projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
panfrost: Redirect cmdstream includes through GenXML
[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 FILE *pandecode_dump_stream;
61
62 /* Semantic logging type.
63 *
64 * Raw: for raw messages to be printed as is.
65 * Message: for helpful information to be commented out in replays.
66 * Property: for properties of a struct
67 *
68 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
69 */
70
71 enum pandecode_log_type {
72 PANDECODE_RAW,
73 PANDECODE_MESSAGE,
74 PANDECODE_PROPERTY
75 };
76
77 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
78 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
79 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
80
81 unsigned pandecode_indent = 0;
82
83 static void
84 pandecode_make_indent(void)
85 {
86 for (unsigned i = 0; i < pandecode_indent; ++i)
87 fprintf(pandecode_dump_stream, " ");
88 }
89
90 static void
91 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
92 {
93 va_list ap;
94
95 pandecode_make_indent();
96
97 if (type == PANDECODE_MESSAGE)
98 fprintf(pandecode_dump_stream, "// ");
99 else if (type == PANDECODE_PROPERTY)
100 fprintf(pandecode_dump_stream, ".");
101
102 va_start(ap, format);
103 vfprintf(pandecode_dump_stream, format, ap);
104 va_end(ap);
105
106 if (type == PANDECODE_PROPERTY)
107 fprintf(pandecode_dump_stream, ",\n");
108 }
109
110 static void
111 pandecode_log_cont(const char *format, ...)
112 {
113 va_list ap;
114
115 va_start(ap, format);
116 vfprintf(pandecode_dump_stream, format, ap);
117 va_end(ap);
118 }
119
120 /* To check for memory safety issues, validates that the given pointer in GPU
121 * memory is valid, containing at least sz bytes. The goal is to eliminate
122 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
123 * overruns) by statically validating pointers.
124 */
125
126 static void
127 pandecode_validate_buffer(mali_ptr addr, size_t sz)
128 {
129 if (!addr) {
130 pandecode_msg("XXX: null pointer deref");
131 return;
132 }
133
134 /* Find a BO */
135
136 struct pandecode_mapped_memory *bo =
137 pandecode_find_mapped_gpu_mem_containing(addr);
138
139 if (!bo) {
140 pandecode_msg("XXX: invalid memory dereference\n");
141 return;
142 }
143
144 /* Bounds check */
145
146 unsigned offset = addr - bo->gpu_va;
147 unsigned total = offset + sz;
148
149 if (total > bo->length) {
150 pandecode_msg("XXX: buffer overrun. "
151 "Chunk of size %zu at offset %d in buffer of size %zu. "
152 "Overrun by %zu bytes. \n",
153 sz, offset, bo->length, total - bo->length);
154 return;
155 }
156 }
157
158 struct pandecode_flag_info {
159 u64 flag;
160 const char *name;
161 };
162
163 static void
164 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
165 u64 flags)
166 {
167 bool decodable_flags_found = false;
168
169 for (int i = 0; flag_info[i].name; i++) {
170 if ((flags & flag_info[i].flag) != flag_info[i].flag)
171 continue;
172
173 if (!decodable_flags_found) {
174 decodable_flags_found = true;
175 } else {
176 pandecode_log_cont(" | ");
177 }
178
179 pandecode_log_cont("%s", flag_info[i].name);
180
181 flags &= ~flag_info[i].flag;
182 }
183
184 if (decodable_flags_found) {
185 if (flags)
186 pandecode_log_cont(" | 0x%" PRIx64, flags);
187 } else {
188 pandecode_log_cont("0x%" PRIx64, flags);
189 }
190 }
191
192 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
193 static const struct pandecode_flag_info gl_enable_flag_info[] = {
194 FLAG_INFO(OCCLUSION_QUERY),
195 FLAG_INFO(OCCLUSION_PRECISE),
196 FLAG_INFO(FRONT_CCW_TOP),
197 FLAG_INFO(CULL_FACE_FRONT),
198 FLAG_INFO(CULL_FACE_BACK),
199 {}
200 };
201 #undef FLAG_INFO
202
203 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
204 static const struct pandecode_flag_info clear_flag_info[] = {
205 FLAG_INFO(FAST),
206 FLAG_INFO(SLOW),
207 FLAG_INFO(SLOW_STENCIL),
208 {}
209 };
210 #undef FLAG_INFO
211
212 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
213 static const struct pandecode_flag_info mask_flag_info[] = {
214 FLAG_INFO(R),
215 FLAG_INFO(G),
216 FLAG_INFO(B),
217 FLAG_INFO(A),
218 {}
219 };
220 #undef FLAG_INFO
221
222 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
223 static const struct pandecode_flag_info u3_flag_info[] = {
224 FLAG_INFO(HAS_MSAA),
225 FLAG_INFO(PER_SAMPLE),
226 FLAG_INFO(CAN_DISCARD),
227 FLAG_INFO(HAS_BLEND_SHADER),
228 FLAG_INFO(DEPTH_WRITEMASK),
229 FLAG_INFO(DEPTH_CLIP_NEAR),
230 FLAG_INFO(DEPTH_CLIP_FAR),
231 {}
232 };
233
234 static const struct pandecode_flag_info u4_flag_info[] = {
235 FLAG_INFO(NO_MSAA),
236 FLAG_INFO(NO_DITHER),
237 FLAG_INFO(DEPTH_RANGE_A),
238 FLAG_INFO(DEPTH_RANGE_B),
239 FLAG_INFO(STENCIL_TEST),
240 FLAG_INFO(ALPHA_TO_COVERAGE),
241 {}
242 };
243 #undef FLAG_INFO
244
245 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
246 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
247 FLAG_INFO(SRGB),
248 {}
249 };
250 #undef FLAG_INFO
251
252 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
253 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
254 FLAG_INFO(YTR),
255 {}
256 };
257 #undef FLAG_INFO
258
259 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
260 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
261 FLAG_INFO(PRESENT),
262 {}
263 };
264 #undef FLAG_INFO
265
266 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
267 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
268 FLAG_INFO(ZS),
269 {}
270 };
271 #undef FLAG_INFO
272
273 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
274 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
275 FLAG_INFO(WRITES_Z),
276 FLAG_INFO(EARLY_Z),
277 FLAG_INFO(READS_TILEBUFFER),
278 FLAG_INFO(WRITES_GLOBAL),
279 FLAG_INFO(READS_ZS),
280 {}
281 };
282
283 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
284 FLAG_INFO(WRITES_S),
285 FLAG_INFO(SUPPRESS_INF_NAN),
286 {}
287 };
288 #undef FLAG_INFO
289
290 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
291 static const struct pandecode_flag_info shader_bifrost_info [] = {
292 FLAG_INFO(FULL_THREAD),
293 FLAG_INFO(EARLY_Z),
294 FLAG_INFO(FIRST_ATEST),
295 {}
296 };
297
298 #undef FLAG_INFO
299
300 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
301 static const struct pandecode_flag_info mfbd_flag_info [] = {
302 FLAG_INFO(DEPTH_WRITE),
303 FLAG_INFO(EXTRA),
304 {}
305 };
306 #undef FLAG_INFO
307
308 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
309 static const struct pandecode_flag_info sampler_flag_info [] = {
310 FLAG_INFO(MAG_NEAREST),
311 FLAG_INFO(MIN_NEAREST),
312 FLAG_INFO(MIP_LINEAR_1),
313 FLAG_INFO(MIP_LINEAR_2),
314 FLAG_INFO(NORM_COORDS),
315 {}
316 };
317 #undef FLAG_INFO
318
