_stride);
877 }
878
879 if (s->depth_stride_zero ||
880 s->stencil_stride_zero ||
881 s->zero7 || s->zero8) {
882 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
883 pandecode_prop("depth_stride_zero = 0x%x",
884 s->depth_stride_zero);
885 pandecode_prop("stencil_stride_zero = 0x%x",
886 s->stencil_stride_zero);
887 pandecode_prop("zero7 = 0x%" PRIx32,
888 s->zero7);
889 pandecode_prop("zero8 = 0x%" PRIx32,
890 s->zero8);
891 }
892
893 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
894 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
895 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
896 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
897 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
898 }
899
900 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
901 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
902 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
903 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
904 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
905 }
906
907 if (s->clear_stencil) {
908 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
909 }
910
911 const struct midgard_tiler_descriptor t = s->tiler;
912
913 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
914 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
915
916 pandecode_indent--;
917 pandecode_log("};\n");
918
919 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
920 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
921 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
922
923 pandecode_log_cont(".zero3 = {");
924
925 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
926 pandecode_log_cont("%X, ", s->zero3[i]);
927
928 pandecode_log_cont("},\n");
929
930 pandecode_log_cont(".zero6 = {");
931
932 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
933 pandecode_log_cont("%X, ", s->zero6[i]);
934
935 pandecode_log_cont("},\n");
936
937 return info;
938 }
939
940 static void
941 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
942 {
943 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
944 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
945
946 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
947 pandecode_indent++;
948 pandecode_shared_memory(s, true);
949 pandecode_indent--;
950 pandecode_log("},\n");
951 }
952
953 /* Extracts the number of components associated with a Mali format */
954
955 static unsigned
956 pandecode_format_component_count(enum mali_format fmt)
957 {
958 /* Mask out the format class */
959 unsigned top = fmt & 0b11100000;
960
961 switch (top) {
962 case MALI_FORMAT_SNORM:
963 case MALI_FORMAT_UINT:
964 case MALI_FORMAT_UNORM:
965 case MALI_FORMAT_SINT:
966 return ((fmt >> 3) & 3) + 1;
967 default:
968 /* TODO: Validate */
969 return 4;
970 }
971 }
972
973 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
974
975 static unsigned
976 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
977 {
978 unsigned mask = 0;
979 assert(MALI_CHANNEL_RED == 0);
980
981 for (unsigned c = 0; c < 4; ++c) {
982 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
983
984 if (chan <= MALI_CHANNEL_ALPHA)
985 mask |= (1 << chan);
986 }
987
988 return mask;
989 }
990
991 /* Validates that a (format, swizzle) pair is valid, in the sense that the
992 * swizzle doesn't access any components that are undefined in the format.
993 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
994 * omitted */
995
996 static bool
997 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
998 {
999 unsigned nr_comp = pandecode_format_component_count(fmt);
1000 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
1001 unsigned valid_mask = (1 << nr_comp) - 1;
1002 unsigned invalid_mask = ~valid_mask;
1003
1004 if (access_mask & invalid_mask) {
1005 pandecode_msg("XXX: invalid components accessed\n");
1006 return false;
1007 }
1008
1009 /* Check for the default non-swizzling swizzle so we can suppress
1010 * useless printing for the defaults */
1011
1012 unsigned default_swizzles[4] = {
1013 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1014 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1015 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
1016 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
1017 };
1018
1019 return (swizzle == default_swizzles[nr_comp - 1]);
1020 }
1021
1022 /* Maps MALI_RGBA32F to rgba32f, etc */
1023
1024 static void
1025 pandecode_format_short(enum mali_format fmt, bool srgb)
1026 {
1027 /* We want a type-like format, so cut off the initial MALI_ */
1028 char *format = pandecode_format(fmt);
1029 format += strlen("MALI_");
1030
1031 unsigned len = strlen(format);
1032 char *lower_format = calloc(1, len + 1);
1033
1034 for (unsigned i = 0; i < len; ++i)
1035 lower_format[i] = tolower(format[i]);
1036
1037 /* Sanity check sRGB flag is applied to RGB, per the name */
1038 if (srgb && lower_format[0] != 'r')
1039 pandecode_msg("XXX: sRGB applied to non-colour format\n");
1040
1041 /* Just prefix with an s, so you get formats like srgba8_unorm */
1042 if (srgb)
1043 pandecode_log_cont("s");
1044
1045 pandecode_log_cont("%s", lower_format);
1046 free(lower_format);
1047 }
1048
1049 static void
1050 pandecode_swizzle(unsigned swizzle, enum mali_format format)
1051 {
1052 /* First, do some validation */
1053 bool trivial_swizzle = pandecode_validate_format_swizzle(
1054 format, swizzle);
1055
1056 if (trivial_swizzle)
1057 return;
1058
1059 /* Next, print the swizzle */
1060 pandecode_log_cont(".");
1061
1062 static const char components[] = "rgba01";
1063
1064 for (unsigned c = 0; c < 4; ++c) {
1065 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
1066
1067 if (chan >= MALI_CHANNEL_RESERVED_0) {
1068 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
1069 continue;
1070 }
1071 pandecode_log_cont("%c", components[chan]);
1072 }
1073 }
1074
1075 static void
1076 pandecode_rt_format(struct mali_rt_format format)
1077 {
1078 pandecode_log(".format = {\n");
1079 pandecode_indent++;
1080
1081 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1082 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1083 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1084 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1085
1086 pandecode_prop("block = %s", pandecode_block_format(format.block));
1087
1088 /* TODO: Map formats so we can check swizzles and print nicely */
1089 pandecode_log("swizzle");
1090 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1091 pandecode_log_cont(",\n");
1092
1093 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1094 (format.nr_channels + 1));
1095
1096 pandecode_log(".flags = ");
1097 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1098 pandecode_log_cont(",\n");
1099
1100 pandecode_prop("msaa = %s", pandecode_msaa_mode(format.msaa));
1101
1102 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1103 * which is a faster hardware-based alternative to the wallpaper method
1104 * to preserve framebuffer contents across frames. In practice, MFBD
1105 * preload is buggy on Midgard, and so this is a chicken bit. If this
1106 * bit isn't set, most likely something broke unrelated to preload */
1107
1108 if (!format.no_preload) {
1109 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1110 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1111 }
1112
1113 if (format.zero)
1114 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1115
1116 pandecode_indent--;
1117 pandecode_log("},\n");
1118 }
1119
1120 static void
1121 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1122 {
1123 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1124 pandecode_indent++;
1125
1126 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1127 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1128 struct pandecode_mapped_memory *mem =
1129 pandecode_find_mapped_gpu_mem_containing(rt_va);
1130 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1131
1132 pandecode_log("{\n");
1133 pandecode_indent++;
1134
1135 pandecode_rt_format(rt->format);
1136
1137 if (rt->format.block == MALI_BLOCK_AFBC) {
1138 pandecode_log(".afbc = {\n");
1139 pandecode_indent++;
1140
1141 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1142 pandecode_prop("metadata = %s", a);
1143 free(a);
1144
1145 pandecode_prop("stride = %d", rt->afbc.stride);
1146
1147 pandecode_log(".flags = ");
1148 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1149 pandecode_log_cont(",\n");
1150
1151 pandecode_indent--;
1152 pandecode_log("},\n");
1153 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1154 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1155 rt->afbc.metadata,
1156 rt->afbc.stride,
1157 rt->afbc.flags);
1158 }
1159
1160 MEMORY_PROP(rt, framebuffer);
1161 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1162
1163 if (rt->layer_stride)
1164 pandecode_prop("layer_stride = %d", rt->layer_stride);
1165
1166 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1167 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1168 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1169 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1170 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1171 }
1172
1173 if (rt->zero1 || rt->zero2) {
1174 pandecode_msg("XXX: render target zeros tripped\n");
1175 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1176 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1177 }
1178
1179 pandecode_indent--;
1180 pandecode_log("},\n");
1181 }
1182
1183 pandecode_indent--;
1184 pandecode_log("};\n");
1185 }
1186
1187 static struct pandecode_fbd
1188 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1189 {
1190 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1191 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1192
1193 struct pandecode_fbd info;
1194
1195 if (is_bifrost && fb->msaa.sample_locations) {
1196 /* The blob stores all possible sample locations in a single buffer
1197 * allocated on startup, and just switches the pointer when switching
1198 * MSAA state. For now, we just put the data into the cmdstream, but we
1199 * should do something like what the blob does with a real driver.