319 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
320 static const struct pandecode_flag_info sfbd_unk1_info [] = {
321 FLAG_INFO(MSAA_8),
322 FLAG_INFO(MSAA_A),
323 {}
324 };
325 #undef FLAG_INFO
326
327 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
328 static const struct pandecode_flag_info sfbd_unk2_info [] = {
329 FLAG_INFO(MSAA_B),
330 FLAG_INFO(SRGB),
331 {}
332 };
333 #undef FLAG_INFO
334
335 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
336 static char *pandecode_format(enum mali_format format)
337 {
338 static char unk_format_str[10];
339
340 switch (format) {
341 DEFINE_CASE(ETC2_RGB8);
342 DEFINE_CASE(ETC2_R11_UNORM);
343 DEFINE_CASE(ETC2_RGBA8);
344 DEFINE_CASE(ETC2_RG11_UNORM);
345 DEFINE_CASE(ETC2_R11_SNORM);
346 DEFINE_CASE(ETC2_RG11_SNORM);
347 DEFINE_CASE(ETC2_RGB8A1);
348 DEFINE_CASE(NXR);
349 DEFINE_CASE(BC1_UNORM);
350 DEFINE_CASE(BC2_UNORM);
351 DEFINE_CASE(BC3_UNORM);
352 DEFINE_CASE(BC4_UNORM);
353 DEFINE_CASE(BC4_SNORM);
354 DEFINE_CASE(BC5_UNORM);
355 DEFINE_CASE(BC5_SNORM);
356 DEFINE_CASE(BC6H_UF16);
357 DEFINE_CASE(BC6H_SF16);
358 DEFINE_CASE(BC7_UNORM);
359 DEFINE_CASE(ASTC_3D_LDR);
360 DEFINE_CASE(ASTC_3D_HDR);
361 DEFINE_CASE(ASTC_2D_LDR);
362 DEFINE_CASE(ASTC_2D_HDR);
363 DEFINE_CASE(RGB565);
364 DEFINE_CASE(RGB5_X1_UNORM);
365 DEFINE_CASE(RGB5_A1_UNORM);
366 DEFINE_CASE(RGB10_A2_UNORM);
367 DEFINE_CASE(RGB10_A2_SNORM);
368 DEFINE_CASE(RGB10_A2UI);
369 DEFINE_CASE(RGB10_A2I);
370 DEFINE_CASE(RGB332_UNORM);
371 DEFINE_CASE(RGB233_UNORM);
372 DEFINE_CASE(Z24X8_UNORM);
373 DEFINE_CASE(R32_FIXED);
374 DEFINE_CASE(RG32_FIXED);
375 DEFINE_CASE(RGB32_FIXED);
376 DEFINE_CASE(RGBA32_FIXED);
377 DEFINE_CASE(R11F_G11F_B10F);
378 DEFINE_CASE(R9F_G9F_B9F_E5F);
379 DEFINE_CASE(VARYING_POS);
380 DEFINE_CASE(VARYING_DISCARD);
381
382 DEFINE_CASE(R8_SNORM);
383 DEFINE_CASE(R16_SNORM);
384 DEFINE_CASE(R32_SNORM);
385 DEFINE_CASE(RG8_SNORM);
386 DEFINE_CASE(RG16_SNORM);
387 DEFINE_CASE(RG32_SNORM);
388 DEFINE_CASE(RGB8_SNORM);
389 DEFINE_CASE(RGB16_SNORM);
390 DEFINE_CASE(RGB32_SNORM);
391 DEFINE_CASE(RGBA8_SNORM);
392 DEFINE_CASE(RGBA16_SNORM);
393 DEFINE_CASE(RGBA32_SNORM);
394
395 DEFINE_CASE(R8UI);
396 DEFINE_CASE(R16UI);
397 DEFINE_CASE(R32UI);
398 DEFINE_CASE(RG8UI);
399 DEFINE_CASE(RG16UI);
400 DEFINE_CASE(RG32UI);
401 DEFINE_CASE(RGB8UI);
402 DEFINE_CASE(RGB16UI);
403 DEFINE_CASE(RGB32UI);
404 DEFINE_CASE(RGBA8UI);
405 DEFINE_CASE(RGBA16UI);
406 DEFINE_CASE(RGBA32UI);
407
408 DEFINE_CASE(R8_UNORM);
409 DEFINE_CASE(R16_UNORM);
410 DEFINE_CASE(R32_UNORM);
411 DEFINE_CASE(R32F);
412 DEFINE_CASE(RG8_UNORM);
413 DEFINE_CASE(RG16_UNORM);
414 DEFINE_CASE(RG32_UNORM);
415 DEFINE_CASE(RG32F);
416 DEFINE_CASE(RGB8_UNORM);
417 DEFINE_CASE(RGB16_UNORM);
418 DEFINE_CASE(RGB32_UNORM);
419 DEFINE_CASE(RGB32F);
420 DEFINE_CASE(RGBA4_UNORM);
421 DEFINE_CASE(RGBA8_UNORM);
422 DEFINE_CASE(RGBA16_UNORM);
423 DEFINE_CASE(RGBA32_UNORM);
424 DEFINE_CASE(RGBA32F);
425
426 DEFINE_CASE(R8I);
427 DEFINE_CASE(R16I);
428 DEFINE_CASE(R32I);
429 DEFINE_CASE(RG8I);
430 DEFINE_CASE(R16F);
431 DEFINE_CASE(RG16I);
432 DEFINE_CASE(RG32I);
433 DEFINE_CASE(RG16F);
434 DEFINE_CASE(RGB8I);
435 DEFINE_CASE(RGB16I);
436 DEFINE_CASE(RGB32I);
437 DEFINE_CASE(RGB16F);
438 DEFINE_CASE(RGBA8I);
439 DEFINE_CASE(RGBA16I);
440 DEFINE_CASE(RGBA32I);
441 DEFINE_CASE(RGBA16F);
442
443 DEFINE_CASE(RGBA4);
444 DEFINE_CASE(RGBA8_2);
445 DEFINE_CASE(RGB10_A2_2);
446 default:
447 snprintf(unk_format_str, sizeof(unk_format_str), "MALI_0x%02x", format);
448 return unk_format_str;
449 }
450 }
451
452 #undef DEFINE_CASE
453
454 static char *
455 pandecode_job_type(enum mali_job_type type)
456 {
457 #define DEFINE_CASE(name) case JOB_TYPE_ ## name: return "JOB_TYPE_" #name
458
459 switch (type) {
460 DEFINE_CASE(NULL);
461 DEFINE_CASE(WRITE_VALUE);
462 DEFINE_CASE(CACHE_FLUSH);
463 DEFINE_CASE(COMPUTE);
464 DEFINE_CASE(VERTEX);
465 DEFINE_CASE(TILER);
466 DEFINE_CASE(FUSED);
467 DEFINE_CASE(FRAGMENT);
468
469 case JOB_NOT_STARTED:
470 return "NOT_STARTED";
471
472 default:
473 pandecode_log("Warning! Unknown job type %x\n", type);
474 return "!?!?!?";
475 }
476
477 #undef DEFINE_CASE
478 }
479
480 static char *
481 pandecode_draw_mode(enum mali_draw_mode mode)
482 {
483 #define DEFINE_CASE(name) case MALI_ ## name: return "MALI_" #name
484
485 switch (mode) {
486 DEFINE_CASE(DRAW_NONE);
487 DEFINE_CASE(POINTS);
488 DEFINE_CASE(LINES);
489 DEFINE_CASE(TRIANGLES);
490 DEFINE_CASE(TRIANGLE_STRIP);
491 DEFINE_CASE(TRIANGLE_FAN);
492 DEFINE_CASE(LINE_STRIP);
493 DEFINE_CASE(LINE_LOOP);
494 DEFINE_CASE(POLYGON);
495 DEFINE_CASE(QUADS);
496 DEFINE_CASE(QUAD_STRIP);
497
498 default:
499 pandecode_msg("XXX: invalid draw mode %X\n", mode);
500 return "";
501 }
502
503 #undef DEFINE_CASE
504 }
505
506 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
507 static char *
508 pandecode_func(enum mali_func mode)
509 {
510 switch (mode) {
511 DEFINE_CASE(NEVER);
512 DEFINE_CASE(LESS);
513 DEFINE_CASE(EQUAL);
514 DEFINE_CASE(LEQUAL);
515 DEFINE_CASE(GREATER);
516 DEFINE_CASE(NOTEQUAL);
517 DEFINE_CASE(GEQUAL);
518 DEFINE_CASE(ALWAYS);
519
520 default:
521 pandecode_msg("XXX: invalid func %X\n", mode);
522 return "";
523 }
524 }
525 #undef DEFINE_CASE
526
527 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
528 static char *
529 pandecode_msaa_mode(enum mali_msaa_mode mode)
530 {
531 switch (mode) {
532 DEFINE_CASE(SINGLE);
533 DEFINE_CASE(AVERAGE);
534 DEFINE_CASE(MULTIPLE);
535 DEFINE_CASE(LAYERED);
536 default:
537 unreachable("Impossible");
538 return "";
539 }
540 }
541 #undef DEFINE_CASE
542
543 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
544 static char *
545 pandecode_stencil_op(enum mali_stencil_op op)
546 {
547 switch (op) {
548 DEFINE_CASE(KEEP);
549 DEFINE_CASE(REPLACE);
550 DEFINE_CASE(ZERO);
551 DEFINE_CASE(INVERT);
552 DEFINE_CASE(INCR_WRAP);
553 DEFINE_CASE(DECR_WRAP);
554 DEFINE_CASE(INCR);
555 DEFINE_CASE(DECR);
556
557 default:
558 pandecode_msg("XXX: invalid stencil op %X\n", op);
559 return "";
560 }
561 }
562
563 #undef DEFINE_CASE
564
565 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
566 {
567 switch(mode) {
568 /* TODO: Combine to just "instanced" once this can be done
569 * unambiguously in all known cases */
570 case MALI_ATTR_POT_DIVIDE:
571 return "instanced_pot";
572 case MALI_ATTR_MODULO:
573 return "instanced_mod";
574 case MALI_ATTR_NPOT_DIVIDE:
575 return "instanced_npot";
576 case MALI_ATTR_IMAGE:
577 return "image";
578 default:
579 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
580 return "";
581 }
582 }
583
584 static const char *
585 pandecode_special_record(uint64_t v, bool* attribute)
586 {
587 switch(v) {
588 case MALI_ATTR_VERTEXID:
589 *attribute = true;
590 return "gl_VertexID";
591 case MALI_ATTR_INSTANCEID:
592 *attribute = true;
593 return "gl_InstanceID";
594 case MALI_VARYING_FRAG_COORD:
595 return "gl_FragCoord";
596 case MALI_VARYING_FRONT_FACING:
597 return "gl_FrontFacing";
598 case MALI_VARYING_POINT_COORD:
599 return "gl_PointCoord";
600 default:
601 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
602 return "";
603 }
604 }
605
606 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
607 static char *
608 pandecode_wrap_mode(enum mali_wrap_mode op)
609 {
610 switch (op) {
611 DEFINE_CASE(REPEAT);
612 DEFINE_CASE(CLAMP_TO_EDGE);
613 DEFINE_CASE(CLAMP);
614 DEFINE_CASE(CLAMP_TO_BORDER);
615 DEFINE_CASE(MIRRORED_REPEAT);
616 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE);
617 DEFINE_CASE(MIRRORED_CLAMP);
618 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER);
619
620 default:
621 pandecode_msg("XXX: invalid wrap mode %X\n", op);
622 return "";
623 }
624 }
625 #undef DEFINE_CASE
626
627 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
628 static char *
629 pandecode_block_format(enum mali_block_format fmt)
630 {
631 switch (fmt) {
632 DEFINE_CASE(TILED);
633 DEFINE_CASE(UNKNOWN);
634 DEFINE_CASE(LINEAR);
635 DEFINE_CASE(AFBC);
636
637 default:
638 unreachable("Invalid case");
639 }
640 }
641 #undef DEFINE_CASE
642
643 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
644 static char *
645 pandecode_exception_access(unsigned access)
646 {
647 switch (access) {
648 DEFINE_CASE(NONE);
649 DEFINE_CASE(EXECUTE);
650 DEFINE_CASE(READ);
651 DEFINE_CASE(WRITE);
652
653 default:
654 unreachable("Invalid case");
655 }
656 }
657 #undef DEFINE_CASE
658
659 /* Midgard's tiler descriptor is embedded within the
660 * larger FBD */
661
662 static void
663 pandecode_midgard_tiler_descriptor(
664 const struct midgard_tiler_descriptor *t,
665 unsigned width,
666 unsigned height,
667 bool is_fragment,
668 bool has_hierarchy)
669 {
670 pandecode_log(".tiler = {\n");
671 pandecode_indent++;
672
673 if (t->hierarchy_mask == MALI_TILER_DISABLED)
674 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
675 else
676 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
677
678 /* We know this name from the kernel, but we never see it nonzero */
679
680 if (t->flags)
681 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
682
683 MEMORY_PROP(t, polygon_list);
684
685 /* The body is offset from the base of the polygon list */
686 //assert(t->polygon_list_body > t->polygon_list);
687 unsigned body_offset = t->polygon_list_body - t->polygon_list;
688
689 /* It needs to fit inside the reported size */
690 //assert(t->polygon_list_size >= body_offset);
691
692 /* Now that we've sanity checked, we'll try to calculate the sizes
693 * ourselves for comparison */
694
695 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
696 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
697
698 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
699 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
700 ref_header, ref_size);
701 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
702 pandecode_msg("body offset %d\n", body_offset);
703 }
704
705 /* The tiler heap has a start and end specified -- it should be
706 * identical to what we have in the BO. The exception is if tiling is
707 * disabled. */
708
709 MEMORY_PROP(t, heap_start);
710 assert(t->heap_end >= t->heap_start);
711
712 struct pandecode_mapped_memory *heap =
713 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
714
715 unsigned heap_size = t->heap_end - t->heap_start;
716
717 /* Tiling is enabled with a special flag */
718 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
719 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
720
721 bool tiling_enabled = hierarchy_mask;
722
723 if (tiling_enabled) {
724 /* When tiling is enabled, the heap should be a tight fit */
725 unsigned heap_offset = t->heap_start - heap->gpu_va;
726 if ((heap_offset + heap_size) != heap->length) {
727 pandecode_msg("XXX: heap size %u (expected %zu)\n",
728 heap_size, heap->length - heap_offset);
729 }
730
731 /* We should also have no other flags */
732 if (tiler_flags)
733 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
734 } else {
735 /* When tiling is disabled, we should have that flag and no others */
736
737 if (tiler_flags != MALI_TILER_DISABLED) {
738 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
739 tiler_flags);
740 }
741
742 /* We should also have an empty heap */
743 if (heap_size) {
744 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
745 heap_size);
746 }
747
748 /* Disabled tiling is used only for clear-only jobs, which are
749 * purely FRAGMENT, so we should never see this for
750 * non-FRAGMENT descriptors. */
751
752 if (!is_fragment)
753 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
754 }
755
756 /* We've never seen weights used in practice, but we know from the
757 * kernel these fields is there */
758
759 bool nonzero_weights = false;
760
761 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
762 nonzero_weights |= t->weights[w] != 0x0;
763 }
764
765 if (nonzero_weights) {
766 pandecode_log(".weights = { ");
767
768 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
769 pandecode_log_cont("%d, ", t->weights[w]);
770 }
771
772 pandecode_log("},");
773 }
774
775 pandecode_indent--;
776 pandecode_log("}\n");
777 }
778
779 /* TODO: The Bifrost tiler is not understood at all yet */
780
781 static void
782 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
783 {
784 pandecode_log(".tiler = {\n");
785 pandecode_indent++;
786
787 MEMORY_PROP(fb, tiler_meta);
788
789 for (int i = 0; i < 16; i++) {
790 if (fb->zeros[i] != 0) {
791 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
792 i, fb->zeros[i]);
793 }
794 }
795
796 pandecode_log("},\n");
797
798 pandecode_indent--;
799 pandecode_log("}\n");
800
801 }
802
803 /* Information about the framebuffer passed back for
804 * additional analysis */
805
806 struct pandecode_fbd {
807 unsigned width;
808 unsigned height;
809 unsigned rt_count;
810 bool has_extra;
811 };
812
813 static void
814 pandecode_sfbd_format(struct mali_sfbd_format format)
815 {
816 pandecode_log(".format = {\n");
817 pandecode_indent++;
818
819 pandecode_log(".unk1 = ");
820 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
821 pandecode_log_cont(",\n");
822
823 /* TODO: Map formats so we can check swizzles and print nicely */
824 pandecode_log("swizzle");
825 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
826 pandecode_log_cont(",\n");
827
828 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