1200 *
1201 * There seem to be 32 slots for sample locations, followed by another
1202 * 16. The second 16 is just the center location followed by 15 zeros
1203 * in all the cases I've identified (maybe shader vs. depth/color
1204 * samples?).
1205 */
1206
1207 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1208
1209 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1210
1211 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1212 pandecode_indent++;
1213
1214 for (int i = 0; i < 32 + 16; i++) {
1215 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1216 }
1217
1218 pandecode_indent--;
1219 pandecode_log("};\n");
1220 }
1221
1222 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1223 pandecode_indent++;
1224
1225 if (is_bifrost) {
1226 pandecode_log(".msaa = {\n");
1227 pandecode_indent++;
1228
1229 if (fb->msaa.sample_locations)
1230 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1231 else
1232 pandecode_msg("XXX: sample_locations missing\n");
1233
1234 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1235 pandecode_msg("XXX: multisampling zero tripped\n");
1236 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1237 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1238 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1239 }
1240
1241 pandecode_indent--;
1242 pandecode_log("},\n");
1243 } else {
1244 pandecode_log(".shared_memory = {\n");
1245 pandecode_indent++;
1246 pandecode_shared_memory(&fb->shared_memory, is_compute);
1247 pandecode_indent--;
1248 pandecode_log("},\n");
1249 }
1250
1251 info.width = fb->width1 + 1;
1252 info.height = fb->height1 + 1;
1253 info.rt_count = fb->rt_count_1 + 1;
1254
1255 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1256 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1257 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1258 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1259
1260 pandecode_prop("unk1 = 0x%x", fb->unk1);
1261 pandecode_prop("unk2 = 0x%x", fb->unk2);
1262 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1263 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1264
1265 pandecode_log(".mfbd_flags = ");
1266 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1267 pandecode_log_cont(",\n");
1268
1269 if (fb->clear_stencil)
1270 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1271
1272 if (fb->clear_depth)
1273 pandecode_prop("clear_depth = %f", fb->clear_depth);
1274
1275 if (!is_compute)
1276 if (is_bifrost)
1277 pandecode_bifrost_tiler_descriptor(fb);
1278 else {
1279 const struct midgard_tiler_descriptor t = fb->tiler;
1280 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1281 }
1282 else
1283 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1284
1285 if (fb->zero3 || fb->zero4) {
1286 pandecode_msg("XXX: framebuffer zeros tripped\n");
1287 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1288 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1289 }
1290
1291 pandecode_indent--;
1292 pandecode_log("};\n");
1293
1294 gpu_va += sizeof(struct mali_framebuffer);
1295
1296 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1297
1298 if (info.has_extra) {
1299 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1300 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1301
1302 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1303 pandecode_indent++;
1304
1305 MEMORY_PROP(fbx, checksum);
1306
1307 if (fbx->checksum_stride)
1308 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1309
1310 pandecode_log(".flags_hi = ");
1311 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1312 pandecode_log_cont(",\n");
1313
1314 pandecode_log(".flags_lo = ");
1315 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1316 pandecode_log_cont(",\n");
1317
1318 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1319 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1320
1321 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1322 pandecode_log(".ds_afbc = {\n");
1323 pandecode_indent++;
1324
1325 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1326 pandecode_prop("depth_stencil_afbc_stride = %d",
1327 fbx->ds_afbc.depth_stencil_afbc_stride);
1328 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1329
1330 pandecode_log(".flags = ");
1331 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1332 pandecode_log_cont(",\n");
1333
1334 if (fbx->ds_afbc.padding) {
1335 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1336 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1337 }
1338
1339 pandecode_indent--;
1340 pandecode_log("},\n");
1341 } else {
1342 pandecode_log(".ds_linear = {\n");
1343 pandecode_indent++;
1344
1345 if (fbx->ds_linear.depth) {
1346 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1347 pandecode_prop("depth_stride = %d",
1348 fbx->ds_linear.depth_stride);
1349 pandecode_prop("depth_layer_stride = %d",
1350 fbx->ds_linear.depth_layer_stride);
1351 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1352 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1353 }
1354
1355 if (fbx->ds_linear.stencil) {
1356 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1357 pandecode_prop("stencil_stride = %d",
1358 fbx->ds_linear.stencil_stride);
1359 pandecode_prop("stencil_layer_stride = %d",
1360 fbx->ds_linear.stencil_layer_stride);
1361 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1362 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1363 }
1364
1365 if (fbx->ds_linear.depth_stride_zero ||
1366 fbx->ds_linear.stencil_stride_zero) {
1367 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1368 pandecode_prop("depth_stride_zero = 0x%x",
1369 fbx->ds_linear.depth_stride_zero);
1370 pandecode_prop("stencil_stride_zero = 0x%x",
1371 fbx->ds_linear.stencil_stride_zero);
1372 }
1373
1374 pandecode_indent--;
1375 pandecode_log("},\n");
1376 }
1377
1378 if (fbx->clear_color_1 | fbx->clear_color_2) {
1379 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1380 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1381 }
1382
1383 if (fbx->zero3) {
1384 pandecode_msg("XXX: fb_extra zeros tripped\n");
1385 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1386 }
1387
1388 pandecode_indent--;
1389 pandecode_log("};\n");
1390
1391 gpu_va += sizeof(struct mali_framebuffer_extra);
1392 }
1393
1394 if (is_fragment)
1395 pandecode_render_target(gpu_va, job_no, fb);
1396
1397 return info;
1398 }
1399
1400 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1401 * count */
1402
1403 static void
1404 pandecode_padded_vertices(unsigned shift, unsigned k)
1405 {
1406 unsigned odd = 2*k + 1;
1407 unsigned pot = 1 << shift;
1408 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1409 }
1410
1411 /* Given a magic divisor, recover what we were trying to divide by.
1412 *
1413 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1414 * 0 <= m < 2^N, for N bits in m.
1415 *
1416 * Let q represent the number we would like to divide by.
1417 *
1418 * By definition of a magic divisor for N-bit unsigned integers (a number you
1419 * multiply by to magically get division), m is a number such that:
1420 *
1421 * (m * x) & (2^N - 1) = floor(x/q).
1422 * for all x on Z where 0 <= x < 2^N
1423 *
1424 * Ignore the case where any of the above values equals zero; it is irrelevant
1425 * for our purposes (instanced arrays).
1426 *
1427 * Choose x = q. Then:
1428 *
1429 * (m * x) & (2^N - 1) = floor(x/q).
1430 * (m * q) & (2^N - 1) = floor(q/q).
1431 *
1432 * floor(q/q) = floor(1) = 1, therefore:
1433 *
1434 * (m * q) & (2^N - 1) = 1
1435 *
1436 * Recall the identity that the bitwise AND of one less than a power-of-two
1437 * equals the modulo with that power of two, i.e. for all x:
1438 *
1439 * x & (2^N - 1) = x % N
1440 *
1441 * Therefore:
1442 *
1443 * mq % (2^N) = 1
1444 *
1445 * By definition, a modular multiplicative inverse of a number m is the number
1446 * q such that with respect to a modulos M:
1447 *
1448 * mq % M = 1
1449 *
1450 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1451 *
1452 */
1453
1454 static void
1455 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1456 {
1457 #if 0
1458 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1459 * shift) the most lame way possible... just repeatedly add.
1460 * Asymptoptically slow but nobody cares in practice, unless you have
1461 * massive numbers of vertices or high divisors. */
1462
1463 unsigned inverse = 0;
1464
1465 /* Magic implicitly has the highest bit set */
1466 magic |= (1 << 31);
1467
1468 /* Depending on rounding direction */
1469 if (extra)
1470 magic++;
1471
1472 for (;;) {
1473 uint32_t product = magic * inverse;
1474
1475 if (shift) {
1476 product >>= shift;
1477 }
1478
1479 if (product == 1)
1480 break;
1481
1482 ++inverse;
1483 }
1484
1485 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1486
1487 /* Recall we're supposed to divide by (gl_level_divisor *
1488 * padded_num_vertices) */
1489
1490 unsigned padded_num_vertices = inverse / orig_divisor;
1491
1492 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1493 #endif
1494 }
1495
1496 static void
1497 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1498 mali_ptr addr, int job_no, char *suffix,
1499 int count, bool varying, enum mali_job_type job_type)
1500 {
1501 char *prefix = varying ? "varying" : "attribute";
1502 assert(addr);
1503
1504 if (!count) {
1505 pandecode_msg("warn: No %s records\n", prefix);
1506 return;
1507 }
1508
1509 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1510
1511 for (int i = 0; i < count; ++i) {
1512 /* First, check for special records */
1513 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1514 if (attr[i].size)
1515 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1516
1517 if (attr[i].stride) {
1518 /* gl_InstanceID passes a magic divisor in the
1519 * stride field to divide by the padded vertex
1520 * count. No other records should do so, so
1521 * stride should otherwise be zero. Note that
1522 * stride in the usual attribute sense doesn't
1523 * apply to special records. */
1524
1525 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1526
1527 pandecode_log_cont("/* %smagic divisor = %X */ ",
1528 has_divisor ? "" : "XXX: ", attr[i].stride);
1529 }
1530
1531 if (attr[i].shift || attr[i].extra_flags) {
1532 /* Attributes use these fields for
1533 * instancing/padding/etc type issues, but
1534 * varyings don't */
1535
1536 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1537 varying ? "XXX: " : "",
1538 attr[i].shift, attr[i].extra_flags);
1539 }
1540
1541 /* Print the special record name */
1542 bool attribute = false;
1543 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1544
1545 /* Sanity check */
1546 if (attribute == varying)
1547 pandecode_msg("XXX: mismatched special record\n");
1548
1549 continue;
1550 }
1551
1552 enum mali_attr_mode mode = attr[i].elements & 7;
1553
1554 if (mode == MALI_ATTR_UNUSED)
1555 pandecode_msg("XXX: unused attribute record\n");
1556
1557 /* For non-linear records, we need to print the type of record */
1558 if (mode != MALI_ATTR_LINEAR)
1559 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1560
1561 /* Print the name to link with attr_meta */
1562 pandecode_log_cont("%s_%d", prefix, i);
1563
1564 /* Print the stride and size */
1565 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1566
1567 /* TODO: Sanity check the quotient itself. It must be equal to
1568 * (or be greater than, if the driver added padding) the padded
1569 * vertex count. */
1570
1571 /* Finally, print the pointer */
1572 mali_ptr raw_elements = attr[i].elements & ~7;
1573 char *a = pointer_as_memory_reference(raw_elements);
1574 pandecode_log_cont(" = (%s);\n", a);
1575 free(a);
1576
1577 /* Check the pointer */
1578 pandecode_validate_buffer(raw_elements, attr[i].size);
1579
1580 /* shift/extra_flags exist only for instanced */
1581 if (attr[i].shift | attr[i].extra_flags) {
1582 /* These are set to random values by the blob for
1583 * varyings, most likely a symptom of uninitialized
1584 * memory where the hardware masked the bug. As such we
1585 * put this at a warning, not an error. */
1586
1587 if (mode == MALI_ATTR_LINEAR)
1588 pandecode_msg("warn: instancing fields set for linear\n");
1589
1590 pandecode_prop("shift = %d", attr[i].shift);
1591 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1592 }
1593
1594 /* Decode further where possible */
1595
1596 if (mode == MALI_ATTR_MODULO) {
1597 pandecode_padded_vertices(
1598 attr[i].shift,
1599 attr[i].extra_flags);
1600 }
1601
1602 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1603 i++;
1604 pandecode_log("{\n");
1605 pandecode_indent++;
1606 pandecode_prop("unk = 0x%x", attr[i].unk);
1607 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1608 if (attr[i].zero != 0)
1609 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1610 pandecode_prop("divisor = %d", attr[i].divisor);
1611 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1612 pandecode_indent--;
1613 pandecode_log("}, \n");
1614 }
1615
1616 }
1617
1618 pandecode_log("\n");
1619 }
1620
1621 static mali_ptr
1622 pandecode_shader_address(const char *name, mali_ptr ptr)
1623 {
1624 /* TODO: Decode flags */
1625 mali_ptr shader_ptr = ptr & ~15;
1626
1627 char *a = pointer_as_memory_reference(shader_ptr);
1628 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1629 free(a);
1630
1631 return shader_ptr;
1632 }
1633
1634 static void
1635 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1636 {
1637 unsigned any_nonzero =
1638 stencil->ref | stencil->mask | stencil->func |
1639 stencil->sfail | stencil->dpfail | stencil->dppass;
1640
1641 if (any_nonzero == 0)
1642 return;
1643
1644 const char *func = mali_func_as_str(stencil->func);
1645 const char *sfail = pandecode_stencil_op(stencil->sfail);
1646 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1647 const char *dppass = pandecode_stencil_op(stencil->dppass);
1648
1649 if (stencil->zero)
1650 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1651
1652 pandecode_log(".stencil_%s = {\n", name);
1653 pandecode_indent++;
1654 pandecode_prop("ref = %d", stencil->ref);
1655 pandecode_prop("mask = 0x%02X", stencil->mask);
1656 pandecode_prop("func = %s", func);
1657 pandecode_prop("sfail = %s", sfail);
1658 pandecode_prop("dpfail = %s", dpfail);
1659 pandecode_prop("dppass = %s", dppass);
1660 pandecode_indent--;
1661 pandecode_log("},\n");
1662 }
1663
1664 static void
1665 pandecode_blend_equation(const struct mali_blend_equation *blend)
1666 {
1667 if (blend->zero1)
1668 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1669
1670 pandecode_log(".equation = {\n");
1671 pandecode_indent++;
1672
1673 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1674 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1675
1676 pandecode_log(".color_mask = ");
1677 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1678 pandecode_log_cont(",\n");
1679
1680 pandecode_indent--;
1681 pandecode_log("},\n");
1682 }
1683
1684 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1685
1686 static unsigned
1687 decode_bifrost_constant(u16 constant)
1688 {
1689 float lo = (float) (constant & 0xFF);
1690 float hi = (float) (constant >> 8);
1691
1692 return (hi / 255.0) + (lo / 65535.0);
1693 }
1694
1695 static mali_ptr
1696 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1697 {
1698 struct bifrost_blend_rt *b =
1699 ((struct bifrost_blend_rt *) descs) + rt_no;
1700
1701 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1702 pandecode_indent++;
1703
1704 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1705 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1706 b->constant, decode_bifrost_constant(b->constant));
1707
1708 /* TODO figure out blend shader enable bit */
1709 pandecode_blend_equation(&b->equation);
1710
1711 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1712 pandecode_prop("index = 0x%" PRIx16, b->index);
1713
1714 pandecode_log(".format = ");
1715 pandecode_format_short(b->format, false);
1716 pandecode_swizzle(b->swizzle, b->format);
1717 pandecode_log_cont(",\n");
1718
1719 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1720 pandecode_prop("format = 0x%" PRIx32, b->format);
1721
1722 if (b->zero1) {
1723 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1724 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1725 }
1726
1727 pandecode_log(".shader_type = ");
1728 switch(b->shader_type) {
1729 case BIFROST_BLEND_F16:
1730 pandecode_log_cont("BIFROST_BLEND_F16");
1731 break;
1732 case BIFROST_BLEND_F32:
1733 pandecode_log_cont("BIFROST_BLEND_F32");
1734 break;
1735 case BIFROST_BLEND_I32:
1736 pandecode_log_cont("BIFROST_BLEND_I32");
1737 break;
1738 case BIFROST_BLEND_U32:
1739 pandecode_log_cont("BIFROST_BLEND_U32");
1740 break;
1741 case BIFROST_BLEND_I16:
1742 pandecode_log_cont("BIFROST_BLEND_I16");
1743 break;
1744 case BIFROST_BLEND_U16:
1745 pandecode_log_cont("BIFROST_BLEND_U16");
1746 break;
1747 }
1748 pandecode_log_cont(",\n");
1749
1750 if (b->zero2) {
1751 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1752 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1753 }
1754
1755 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1756
1757 pandecode_indent--;
1758 pandecode_log("},\n");
1759
1760 return 0;
1761 }
1762
1763 static mali_ptr
1764 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1765 {
1766 /* constant/equation is in a union */
1767 if (!blend->shader)
1768 return 0;
1769
1770 pandecode_log(".blend = {\n");
1771 pandecode_indent++;
1772
1773 if (is_shader) {
1774 pandecode_shader_address("shader", blend->shader);
1775 } else {
1776 pandecode_blend_equation(&blend->equation);
1777 pandecode_prop("constant = %f", blend->constant);
1778 }
1779
1780 pandecode_indent--;
1781 pandecode_log("},\n");
1782
1783 /* Return blend shader to disassemble if present */
1784 return is_shader ? (blend->shader & ~0xF) : 0;
1785 }
1786
1787 static mali_ptr
1788 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1789 {
1790 struct midgard_blend_rt *b =
1791 ((struct midgard_blend_rt *) descs) + rt_no;
1792
1793 /* Flags determine presence of blend shader */
1794 bool is_shader = (b->flags & 0xF) >= 0x2;
1795
1796 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1797 pandecode_indent++;
1798
1799 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1800
1801 union midgard_blend blend = b->blend;
1802 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1803
1804 pandecode_indent--;
1805 pandecode_log("};\n");
1806
1807 return shader;
1808 }
1809
1810 /* Attributes and varyings have descriptor records, which contain information
1811 * about their format and ordering with the attribute/varying buffers. We'll
1812 * want to validate that the combinations specified are self-consistent.