829 (format.nr_channels + 1));
830
831 pandecode_log(".unk2 = ");
832 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
833 pandecode_log_cont(",\n");
834
835 pandecode_prop("block = %s", pandecode_block_format(format.block));
836
837 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
838
839 pandecode_indent--;
840 pandecode_log("},\n");
841 }
842
843 static void
844 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
845 {
846 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
847
848 if (desc->unk0)
849 pandecode_prop("unk0 = 0x%x", desc->unk0);
850
851 if (desc->shared_workgroup_count != 0x1F) {
852 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
853 if (!is_compute)
854 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
855 }
856
857 if (desc->shared_unk1 || desc->shared_shift) {
858 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
859 pandecode_prop("shared_shift = %X", desc->shared_shift);
860
861 if (!is_compute)
862 pandecode_msg("XXX: shared memory configured in noncompute shader");
863 }
864
865 if (desc->shared_zero) {
866 pandecode_msg("XXX: shared memory zero tripped\n");
867 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
868 }
869
870 if (desc->shared_memory && !is_compute)
871 pandecode_msg("XXX: shared memory used in noncompute shader\n");
872
873 MEMORY_PROP(desc, scratchpad);
874 MEMORY_PROP(desc, shared_memory);
875 MEMORY_PROP(desc, unknown1);
876
877 if (desc->scratchpad) {
878 struct pandecode_mapped_memory *smem =
879 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
880
881 pandecode_msg("scratchpad size %u\n", smem->length);
882 }
883
884 }
885
886 static struct pandecode_fbd
887 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
888 {
889 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
890 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
891
892 struct pandecode_fbd info = {
893 .has_extra = false,
894 .rt_count = 1
895 };
896
897 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
898 pandecode_indent++;
899
900 pandecode_log(".shared_memory = {\n");
901 pandecode_indent++;
902 pandecode_shared_memory(&s->shared_memory, false);
903 pandecode_indent--;
904 pandecode_log("},\n");
905
906 pandecode_sfbd_format(s->format);
907
908 info.width = s->width + 1;
909 info.height = s->height + 1;
910
911 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
912 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
913
914 MEMORY_PROP(s, checksum);
915
916 if (s->checksum_stride)
917 pandecode_prop("checksum_stride = %d", s->checksum_stride);
918
919 MEMORY_PROP(s, framebuffer);
920 pandecode_prop("stride = %d", s->stride);
921
922 /* Earlier in the actual commandstream -- right before width -- but we
923 * delay to flow nicer */
924
925 pandecode_log(".clear_flags = ");
926 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
927 pandecode_log_cont(",\n");
928
929 if (s->depth_buffer) {
930 MEMORY_PROP(s, depth_buffer);
931 pandecode_prop("depth_stride = %d", s->depth_stride);
932 }
933
934 if (s->stencil_buffer) {
935 MEMORY_PROP(s, stencil_buffer);
936 pandecode_prop("stencil_stride = %d", s->stencil_stride);
937 }
938
939 if (s->depth_stride_zero ||
940 s->stencil_stride_zero ||
941 s->zero7 || s->zero8) {
942 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
943 pandecode_prop("depth_stride_zero = 0x%x",
944 s->depth_stride_zero);
945 pandecode_prop("stencil_stride_zero = 0x%x",
946 s->stencil_stride_zero);
947 pandecode_prop("zero7 = 0x%" PRIx32,
948 s->zero7);
949 pandecode_prop("zero8 = 0x%" PRIx32,
950 s->zero8);
951 }
952
953 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
954 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
955 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
956 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
957 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
958 }
959
960 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
961 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
962 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
963 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
964 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
965 }
966
967 if (s->clear_stencil) {
968 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
969 }
970
971 const struct midgard_tiler_descriptor t = s->tiler;
972
973 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
974 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
975
976 pandecode_indent--;
977 pandecode_log("};\n");
978
979 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
980 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
981 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
982
983 pandecode_log_cont(".zero3 = {");
984
985 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
986 pandecode_log_cont("%X, ", s->zero3[i]);
987
988 pandecode_log_cont("},\n");
989
990 pandecode_log_cont(".zero6 = {");
991
992 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
993 pandecode_log_cont("%X, ", s->zero6[i]);
994
995 pandecode_log_cont("},\n");
996
997 return info;
998 }
999
1000 static void
1001 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
1002 {
1003 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1004 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
1005
1006 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
1007 pandecode_indent++;
1008 pandecode_shared_memory(s, true);
1009 pandecode_indent--;
1010 pandecode_log("},\n");
1011 }
1012
1013 /* Extracts the number of components associated with a Mali format */
1014
1015 static unsigned
1016 pandecode_format_component_count(enum mali_format fmt)
1017 {
1018 /* Mask out the format class */
1019 unsigned top = fmt & 0b11100000;
1020
1021 switch (top) {
1022 case MALI_FORMAT_SNORM:
1023 case MALI_FORMAT_UINT:
1024 case MALI_FORMAT_UNORM:
1025 case MALI_FORMAT_SINT:
1026 return ((fmt >> 3) & 3) + 1;
1027 default:
1028 /* TODO: Validate */
1029 return 4;
1030 }
1031 }
1032
1033 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
1034
1035 static unsigned
1036 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
1037 {
1038 unsigned mask = 0;
1039 assert(MALI_CHANNEL_RED == 0);
1040
1041 for (unsigned c = 0; c < 4; ++c) {
1042 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
1043
1044 if (chan <= MALI_CHANNEL_ALPHA)
1045 mask |= (1 << chan);
1046 }
1047
1048 return mask;
1049 }
1050
1051 /* Validates that a (format, swizzle) pair is valid, in the sense that the
1052 * swizzle doesn't access any components that are undefined in the format.
1053 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
1054 * omitted */
1055
1056 static bool