1813 */
1814
1815 static int
1816 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1817 {
1818 char base[128];
1819 char *prefix = varying ? "varying" : "attribute";
1820 unsigned max_index = 0;
1821 snprintf(base, sizeof(base), "%s_meta", prefix);
1822
1823 struct mali_attr_meta *attr_meta;
1824 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1825
1826 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1827
1828 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1829 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1830 sizeof(*attr_mem));
1831
1832 /* If the record is discard, it should be zero for everything else */
1833
1834 if (attr_meta->format == MALI_VARYING_DISCARD) {
1835 uint64_t zero =
1836 attr_meta->index |
1837 attr_meta->unknown1 |
1838 attr_meta->unknown3 |
1839 attr_meta->src_offset;
1840
1841 if (zero)
1842 pandecode_msg("XXX: expected empty record for varying discard\n");
1843
1844 /* We want to look for a literal 0000 swizzle -- this
1845 * is not encoded with all zeroes, however */
1846
1847 enum mali_channel z = MALI_CHANNEL_ZERO;
1848 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1849 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1850
1851 if (!good_swizzle)
1852 pandecode_msg("XXX: expected zero swizzle for discard\n");
1853
1854 if (!varying)
1855 pandecode_msg("XXX: cannot discard attribute\n");
1856
1857 /* If we're all good, omit the record */
1858 if (!zero && varying && good_swizzle) {
1859 pandecode_log("/* discarded varying */\n");
1860 continue;
1861 }
1862 }
1863
1864 if (attr_meta->index > max_index)
1865 max_index = attr_meta->index;
1866
1867 if (attr_meta->unknown1 != 0x2) {
1868 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1869 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1870 }
1871
1872 if (attr_meta->unknown3) {
1873 pandecode_msg("XXX: unexpected unknown3 set\n");
1874 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1875 }
1876
1877 pandecode_format_short(attr_meta->format, false);
1878 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1879
1880 if (attr_meta->src_offset)
1881 pandecode_log_cont("[%u]", attr_meta->src_offset);
1882
1883 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1884
1885 pandecode_log_cont(";\n");
1886 }
1887
1888 pandecode_log("\n");
1889
1890 return count ? (max_index + 1) : 0;
1891 }
1892
1893 /* return bits [lo, hi) of word */
1894 static u32
1895 bits(u32 word, u32 lo, u32 hi)
1896 {
1897 if (hi - lo >= 32)
1898 return word; // avoid undefined behavior with the shift
1899
1900 return (word >> lo) & ((1 << (hi - lo)) - 1);
1901 }
1902
1903 static void
1904 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1905 {
1906 pandecode_log(".prefix = {\n");
1907 pandecode_indent++;
1908
1909 /* Decode invocation_count. See the comment before the definition of
1910 * invocation_count for an explanation.
1911 */
1912
1913 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1914 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1915 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1916 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1917 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1918 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1919
1920 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1921 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1922 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1923
1924 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1925 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1926 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1927
1928 /* Even though we have this decoded, we want to ensure that the
1929 * representation is "unique" so we don't lose anything by printing only
1930 * the final result. More specifically, we need to check that we were
1931 * passed something in canonical form, since the definition per the
1932 * hardware is inherently not unique. How? Well, take the resulting
1933 * decode and pack it ourselves! If it is bit exact with what we
1934 * decoded, we're good to go. */
1935
1936 struct mali_vertex_tiler_prefix ref;
1937 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1938
1939 bool canonical =
1940 (p->invocation_count == ref.invocation_count) &&
1941 (p->invocation_shifts == ref.invocation_shifts);
1942
1943 if (!canonical) {
1944 pandecode_msg("XXX: non-canonical workgroups packing\n");
1945 pandecode_msg("expected: %X, %X",
1946 ref.invocation_count,
1947 ref.invocation_shifts);
1948
1949 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1950 pandecode_prop("size_y_shift = %d", size_y_shift);
1951 pandecode_prop("size_z_shift = %d", size_z_shift);
1952 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1953 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1954 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1955 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1956 }
1957
1958 /* Regardless, print the decode */
1959 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1960 size_x, size_y, size_z,
1961 groups_x, groups_y, groups_z);
1962
1963 /* TODO: Decode */
1964 if (p->unknown_draw)
1965 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1966
1967 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1968
1969 if (p->draw_mode != MALI_DRAW_MODE_NONE)
1970 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p->draw_mode));
1971
1972 /* Index count only exists for tiler jobs anyway */
1973
1974 if (p->index_count)
1975 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1976
1977
1978 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
1979 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
1980
1981 /* Validate an index buffer is present if we need one. TODO: verify
1982 * relationship between invocation_count and index_count */
1983
1984 if (p->indices) {
1985 unsigned count = p->index_count;
1986
1987 /* Grab the size */
1988 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
1989
1990 /* Ensure we got a size, and if so, validate the index buffer
1991 * is large enough to hold a full set of indices of the given
1992 * size */
1993
1994 if (!index_raw_size)
1995 pandecode_msg("XXX: index size missing\n");
1996 else
1997 pandecode_validate_buffer(p->indices, count * size);
1998 } else if (index_raw_size)
1999 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
2000
2001 if (p->offset_bias_correction)
2002 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
2003
2004 /* TODO: Figure out what this is. It's not zero */
2005 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
2006
2007 pandecode_indent--;
2008 pandecode_log("},\n");
2009 }
2010
2011 static void
2012 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
2013 {
2014 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
2015 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
2016
2017 for (int i = 0; i < ubufs_count; i++) {
2018 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
2019 mali_ptr addr = (ubufs[i] >> 10) << 2;
2020
2021 pandecode_validate_buffer(addr, size);
2022
2023 char *ptr = pointer_as_memory_reference(addr);
2024 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
2025 free(ptr);
2026 }
2027
2028 pandecode_log("\n");
2029 }
2030
2031 static void
2032 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
2033 {
2034 pandecode_validate_buffer(uniforms, uniform_count * 16);
2035
2036 char *ptr = pointer_as_memory_reference(uniforms);
2037 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
2038 free(ptr);
2039 }
2040
2041 static const char *
2042 shader_type_for_job(unsigned type)
2043 {
2044 switch (type) {
2045 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
2046 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
2047 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
2048 default:
2049 return "UNKNOWN";
2050 }
2051 }
2052
2053 static unsigned shader_id = 0;
2054
2055 static struct midgard_disasm_stats
2056 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
2057 bool is_bifrost, unsigned gpu_id)
2058 {
2059 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
2060 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
2061
2062 /* Compute maximum possible size */
2063 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
2064
2065 /* Print some boilerplate to clearly denote the assembly (which doesn't
2066 * obey indentation rules), and actually do the disassembly! */
2067
2068 pandecode_log_cont("\n\n");
2069
2070 struct midgard_disasm_stats stats;
2071
2072 if (is_bifrost) {
2073 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
2074
2075 /* TODO: Extend stats to Bifrost */
2076 stats.texture_count = -128;
2077 stats.sampler_count = -128;
2078 stats.attribute_count = -128;
2079 stats.varying_count = -128;
2080 stats.uniform_count = -128;
2081 stats.uniform_buffer_count = -128;
2082 stats.work_count = -128;
2083
2084 stats.instruction_count = 0;
2085 stats.bundle_count = 0;
2086 stats.quadword_count = 0;
2087 stats.helper_invocations = false;
2088 } else {
2089 stats = disassemble_midgard(pandecode_dump_stream,
2090 code, sz, gpu_id,
2091 type == MALI_JOB_TYPE_TILER ?