1057 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
1058 {
1059 unsigned nr_comp = pandecode_format_component_count(fmt);
1060 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
1061 unsigned valid_mask = (1 << nr_comp) - 1;
1062 unsigned invalid_mask = ~valid_mask;
1063
1064 if (access_mask & invalid_mask) {
1065 pandecode_msg("XXX: invalid components accessed\n");
1066 return false;
1067 }
1068
1069 /* Check for the default non-swizzling swizzle so we can suppress
1070 * useless printing for the defaults */
1071
1072 unsigned default_swizzles[4] = {
1073 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1074 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1075 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
1076 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
1077 };
1078
1079 return (swizzle == default_swizzles[nr_comp - 1]);
1080 }
1081
1082 /* Maps MALI_RGBA32F to rgba32f, etc */
1083
1084 static void
1085 pandecode_format_short(enum mali_format fmt, bool srgb)
1086 {
1087 /* We want a type-like format, so cut off the initial MALI_ */
1088 char *format = pandecode_format(fmt);
1089 format += strlen("MALI_");
1090
1091 unsigned len = strlen(format);
1092 char *lower_format = calloc(1, len + 1);
1093
1094 for (unsigned i = 0; i < len; ++i)
1095 lower_format[i] = tolower(format[i]);
1096
1097 /* Sanity check sRGB flag is applied to RGB, per the name */
1098 if (srgb && lower_format[0] != 'r')
1099 pandecode_msg("XXX: sRGB applied to non-colour format\n");
1100
1101 /* Just prefix with an s, so you get formats like srgba8_unorm */
1102 if (srgb)
1103 pandecode_log_cont("s");
1104
1105 pandecode_log_cont("%s", lower_format);
1106 free(lower_format);
1107 }
1108
1109 static void
1110 pandecode_swizzle(unsigned swizzle, enum mali_format format)
1111 {
1112 /* First, do some validation */
1113 bool trivial_swizzle = pandecode_validate_format_swizzle(
1114 format, swizzle);
1115
1116 if (trivial_swizzle)
1117 return;
1118
1119 /* Next, print the swizzle */
1120 pandecode_log_cont(".");
1121
1122 static const char components[] = "rgba01";
1123
1124 for (unsigned c = 0; c < 4; ++c) {
1125 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
1126
1127 if (chan >= MALI_CHANNEL_RESERVED_0) {
1128 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
1129 continue;
1130 }
1131 pandecode_log_cont("%c", components[chan]);
1132 }
1133 }
1134
1135 static void
1136 pandecode_rt_format(struct mali_rt_format format)
1137 {
1138 pandecode_log(".format = {\n");
1139 pandecode_indent++;
1140
1141 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1142 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1143 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1144 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1145
1146 pandecode_prop("block = %s", pandecode_block_format(format.block));
1147
1148 /* TODO: Map formats so we can check swizzles and print nicely */
1149 pandecode_log("swizzle");
1150 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1151 pandecode_log_cont(",\n");
1152
1153 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1154 (format.nr_channels + 1));
1155
1156 pandecode_log(".flags = ");
1157 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1158 pandecode_log_cont(",\n");
1159
1160 pandecode_prop("msaa = %s", pandecode_msaa_mode(format.msaa));
1161
1162 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1163 * which is a faster hardware-based alternative to the wallpaper method
1164 * to preserve framebuffer contents across frames. In practice, MFBD
1165 * preload is buggy on Midgard, and so this is a chicken bit. If this
1166 * bit isn't set, most likely something broke unrelated to preload */
1167
1168 if (!format.no_preload) {
1169 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1170 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1171 }
1172
1173 if (format.zero)
1174 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1175
1176 pandecode_indent--;
1177 pandecode_log("},\n");
1178 }
1179
1180 static void
1181 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1182 {
1183 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1184 pandecode_indent++;
1185
1186 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1187 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1188 struct pandecode_mapped_memory *mem =
1189 pandecode_find_mapped_gpu_mem_containing(rt_va);
1190 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1191
1192 pandecode_log("{\n");
1193 pandecode_indent++;
1194
1195 pandecode_rt_format(rt->format);
1196
1197 if (rt->format.block == MALI_BLOCK_AFBC) {
1198 pandecode_log(".afbc = {\n");
1199 pandecode_indent++;
1200
1201 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1202 pandecode_prop("metadata = %s", a);
1203 free(a);
1204
1205 pandecode_prop("stride = %d", rt->afbc.stride);
1206
1207 pandecode_log(".flags = ");
1208 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1209 pandecode_log_cont(",\n");
1210
1211 pandecode_indent--;
1212 pandecode_log("},\n");
1213 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1214 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1215 rt->afbc.metadata,
1216 rt->afbc.stride,
1217 rt->afbc.flags);
1218 }
1219
1220 MEMORY_PROP(rt, framebuffer);
1221 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1222
1223 if (rt->layer_stride)
1224 pandecode_prop("layer_stride = %d", rt->layer_stride);
1225
1226 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1227 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1228 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1229 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1230 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1231 }
1232
1233 if (rt->zero1 || rt->zero2) {
1234 pandecode_msg("XXX: render target zeros tripped\n");
1235 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1236 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1237 }
1238
1239 pandecode_indent--;
1240 pandecode_log("},\n");
1241 }
1242
1243 pandecode_indent--;
1244 pandecode_log("};\n");
1245 }
1246
1247 static struct pandecode_fbd
1248 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1249 {
1250 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1251 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1252
1253 struct pandecode_fbd info;
1254
1255 if (is_bifrost && fb->msaa.sample_locations) {
1256 /* The blob stores all possible sample locations in a single buffer
1257 * allocated on startup, and just switches the pointer when switching
1258 * MSAA state. For now, we just put the data into the cmdstream, but we
1259 * should do something like what the blob does with a real driver.