2092 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
2093 }
2094
2095 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2096 * driver-internal purposes with the blob and interfere */
2097
2098 bool should_shaderdb = type != MALI_JOB_TYPE_COMPUTE;
2099
2100 if (should_shaderdb) {
2101 unsigned nr_threads =
2102 (stats.work_count <= 4) ? 4 :
2103 (stats.work_count <= 8) ? 2 :
2104 1;
2105
2106 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2107 "%u inst, %u bundles, %u quadwords, "
2108 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2109 shader_id++,
2110 shader_type_for_job(type),
2111 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2112 stats.work_count, nr_threads);
2113 }
2114
2115
2116 return stats;
2117 }
2118
2119 static void
2120 pandecode_texture_payload(mali_ptr payload,
2121 enum mali_texture_type type,
2122 enum mali_texture_layout layout,
2123 bool manual_stride,
2124 uint8_t levels,
2125 uint16_t depth,
2126 uint16_t array_size,
2127 struct pandecode_mapped_memory *tmem)
2128 {
2129 pandecode_log(".payload = {\n");
2130 pandecode_indent++;
2131
2132 /* A bunch of bitmap pointers follow.
2133 * We work out the correct number,
2134 * based on the mipmap/cubemap
2135 * properties, but dump extra
2136 * possibilities to futureproof */
2137
2138 int bitmap_count = levels + 1;
2139
2140 /* Miptree for each face */
2141 if (type == MALI_TEX_CUBE)
2142 bitmap_count *= 6;
2143
2144 /* Array of layers */
2145 bitmap_count *= (depth + 1);
2146
2147 /* Array of textures */
2148 bitmap_count *= (array_size + 1);
2149
2150 /* Stride for each element */
2151 if (manual_stride)
2152 bitmap_count *= 2;
2153
2154 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2155 payload, sizeof(mali_ptr) * bitmap_count);
2156 for (int i = 0; i < bitmap_count; ++i) {
2157 /* How we dump depends if this is a stride or a pointer */
2158
2159 if (manual_stride && (i & 1)) {
2160 /* signed 32-bit snuck in as a 64-bit pointer */
2161 uint64_t stride_set = pointers_and_strides[i];
2162 uint32_t clamped_stride = stride_set;
2163 int32_t stride = clamped_stride;
2164 assert(stride_set == clamped_stride);
2165 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2166 } else {
2167 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2168 pandecode_log("%s, \n", a);
2169 free(a);
2170 }
2171 }
2172
2173 pandecode_indent--;
2174 pandecode_log("},\n");
2175 }
2176
2177 static void
2178 pandecode_texture(mali_ptr u,
2179 struct pandecode_mapped_memory *tmem,
2180 unsigned job_no, unsigned tex)
2181 {
2182 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2183
2184 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2185 pandecode_indent++;
2186
2187 pandecode_prop("width = %" PRId32, t->width);
2188 pandecode_prop("height = %" PRId32, t->height);
2189 pandecode_prop("depth = %" PRId32, t->depth);
2190 pandecode_prop("array_size = %" PRId32, t->array_size);
2191
2192 pandecode_log("\n");
2193 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2194 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2195 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2196 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2197 pandecode_prop("f.type = %" PRId32, t->format.type);
2198 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2199 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2200 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2201 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2202 pandecode_log("\n");
2203
2204 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2205 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2206 pandecode_prop("levels = %" PRId32, t->levels);
2207 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2208 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2209
2210 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2211 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2212 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2213 pandecode_log("\n");
2214
2215 struct mali_texture_format f = t->format;
2216
2217 /* See the definiton of enum mali_texture_type */
2218
2219 bool is_cube = f.type == MALI_TEX_CUBE;
2220 unsigned dimension = is_cube ? 2 : f.type;
2221
2222 pandecode_make_indent();
2223
2224 /* Print the layout. Default is linear; a modifier can denote AFBC or
2225 * u-interleaved/tiled modes */
2226
2227 if (f.layout == MALI_TEXTURE_AFBC)
2228 pandecode_log_cont("afbc");
2229 else if (f.layout == MALI_TEXTURE_TILED)
2230 pandecode_log_cont("tiled");
2231 else if (f.layout == MALI_TEXTURE_LINEAR)
2232 pandecode_log_cont("linear");
2233 else
2234 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2235
2236 pandecode_swizzle(t->swizzle, f.format);
2237 pandecode_log_cont(" ");
2238
2239 /* Distinguish cube/2D with modifier */
2240
2241 if (is_cube)
2242 pandecode_log_cont("cube ");
2243
2244 pandecode_format_short(f.format, f.srgb);
2245 pandecode_swizzle(f.swizzle, f.format);
2246
2247 /* All four width/height/depth/array_size dimensions are present
2248 * regardless of the type of texture, but it is an error to have
2249 * non-zero dimensions for unused dimensions. Verify this. array_size
2250 * can always be set, as can width. Depth used for MSAA. */
2251
2252 if (t->height && dimension < 2)
2253 pandecode_msg("XXX: nonzero height for <2D texture\n");
2254
2255 /* Print only the dimensions that are actually there */
2256
2257 pandecode_log_cont(": %d", t->width + 1);
2258
2259 if (t->height || t->depth)
2260 pandecode_log_cont("x%u", t->height + 1);
2261
2262 if (t->depth)
2263 pandecode_log_cont("x%u", t->depth + 1);
2264
2265 if (t->array_size)
2266 pandecode_log_cont("[%u]", t->array_size + 1);
2267
2268 if (t->levels)
2269 pandecode_log_cont(" mip %u", t->levels);
2270
2271 pandecode_log_cont("\n");
2272
2273 if (f.unknown1 | f.zero) {
2274 pandecode_msg("XXX: texture format zero tripped\n");
2275 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2276 pandecode_prop("zero = %" PRId32, f.zero);
2277 }
2278
2279 if (!f.unknown2) {
2280 pandecode_msg("XXX: expected unknown texture bit set\n");
2281 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2282 }
2283
2284 if (t->swizzle_zero) {
2285 pandecode_msg("XXX: swizzle zero tripped\n");
2286 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2287 }
2288
2289 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2290 pandecode_msg("XXX: texture zero tripped\n");
2291 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2292 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2293 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2294 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2295 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2296 }
2297
2298 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2299
2300 pandecode_indent--;
2301 pandecode_log("};\n");
2302 }
2303
2304 static void
2305 pandecode_bifrost_texture(
2306 const struct bifrost_texture_descriptor *t,
2307 unsigned job_no,
2308 unsigned tex)
2309 {
2310 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2311 pandecode_indent++;
2312
2313 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2314 pandecode_prop("type = %" PRId32, t->type);
2315
2316 if (t->zero) {
2317 pandecode_msg("XXX: zero tripped\n");
2318 pandecode_prop("zero = 0x%" PRIx32, t->zero);
2319 }
2320
2321 pandecode_prop("format_swizzle = 0x%" PRIx32, t->format_swizzle);
2322 pandecode_prop("format = 0x%" PRIx32, t->format);
2323 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2324 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2325 pandecode_prop("width = %" PRId32, t->width);
2326 pandecode_prop("height = %" PRId32, t->height);
2327 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2328 pandecode_prop("levels = %" PRId32, t->levels);
2329 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2330 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2331 pandecode_prop("level_2 = %" PRId32, t->level_2);
2332 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2333 pandecode_prop("array_size = %" PRId32, t->array_size);
2334 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2335 pandecode_prop("depth = %" PRId32, t->depth);
2336 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2337 pandecode_log("\n");
2338
2339 /* See the definiton of enum mali_texture_type */
2340
2341 bool is_cube = t->type == MALI_TEX_CUBE;
2342 unsigned dimension = is_cube ? 2 : t->type;
2343
2344 /* Print the layout. Default is linear; a modifier can denote AFBC or
2345 * u-interleaved/tiled modes */
2346
2347 if (t->layout == MALI_TEXTURE_AFBC)
2348 pandecode_log_cont("afbc");
2349 else if (t->layout == MALI_TEXTURE_TILED)
2350 pandecode_log_cont("tiled");
2351 else if (t->layout == MALI_TEXTURE_LINEAR)
2352 pandecode_log_cont("linear");
2353 else
2354 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2355
2356 pandecode_swizzle(t->swizzle, t->format);
2357 pandecode_log_cont(" ");
2358
2359 /* Distinguish cube/2D with modifier */
2360
2361 if (is_cube)
2362 pandecode_log_cont("cube ");
2363
2364 pandecode_format_short(t->format, t->srgb);
2365
2366 /* All four width/height/depth/array_size dimensions are present
2367 * regardless of the type of texture, but it is an error to have
2368 * non-zero dimensions for unused dimensions. Verify this. array_size
2369 * can always be set, as can width. */
2370
2371 if (t->height && dimension < 2)
2372 pandecode_msg("XXX: nonzero height for <2D texture\n");
2373
2374 if (t->depth && dimension < 3)
2375 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2376
2377 /* Print only the dimensions that are actually there */
2378
2379 pandecode_log_cont(": %d", t->width + 1);
2380
2381 if (dimension >= 2)
2382 pandecode_log_cont("x%u", t->height + 1);
2383
2384 if (dimension >= 3)
2385 pandecode_log_cont("x%u", t->depth + 1);
2386
2387 if (t->array_size)
2388 pandecode_log_cont("[%u]", t->array_size + 1);
2389
2390 if (t->levels)
2391 pandecode_log_cont(" mip %u", t->levels);
2392
2393 pandecode_log_cont("\n");
2394
2395 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2396 if (t->payload) {
2397 pandecode_texture_payload(t->payload, t->type, t->layout,
2398 true, t->levels, t->depth,
2399 t->array_size, tmem);
2400 }
2401
2402 pandecode_indent--;
2403 pandecode_log("};\n");
2404 }
2405
2406 /* 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 */
2407
2408 static void
2409 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2410 {
2411 /* Nothing to do */
2412 if (claim == truth)
2413 return;
2414
2415 if (fuzzy && (truth < 0))
2416 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2417
2418 if ((truth >= 0) && !fuzzy) {
2419 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2420 (truth < claim) ? "warn" : "XXX",
2421 name, truth, claim);
2422 } else if ((claim > -truth) && !fuzzy) {
2423 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2424 name, -truth, claim);
2425 } else if (fuzzy && (claim < truth))
2426 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2427 name, truth, claim);
2428
2429 pandecode_log(".%s = %" PRId16, name, claim);
2430
2431 if (fuzzy)
2432 pandecode_log_cont(" /* %u used */", truth);
2433
2434 pandecode_log_cont(",\n");
2435 }
2436
2437 static void
2438 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2439 bool is_bifrost, unsigned gpu_id)
2440 {
2441 struct midgard_disasm_stats stats =
2442 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2443
2444 bool has_texture = (stats.texture_count > 0);
2445 bool has_sampler = (stats.sampler_count > 0);
2446 bool has_attribute = (stats.attribute_count > 0);
2447 bool has_varying = (stats.varying_count > 0);
2448 bool has_uniform = (stats.uniform_count > 0);
2449 bool has_ubo = (stats.uniform_buffer_count > 0);
2450
2451 if (has_texture || has_sampler)
2452 pandecode_msg("XXX: blend shader accessing textures\n");
2453
2454 if (has_attribute || has_varying)
2455 pandecode_msg("XXX: blend shader accessing interstage\n");
2456
2457 if (has_uniform || has_ubo)
2458 pandecode_msg("XXX: blend shader accessing uniforms\n");
2459 }
2460
2461 static void
2462 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2463 {
2464 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2465
2466 if (!mmem)
2467 return;
2468
2469 if (is_bifrost) {
2470 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2471
2472 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2473 pandecode_indent++;
2474
2475 for (unsigned tex = 0; tex < texture_count; ++tex)
2476 pandecode_bifrost_texture(&t[tex], job_no, tex);
2477
2478 pandecode_indent--;
2479 pandecode_log("};\n");
2480 } else {
2481 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2482
2483 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2484 pandecode_indent++;
2485
2486 for (int tex = 0; tex < texture_count; ++tex) {
2487 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2488 char *a = pointer_as_memory_reference(*u);
2489 pandecode_log("%s,\n", a);
2490 free(a);
2491 }
2492
2493 pandecode_indent--;
2494 pandecode_log("};\n");
2495
2496 /* Now, finally, descend down into the texture descriptor */
2497 for (unsigned tex = 0; tex < texture_count; ++tex) {
2498 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2499 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2500 if (tmem)
2501 pandecode_texture(*u, tmem, job_no, tex);
2502 }
2503 }
2504 }
2505
2506 static void
2507 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2508 {
2509 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2510
2511 if (!smem)
2512 return;
2513
2514 if (is_bifrost) {
2515 struct bifrost_sampler_descriptor *s;
2516
2517 for (int i = 0; i < sampler_count; ++i) {
2518 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2519
2520 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2521 pandecode_indent++;
2522
2523 if (s->unk1 != 1) {
2524 pandecode_msg("XXX: unk1 tripped\n");
2525 pandecode_prop("unk1 = 0x%x", s->unk1);
2526 }
2527
2528 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2529 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2530 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2531
2532 if (s->unk8 != 0x8) {
2533 pandecode_msg("XXX: unk8 tripped\n");
2534 pandecode_prop("unk8 = 0x%x", s->unk8);
2535 }
2536
2537 pandecode_prop("unk2 = 0x%x", s->unk2);
2538 pandecode_prop("unk3 = 0x%x", s->unk3);
2539 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2540 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2541 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2542 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2543 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2544
2545 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2546 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2547
2548 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2549 pandecode_msg("XXX: sampler zero tripped\n");
2550 pandecode_prop("zero = 0x%" PRIx8 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2551 }
2552
2553 pandecode_indent--;
2554 pandecode_log("};\n");
2555 }
2556 } else {
2557 struct mali_sampler_descriptor *s;
2558
2559 for (int i = 0; i < sampler_count; ++i) {
2560 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2561
2562 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2563 pandecode_indent++;
2564
2565 pandecode_log(".filter_mode = ");
2566 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2567 pandecode_log_cont(",\n");
2568
2569 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2570 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2571
2572 if (s->lod_bias)
2573 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2574
2575 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2576 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2577 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2578
2579 pandecode_prop("compare_func = %s", mali_func_as_str(s->compare_func));
2580
2581 if (s->zero || s->zero2) {
2582 pandecode_msg("XXX: sampler zero tripped\n");
2583 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2584 }
2585
2586 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2587
2588 pandecode_prop("border_color = { %f, %f, %f, %f }",
2589 s->border_color[0],
2590 s->border_color[1],
2591 s->border_color[2],
2592 s->border_color[3]);
2593
2594 pandecode_indent--;
2595 pandecode_log("};\n");
2596 }
2597 }
2598 }
2599
2600 static void
2601 pandecode_vertex_tiler_postfix_pre(
2602 const struct mali_vertex_tiler_postfix *p,
2603 int job_no, enum mali_job_type job_type,
2604 char *suffix, bool is_bifrost, unsigned gpu_id)
2605 {
2606 struct pandecode_mapped_memory *attr_mem;
2607
2608 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2609 * are the only things actually needed from the FBD, vertex/tiler jobs
2610 * no longer reference the FBD -- instead, this field points to some
2611 * info about the scratchpad.