1260 *
1261 * There seem to be 32 slots for sample locations, followed by another
1262 * 16. The second 16 is just the center location followed by 15 zeros
1263 * in all the cases I've identified (maybe shader vs. depth/color
1264 * samples?).
1265 */
1266
1267 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1268
1269 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1270
1271 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1272 pandecode_indent++;
1273
1274 for (int i = 0; i < 32 + 16; i++) {
1275 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1276 }
1277
1278 pandecode_indent--;
1279 pandecode_log("};\n");
1280 }
1281
1282 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1283 pandecode_indent++;
1284
1285 if (is_bifrost) {
1286 pandecode_log(".msaa = {\n");
1287 pandecode_indent++;
1288
1289 if (fb->msaa.sample_locations)
1290 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1291 else
1292 pandecode_msg("XXX: sample_locations missing\n");
1293
1294 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1295 pandecode_msg("XXX: multisampling zero tripped\n");
1296 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1297 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1298 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1299 }
1300
1301 pandecode_indent--;
1302 pandecode_log("},\n");
1303 } else {
1304 pandecode_log(".shared_memory = {\n");
1305 pandecode_indent++;
1306 pandecode_shared_memory(&fb->shared_memory, is_compute);
1307 pandecode_indent--;
1308 pandecode_log("},\n");
1309 }
1310
1311 info.width = fb->width1 + 1;
1312 info.height = fb->height1 + 1;
1313 info.rt_count = fb->rt_count_1 + 1;
1314
1315 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1316 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1317 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1318 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1319
1320 pandecode_prop("unk1 = 0x%x", fb->unk1);
1321 pandecode_prop("unk2 = 0x%x", fb->unk2);
1322 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1323 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1324
1325 pandecode_log(".mfbd_flags = ");
1326 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1327 pandecode_log_cont(",\n");
1328
1329 if (fb->clear_stencil)
1330 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1331
1332 if (fb->clear_depth)
1333 pandecode_prop("clear_depth = %f", fb->clear_depth);
1334
1335 if (!is_compute)
1336 if (is_bifrost)
1337 pandecode_bifrost_tiler_descriptor(fb);
1338 else {
1339 const struct midgard_tiler_descriptor t = fb->tiler;
1340 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1341 }
1342 else
1343 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1344
1345 if (fb->zero3 || fb->zero4) {
1346 pandecode_msg("XXX: framebuffer zeros tripped\n");
1347 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1348 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1349 }
1350
1351 pandecode_indent--;
1352 pandecode_log("};\n");
1353
1354 gpu_va += sizeof(struct mali_framebuffer);
1355
1356 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1357
1358 if (info.has_extra) {
1359 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1360 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1361
1362 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1363 pandecode_indent++;
1364
1365 MEMORY_PROP(fbx, checksum);
1366
1367 if (fbx->checksum_stride)
1368 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1369
1370 pandecode_log(".flags_hi = ");
1371 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1372 pandecode_log_cont(",\n");
1373
1374 pandecode_log(".flags_lo = ");
1375 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1376 pandecode_log_cont(",\n");
1377
1378 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1379 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1380
1381 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1382 pandecode_log(".ds_afbc = {\n");
1383 pandecode_indent++;
1384
1385 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1386 pandecode_prop("depth_stencil_afbc_stride = %d",
1387 fbx->ds_afbc.depth_stencil_afbc_stride);
1388 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1389
1390 pandecode_log(".flags = ");
1391 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1392 pandecode_log_cont(",\n");
1393
1394 if (fbx->ds_afbc.padding) {
1395 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1396 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1397 }
1398
1399 pandecode_indent--;
1400 pandecode_log("},\n");
1401 } else {
1402 pandecode_log(".ds_linear = {\n");
1403 pandecode_indent++;
1404
1405 if (fbx->ds_linear.depth) {
1406 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1407 pandecode_prop("depth_stride = %d",
1408 fbx->ds_linear.depth_stride);
1409 pandecode_prop("depth_layer_stride = %d",
1410 fbx->ds_linear.depth_layer_stride);
1411 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1412 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1413 }
1414
1415 if (fbx->ds_linear.stencil) {
1416 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1417 pandecode_prop("stencil_stride = %d",
1418 fbx->ds_linear.stencil_stride);
1419 pandecode_prop("stencil_layer_stride = %d",
1420 fbx->ds_linear.stencil_layer_stride);
1421 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1422 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1423 }
1424
1425 if (fbx->ds_linear.depth_stride_zero ||
1426 fbx->ds_linear.stencil_stride_zero) {
1427 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1428 pandecode_prop("depth_stride_zero = 0x%x",
1429 fbx->ds_linear.depth_stride_zero);
1430 pandecode_prop("stencil_stride_zero = 0x%x",
1431 fbx->ds_linear.stencil_stride_zero);
1432 }
1433
1434 pandecode_indent--;
1435 pandecode_log("},\n");
1436 }
1437
1438 if (fbx->clear_color_1 | fbx->clear_color_2) {
1439 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1440 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1441 }
1442
1443 if (fbx->zero3) {
1444 pandecode_msg("XXX: fb_extra zeros tripped\n");
1445 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1446 }
1447
1448 pandecode_indent--;
1449 pandecode_log("};\n");
1450
1451 gpu_va += sizeof(struct mali_framebuffer_extra);
1452 }
1453
1454 if (is_fragment)
1455 pandecode_render_target(gpu_va, job_no, fb);
1456
1457 return info;
1458 }
1459
1460 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1461 * count */
1462
1463 static void
1464 pandecode_padded_vertices(unsigned shift, unsigned k)
1465 {
1466 unsigned odd = 2*k + 1;
1467 unsigned pot = 1 << shift;
1468 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1469 }
1470
1471 /* Given a magic divisor, recover what we were trying to divide by.