2612 */
2613
2614 struct pandecode_fbd fbd_info = {
2615 /* Default for Bifrost */
2616 .rt_count = 1
2617 };
2618
2619 if (is_bifrost) {
2620 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2621 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2622 } else if (p->shared_memory & MALI_MFBD)
2623 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
2624 else if (job_type == MALI_JOB_TYPE_COMPUTE)
2625 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2626 else
2627 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2628
2629 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2630 int texture_count = 0, sampler_count = 0;
2631
2632 if (p->shader) {
2633 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2634 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2635
2636 /* Disassemble ahead-of-time to get stats. Initialize with
2637 * stats for the missing-shader case so we get validation
2638 * there, too */
2639
2640 struct midgard_disasm_stats info = {
2641 .texture_count = 0,
2642 .sampler_count = 0,
2643 .attribute_count = 0,
2644 .varying_count = 0,
2645 .work_count = 1,
2646
2647 .uniform_count = -128,
2648 .uniform_buffer_count = 0
2649 };
2650
2651 if (s->shader & ~0xF)
2652 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2653
2654 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2655 pandecode_indent++;
2656
2657 /* Save for dumps */
2658 attribute_count = s->attribute_count;
2659 varying_count = s->varying_count;
2660 texture_count = s->texture_count;
2661 sampler_count = s->sampler_count;
2662
2663 if (is_bifrost) {
2664 uniform_count = s->bifrost2.uniform_count;
2665 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2666 } else {
2667 uniform_count = s->midgard1.uniform_count;
2668 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2669 }
2670
2671 pandecode_shader_address("shader", s->shader);
2672
2673 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2674 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2675 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2676 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2677 pandecode_shader_prop("uniform_buffer_count",
2678 uniform_buffer_count,
2679 info.uniform_buffer_count, true);
2680
2681 if (!is_bifrost) {
2682 pandecode_shader_prop("uniform_count",
2683 uniform_count,
2684 info.uniform_count, false);
2685
2686 pandecode_shader_prop("work_count",
2687 s->midgard1.work_count, info.work_count, false);
2688 }
2689
2690 if (is_bifrost) {
2691 pandecode_log("bifrost1.unk1 = ");
2692 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2693 pandecode_log_cont(",\n");
2694 } else {
2695 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2696
2697 if (helpers != info.helper_invocations) {
2698 pandecode_msg("XXX: expected helpers %u but got %u\n",
2699 info.helper_invocations, helpers);
2700 }
2701
2702 pandecode_log(".midgard1.flags_lo = ");
2703 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2704 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2705 pandecode_log_cont(",\n");
2706
2707 pandecode_log(".midgard1.flags_hi = ");
2708 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2709 pandecode_log_cont(",\n");
2710 }
2711
2712 if (s->depth_units || s->depth_factor) {
2713 pandecode_prop("depth_factor = %f", s->depth_factor);
2714 pandecode_prop("depth_units = %f", s->depth_units);
2715 }
2716
2717 if (s->coverage_mask)
2718 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
2719
2720 if (s->unknown2_2)
2721 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2722
2723 if (s->unknown2_3 || s->unknown2_4) {
2724 pandecode_log(".unknown2_3 = ");
2725
2726 int unknown2_3 = s->unknown2_3;
2727 int unknown2_4 = s->unknown2_4;
2728
2729 /* We're not quite sure what these flags mean without the depth test, if anything */
2730
2731 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2732 const char *func = mali_func_as_str(MALI_GET_DEPTH_FUNC(unknown2_3));
2733 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2734
2735 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2736 }
2737
2738 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2739 pandecode_log_cont(",\n");
2740
2741 pandecode_log(".unknown2_4 = ");
2742 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2743 pandecode_log_cont(",\n");
2744 }
2745
2746 if (s->stencil_mask_front || s->stencil_mask_back) {
2747 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2748 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2749 }
2750
2751 pandecode_stencil("front", &s->stencil_front);
2752 pandecode_stencil("back", &s->stencil_back);
2753
2754 if (is_bifrost) {
2755 pandecode_log(".bifrost2 = {\n");
2756 pandecode_indent++;
2757
2758 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2759 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2760 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2761 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2762
2763 pandecode_indent--;
2764 pandecode_log("},\n");
2765 } else if (s->midgard2.unknown2_7) {
2766 pandecode_log(".midgard2 = {\n");
2767 pandecode_indent++;
2768
2769 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2770 pandecode_indent--;
2771 pandecode_log("},\n");
2772 }
2773
2774 if (s->padding) {
2775 pandecode_msg("XXX: shader padding tripped\n");
2776 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2777 }
2778
2779 if (!is_bifrost) {
2780 /* TODO: Blend shaders routing/disasm */
2781 union midgard_blend blend = s->blend;
2782 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2783 if (shader & ~0xF)
2784 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2785 } else {
2786 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2787 }
2788
2789 pandecode_indent--;
2790 pandecode_log("};\n");
2791
2792 /* MRT blend fields are used whenever MFBD is used, with
2793 * per-RT descriptors */
2794
2795 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2796 void* blend_base = (void *) (s + 1);
2797
2798 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2799 mali_ptr shader = 0;
2800
2801 if (is_bifrost)
2802 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2803 else
2804 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2805
2806 if (shader & ~0xF)
2807 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2808
2809 }
2810 }
2811 } else
2812 pandecode_msg("XXX: missing shader descriptor\n");
2813
2814 if (p->viewport) {
2815 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2816 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2817
2818 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2819 pandecode_indent++;
2820
2821 pandecode_prop("clip_minx = %f", f->clip_minx);
2822 pandecode_prop("clip_miny = %f", f->clip_miny);
2823 pandecode_prop("clip_minz = %f", f->clip_minz);
2824 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2825 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2826 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2827
2828 /* Only the higher coordinates are MALI_POSITIVE scaled */
2829
2830 pandecode_prop("viewport0 = { %d, %d }",
2831 f->viewport0[0], f->viewport0[1]);
2832
2833 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2834 f->viewport1[0] + 1, f->viewport1[1] + 1);
2835
2836 pandecode_indent--;
2837 pandecode_log("};\n");
2838 }
2839
2840 unsigned max_attr_index = 0;
2841
2842 if (p->attribute_meta)
2843 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2844
2845 if (p->attributes) {
2846 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2847 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2848 }
2849
2850 /* Varyings are encoded like attributes but not actually sent; we just
2851 * pass a zero buffer with the right stride/size set, (or whatever)
2852 * since the GPU will write to it itself */
2853
2854 if (p->varying_meta) {
2855 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2856 }
2857
2858 if (p->varyings) {
2859 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2860
2861 /* Number of descriptors depends on whether there are
2862 * non-internal varyings */
2863
2864 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2865 }
2866
2867 if (p->uniform_buffers) {
2868 if (uniform_buffer_count)
2869 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2870 else
2871 pandecode_msg("warn: UBOs specified but not referenced\n");
2872 } else if (uniform_buffer_count)
2873 pandecode_msg("XXX: UBOs referenced but not specified\n");
2874
2875 /* We don't want to actually dump uniforms, but we do need to validate
2876 * that the counts we were given are sane */
2877
2878 if (p->uniforms) {
2879 if (uniform_count)
2880 pandecode_uniforms(p->uniforms, uniform_count);
2881 else
2882 pandecode_msg("warn: Uniforms specified but not referenced\n");
2883 } else if (uniform_count)
2884 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2885
2886 if (p->textures)
2887 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2888
2889 if (p->sampler_descriptor)
2890 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2891 }
2892
2893 static void
2894 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2895 {
2896 pandecode_log(".gl_enables = ");
2897
2898 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2899
2900 pandecode_log_cont(",\n");
2901 }
2902
2903 static void
2904 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2905 {
2906 if (p->shader & 0xF)
2907 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2908
2909 pandecode_log(".postfix = {\n");
2910 pandecode_indent++;
2911
2912 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
2913 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2914 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2915
2916 if (p->zero4) {
2917 pandecode_msg("XXX: vertex only zero tripped");
2918 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2919 }
2920
2921 pandecode_prop("offset_start = 0x%x", p->offset_start);
2922
2923 if (p->zero5) {
2924 pandecode_msg("XXX: vertex only zero tripped");
2925 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2926 }
2927
2928 MEMORY_PROP(p, position_varying);
2929 MEMORY_PROP(p, occlusion_counter);
2930
2931 pandecode_indent--;
2932 pandecode_log("},\n");
2933 }
2934
2935 static void
2936 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2937 {
2938 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2939 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2940
2941 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2942 pandecode_indent++;
2943
2944 if (h->zero) {
2945 pandecode_msg("XXX: tiler heap zero tripped\n");
2946 pandecode_prop("zero = 0x%x", h->zero);
2947 }
2948
2949 pandecode_prop("heap_size = 0x%x", h->heap_size);
2950 MEMORY_PROP(h, tiler_heap_start);
2951 MEMORY_PROP(h, tiler_heap_free);
2952
2953 /* this might point to the beginning of another buffer, when it's
2954 * really the end of the tiler heap buffer, so we have to be careful
2955 * here. but for zero length, we need the same pointer.