1472 *
1473 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1474 * 0 <= m < 2^N, for N bits in m.
1475 *
1476 * Let q represent the number we would like to divide by.
1477 *
1478 * By definition of a magic divisor for N-bit unsigned integers (a number you
1479 * multiply by to magically get division), m is a number such that:
1480 *
1481 * (m * x) & (2^N - 1) = floor(x/q).
1482 * for all x on Z where 0 <= x < 2^N
1483 *
1484 * Ignore the case where any of the above values equals zero; it is irrelevant
1485 * for our purposes (instanced arrays).
1486 *
1487 * Choose x = q. Then:
1488 *
1489 * (m * x) & (2^N - 1) = floor(x/q).
1490 * (m * q) & (2^N - 1) = floor(q/q).
1491 *
1492 * floor(q/q) = floor(1) = 1, therefore:
1493 *
1494 * (m * q) & (2^N - 1) = 1
1495 *
1496 * Recall the identity that the bitwise AND of one less than a power-of-two
1497 * equals the modulo with that power of two, i.e. for all x:
1498 *
1499 * x & (2^N - 1) = x % N
1500 *
1501 * Therefore:
1502 *
1503 * mq % (2^N) = 1
1504 *
1505 * By definition, a modular multiplicative inverse of a number m is the number
1506 * q such that with respect to a modulos M:
1507 *
1508 * mq % M = 1
1509 *
1510 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1511 *
1512 */
1513
1514 static void
1515 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1516 {
1517 #if 0
1518 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1519 * shift) the most lame way possible... just repeatedly add.
1520 * Asymptoptically slow but nobody cares in practice, unless you have
1521 * massive numbers of vertices or high divisors. */
1522
1523 unsigned inverse = 0;
1524
1525 /* Magic implicitly has the highest bit set */
1526 magic |= (1 << 31);
1527
1528 /* Depending on rounding direction */
1529 if (extra)
1530 magic++;
1531
1532 for (;;) {
1533 uint32_t product = magic * inverse;
1534
1535 if (shift) {
1536 product >>= shift;
1537 }
1538
1539 if (product == 1)
1540 break;
1541
1542 ++inverse;
1543 }
1544
1545 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1546
1547 /* Recall we're supposed to divide by (gl_level_divisor *
1548 * padded_num_vertices) */
1549
1550 unsigned padded_num_vertices = inverse / orig_divisor;
1551
1552 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1553 #endif
1554 }
1555
1556 static void
1557 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1558 mali_ptr addr, int job_no, char *suffix,
1559 int count, bool varying, enum mali_job_type job_type)
1560 {
1561 char *prefix = varying ? "varying" : "attribute";
1562 assert(addr);
1563
1564 if (!count) {
1565 pandecode_msg("warn: No %s records\n", prefix);
1566 return;
1567 }
1568
1569 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1570
1571 for (int i = 0; i < count; ++i) {
1572 /* First, check for special records */
1573 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1574 if (attr[i].size)
1575 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1576
1577 if (attr[i].stride) {
1578 /* gl_InstanceID passes a magic divisor in the
1579 * stride field to divide by the padded vertex
1580 * count. No other records should do so, so
1581 * stride should otherwise be zero. Note that
1582 * stride in the usual attribute sense doesn't
1583 * apply to special records. */
1584
1585 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1586
1587 pandecode_log_cont("/* %smagic divisor = %X */ ",
1588 has_divisor ? "" : "XXX: ", attr[i].stride);
1589 }
1590
1591 if (attr[i].shift || attr[i].extra_flags) {
1592 /* Attributes use these fields for
1593 * instancing/padding/etc type issues, but
1594 * varyings don't */
1595
1596 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1597 varying ? "XXX: " : "",
1598 attr[i].shift, attr[i].extra_flags);
1599 }
1600
1601 /* Print the special record name */
1602 bool attribute = false;
1603 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1604
1605 /* Sanity check */
1606 if (attribute == varying)
1607 pandecode_msg("XXX: mismatched special record\n");
1608
1609 continue;
1610 }
1611
1612 enum mali_attr_mode mode = attr[i].elements & 7;
1613
1614 if (mode == MALI_ATTR_UNUSED)
1615 pandecode_msg("XXX: unused attribute record\n");
1616
1617 /* For non-linear records, we need to print the type of record */
1618 if (mode != MALI_ATTR_LINEAR)
1619 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1620
1621 /* Print the name to link with attr_meta */
1622 pandecode_log_cont("%s_%d", prefix, i);
1623
1624 /* Print the stride and size */
1625 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1626
1627 /* TODO: Sanity check the quotient itself. It must be equal to
1628 * (or be greater than, if the driver added padding) the padded
1629 * vertex count. */
1630
1631 /* Finally, print the pointer */
1632 mali_ptr raw_elements = attr[i].elements & ~7;
1633 char *a = pointer_as_memory_reference(raw_elements);
1634 pandecode_log_cont(" = (%s);\n", a);
1635 free(a);
1636
1637 /* Check the pointer */
1638 pandecode_validate_buffer(raw_elements, attr[i].size);
1639
1640 /* shift/extra_flags exist only for instanced */
1641 if (attr[i].shift | attr[i].extra_flags) {
1642 /* These are set to random values by the blob for
1643 * varyings, most likely a symptom of uninitialized
1644 * memory where the hardware masked the bug. As such we
1645 * put this at a warning, not an error. */
1646
1647 if (mode == MALI_ATTR_LINEAR)
1648 pandecode_msg("warn: instancing fields set for linear\n");
1649
1650 pandecode_prop("shift = %d", attr[i].shift);
1651 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1652 }
1653
1654 /* Decode further where possible */
1655
1656 if (mode == MALI_ATTR_MODULO) {
1657 pandecode_padded_vertices(
1658 attr[i].shift,
1659 attr[i].extra_flags);
1660 }
1661
1662 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1663 i++;
1664 pandecode_log("{\n");
1665 pandecode_indent++;
1666 pandecode_prop("unk = 0x%x", attr[i].unk);
1667 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1668 if (attr[i].zero != 0)
1669 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1670 pandecode_prop("divisor = %d", attr[i].divisor);
1671 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1672 pandecode_indent--;
1673 pandecode_log("}, \n");
1674 }
1675
1676 }
1677
1678 pandecode_log("\n");
1679 }
1680
1681 static mali_ptr
1682 pandecode_shader_address(const char *name, mali_ptr ptr)
1683 {
1684 /* TODO: Decode flags */
1685 mali_ptr shader_ptr = ptr & ~15;
1686