2956 */
2957
2958 if (h->tiler_heap_end == h->tiler_heap_start) {
2959 MEMORY_PROP(h, tiler_heap_start);
2960 } else {
2961 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2962 pandecode_prop("tiler_heap_end = %s + 1", a);
2963 free(a);
2964 }
2965
2966 for (int i = 0; i < 10; i++) {
2967 if (h->zeros[i] != 0) {
2968 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2969 i, h->zeros[i]);
2970 }
2971 }
2972
2973 if (h->unk1 != 0x1) {
2974 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2975 pandecode_prop("unk1 = 0x%x", h->unk1);
2976 }
2977
2978 if (h->unk7e007e != 0x7e007e) {
2979 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
2980 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
2981 }
2982
2983 pandecode_indent--;
2984 pandecode_log("};\n");
2985 }
2986
2987 static void
2988 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
2989 {
2990 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2991 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
2992
2993 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
2994
2995 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2996 pandecode_indent++;
2997
2998 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
2999 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
3000
3001 if (t->hierarchy_mask != 0xa &&
3002 t->hierarchy_mask != 0x14 &&
3003 t->hierarchy_mask != 0x28 &&
3004 t->hierarchy_mask != 0x50 &&
3005 t->hierarchy_mask != 0xa0)
3006 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
3007
3008 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
3009
3010 pandecode_prop("flags = 0x%" PRIx16, t->flags);
3011
3012 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
3013 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
3014
3015 if (t->zero0) {
3016 pandecode_msg("XXX: tiler meta zero tripped\n");
3017 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
3018 }
3019
3020 for (int i = 0; i < 12; i++) {
3021 if (t->zeros[i] != 0) {
3022 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
3023 i, t->zeros[i]);
3024 }
3025 }
3026
3027 pandecode_indent--;
3028 pandecode_log("};\n");
3029 }
3030
3031 static void
3032 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
3033 {
3034 if (u.pointer == 0x0)
3035 return;
3036
3037 pandecode_log(".primitive_size = {\n");
3038 pandecode_indent++;
3039
3040 if (constant) {
3041 pandecode_prop("constant = %f", u.constant);
3042 } else {
3043 MEMORY_PROP((&u), pointer);
3044 }
3045
3046 pandecode_indent--;
3047 pandecode_log("},\n");
3048 }
3049
3050 static void
3051 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
3052 {
3053 pandecode_log_cont("{\n");
3054 pandecode_indent++;
3055
3056 /* TODO: gl_PointSize on Bifrost */
3057 pandecode_primitive_size(t->primitive_size, true);
3058
3059 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
3060 || t->zero6) {
3061 pandecode_msg("XXX: tiler only zero tripped\n");
3062 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
3063 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
3064 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
3065 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
3066 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
3067 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
3068 }
3069
3070 pandecode_indent--;
3071 pandecode_log("},\n");
3072 }
3073
3074 static int
3075 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
3076 const struct pandecode_mapped_memory *mem,
3077 mali_ptr payload, int job_no, unsigned gpu_id)
3078 {
3079 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
3080
3081 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
3082
3083 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
3084 pandecode_indent++;
3085
3086 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
3087 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
3088
3089 pandecode_indent--;
3090 pandecode_log("};\n");
3091
3092 return sizeof(*v);
3093 }
3094
3095 static int
3096 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
3097 const struct pandecode_mapped_memory *mem,
3098 mali_ptr payload, int job_no, unsigned gpu_id)
3099 {
3100 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
3101
3102 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
3103 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
3104
3105 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
3106 pandecode_indent++;
3107
3108 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
3109
3110 pandecode_log(".tiler = ");
3111 pandecode_tiler_only_bfr(&t->tiler, job_no);
3112
3113 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3114
3115 pandecode_indent--;
3116 pandecode_log("};\n");
3117
3118 return sizeof(*t);
3119 }
3120
3121 static int
3122 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3123 const struct pandecode_mapped_memory *mem,
3124 mali_ptr payload, int job_no, unsigned gpu_id)
3125 {
3126 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3127 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
3128
3129 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3130
3131 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3132 pandecode_indent++;
3133
3134 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3135 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3136
3137 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3138 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3139
3140 pandecode_indent--;
3141 pandecode_log("};\n");
3142
3143 return sizeof(*v);
3144 }
3145
3146 static int
3147 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3148 mali_ptr payload, int job_no,
3149 bool is_bifrost, unsigned gpu_id)
3150 {
3151 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3152
3153 bool is_mfbd = s->framebuffer & MALI_MFBD;
3154
3155 if (!is_mfbd && is_bifrost)
3156 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3157
3158 struct pandecode_fbd info;
3159
3160 if (is_mfbd)
3161 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3162 else
3163 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3164
3165 /* Compute the tag for the tagged pointer. This contains the type of
3166 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3167 * additional structures follow the MFBD header (an extra payload or
3168 * not, as well as a count of render targets) */
3169
3170 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3171
3172 if (is_mfbd) {
3173 if (info.has_extra)
3174 expected_tag |= MALI_MFBD_TAG_EXTRA;
3175
3176 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3177 }
3178
3179 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3180 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3181 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3182 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3183 }
3184
3185 /* Extract tile coordinates */
3186
3187 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3188 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3189
3190 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3191 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3192
3193 /* For the max, we also want the floored (rather than ceiled) version for checking */
3194
3195 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3196 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3197
3198 /* Validate the coordinates are well-ordered */
3199
3200 if (min_x == max_x)
3201 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3202 else if (min_x > max_x)
3203 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3204
3205 if (min_y == max_y)
3206 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3207 else if (min_y > max_y)
3208 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3209
3210 /* Validate the coordinates fit inside the framebuffer. We use floor,
3211 * rather than ceil, for the max coordinates, since the tile
3212 * coordinates for something like an 800x600 framebuffer will actually
3213 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3214
3215 if ((min_x > info.width) || (max_x_f > info.width))
3216 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3217
3218 if ((min_y > info.height) || (max_y_f > info.height))
3219 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3220
3221 /* After validation, we print */
3222
3223 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3224
3225 /* The FBD is a tagged pointer */
3226
3227 unsigned tag = (s->framebuffer & ~FBD_MASK);
3228
3229 if (tag != expected_tag)
3230 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3231
3232 return sizeof(*s);
3233 }
3234
3235 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3236 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3237 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3238 * might wish to combine this with the bifrost parameter) because some details
3239 * are model-specific even within a particular architecture. Minimal traces
3240 * *only* examine the job descriptors, skipping printing entirely if there is
3241 * no faults, and only descends into the payload if there are faults. This is
3242 * useful for looking for faults without the overhead of invasive traces. */
3243
3244 void
3245 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3246 {
3247 pandecode_dump_file_open();
3248
3249 struct mali_job_descriptor_header *h;
3250 unsigned job_descriptor_number = 0;
3251
3252 do {
3253 struct pandecode_mapped_memory *mem =
3254 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3255
3256 void *payload;
3257
3258 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3259
3260 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3261 * high 32-bits of the 64-bit pointer are reused to store
3262 * something else.
3263 */
3264 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3265 h->job_type != MALI_JOB_TYPE_FRAGMENT ? 4 : 0;
3266 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3267
3268 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3269
3270 int job_no = job_descriptor_number++;
3271
3272 /* If the job is good to go, skip it in minimal mode */
3273 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3274 continue;
3275
3276 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3277 pandecode_indent++;
3278
3279 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
3280
3281 if (h->job_descriptor_size)
3282 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3283
3284 if (h->exception_status && h->exception_status != 0x1)
3285 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3286 h->exception_status,
3287 (h->exception_status >> 16) & 0xFFFF,
3288 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3289 h->exception_status & 0xFF);
3290
3291 if (h->first_incomplete_task)
3292 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3293
3294 if (h->fault_pointer)
3295 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3296
3297 if (h->job_barrier)
3298 pandecode_prop("job_barrier = %d", h->job_barrier);
3299
3300 pandecode_prop("job_index = %d", h->job_index);
3301
3302 if (h->unknown_flags)
3303 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3304
3305 if (h->job_dependency_index_1)
3306 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3307
3308 if (h->job_dependency_index_2)
3309 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3310
3311 pandecode_indent--;
3312 pandecode_log("};\n");
3313
3314 switch (h->job_type) {
3315 case MALI_JOB_TYPE_WRITE_VALUE: {
3316 struct mali_payload_write_value *s = payload;
3317 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3318 pandecode_indent++;
3319 MEMORY_PROP(s, address);
3320
3321 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3322 pandecode_msg("XXX: unknown value descriptor\n");
3323 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3324 }
3325
3326 if (s->reserved) {
3327 pandecode_msg("XXX: set value tripped\n");
3328 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3329 }
3330
3331 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3332 pandecode_indent--;
3333 pandecode_log("};\n");
3334
3335 break;
3336 }
3337
3338 case MALI_JOB_TYPE_TILER:
3339 case MALI_JOB_TYPE_VERTEX:
3340 case MALI_JOB_TYPE_COMPUTE:
3341 if (bifrost) {
3342 if (h->job_type == MALI_JOB_TYPE_TILER)
3343 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3344 else
3345 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3346 } else
3347 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3348
3349 break;
3350
3351 case MALI_JOB_TYPE_FRAGMENT:
3352 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3353 break;
3354
3355 default:
3356 break;
3357 }
3358 } while ((jc_gpu_va = h->next_job));
3359
3360 pandecode_map_read_write();
3361 }