1687 char *a = pointer_as_memory_reference(shader_ptr);
1688 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1689 free(a);
1690
1691 return shader_ptr;
1692 }
1693
1694 static void
1695 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1696 {
1697 unsigned any_nonzero =
1698 stencil->ref | stencil->mask | stencil->func |
1699 stencil->sfail | stencil->dpfail | stencil->dppass;
1700
1701 if (any_nonzero == 0)
1702 return;
1703
1704 const char *func = pandecode_func(stencil->func);
1705 const char *sfail = pandecode_stencil_op(stencil->sfail);
1706 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1707 const char *dppass = pandecode_stencil_op(stencil->dppass);
1708
1709 if (stencil->zero)
1710 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1711
1712 pandecode_log(".stencil_%s = {\n", name);
1713 pandecode_indent++;
1714 pandecode_prop("ref = %d", stencil->ref);
1715 pandecode_prop("mask = 0x%02X", stencil->mask);
1716 pandecode_prop("func = %s", func);
1717 pandecode_prop("sfail = %s", sfail);
1718 pandecode_prop("dpfail = %s", dpfail);
1719 pandecode_prop("dppass = %s", dppass);
1720 pandecode_indent--;
1721 pandecode_log("},\n");
1722 }
1723
1724 static void
1725 pandecode_blend_equation(const struct mali_blend_equation *blend)
1726 {
1727 if (blend->zero1)
1728 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1729
1730 pandecode_log(".equation = {\n");
1731 pandecode_indent++;
1732
1733 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1734 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1735
1736 pandecode_log(".color_mask = ");
1737 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1738 pandecode_log_cont(",\n");
1739
1740 pandecode_indent--;
1741 pandecode_log("},\n");
1742 }
1743
1744 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1745
1746 static unsigned
1747 decode_bifrost_constant(u16 constant)
1748 {
1749 float lo = (float) (constant & 0xFF);
1750 float hi = (float) (constant >> 8);
1751
1752 return (hi / 255.0) + (lo / 65535.0);
1753 }
1754
1755 static mali_ptr
1756 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1757 {
1758 struct bifrost_blend_rt *b =
1759 ((struct bifrost_blend_rt *) descs) + rt_no;
1760
1761 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1762 pandecode_indent++;
1763
1764 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1765 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1766 b->constant, decode_bifrost_constant(b->constant));
1767
1768 /* TODO figure out blend shader enable bit */
1769 pandecode_blend_equation(&b->equation);
1770
1771 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1772 pandecode_prop("index = 0x%" PRIx16, b->index);
1773
1774 pandecode_log(".format = ");
1775 pandecode_format_short(b->format, false);
1776 pandecode_swizzle(b->swizzle, b->format);
1777 pandecode_log_cont(",\n");
1778
1779 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1780 pandecode_prop("format = 0x%" PRIx32, b->format);
1781
1782 if (b->zero1) {
1783 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1784 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1785 }
1786
1787 pandecode_log(".shader_type = ");
1788 switch(b->shader_type) {
1789 case BIFROST_BLEND_F16:
1790 pandecode_log_cont("BIFROST_BLEND_F16");
1791 break;
1792 case BIFROST_BLEND_F32:
1793 pandecode_log_cont("BIFROST_BLEND_F32");
1794 break;
1795 case BIFROST_BLEND_I32:
1796 pandecode_log_cont("BIFROST_BLEND_I32");
1797 break;
1798 case BIFROST_BLEND_U32:
1799 pandecode_log_cont("BIFROST_BLEND_U32");
1800 break;
1801 case BIFROST_BLEND_I16:
1802 pandecode_log_cont("BIFROST_BLEND_I16");
1803 break;
1804 case BIFROST_BLEND_U16:
1805 pandecode_log_cont("BIFROST_BLEND_U16");
1806 break;
1807 }
1808 pandecode_log_cont(",\n");
1809
1810 if (b->zero2) {
1811 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1812 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1813 }
1814
1815 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1816
1817 pandecode_indent--;
1818 pandecode_log("},\n");
1819
1820 return 0;
1821 }
1822
1823 static mali_ptr
1824 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1825 {
1826 /* constant/equation is in a union */
1827 if (!blend->shader)
1828 return 0;
1829
1830 pandecode_log(".blend = {\n");
1831 pandecode_indent++;
1832
1833 if (is_shader) {
1834 pandecode_shader_address("shader", blend->shader);
1835 } else {
1836 pandecode_blend_equation(&blend->equation);
1837 pandecode_prop("constant = %f", blend->constant);
1838 }
1839
1840 pandecode_indent--;
1841 pandecode_log("},\n");
1842
1843 /* Return blend shader to disassemble if present */
1844 return is_shader ? (blend->shader & ~0xF) : 0;
1845 }
1846
1847 static mali_ptr
1848 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1849 {
1850 struct midgard_blend_rt *b =
1851 ((struct midgard_blend_rt *) descs) + rt_no;
1852
1853 /* Flags determine presence of blend shader */
1854 bool is_shader = (b->flags & 0xF) >= 0x2;
1855
1856 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1857 pandecode_indent++;
1858
1859 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1860
1861 union midgard_blend blend = b->blend;
1862 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1863
1864 pandecode_indent--;
1865 pandecode_log("};\n");
1866
1867 return shader;
1868 }
1869
1870 /* Attributes and varyings have descriptor records, which contain information
1871 * about their format and ordering with the attribute/varying buffers. We'll
1872 * want to validate that the combinations specified are self-consistent.
1873 */
1874
1875 static int
1876 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1877 {
1878 char base[128];
1879 char *prefix = varying ? "varying" : "attribute";
1880 unsigned max_index = 0;
1881 snprintf(base, sizeof(base), "%s_meta", prefix);
1882
1883 struct mali_attr_meta *attr_meta;
1884 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1885
1886 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1887
1888 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1889 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1890 sizeof(*attr_mem));
1891
1892 /* If the record is discard, it should be zero for everything else */
1893
1894 if (attr_meta->format == MALI_VARYING_DISCARD) {
1895 uint64_t zero =
1896 attr_meta->index |
1897 attr_meta->unknown1 |
1898 attr_meta->unknown3 |
1899 attr_meta->src_offset;
1900
1901 if (zero)
1902 pandecode_msg("XXX: expected empty record for varying discard\n");
1903
1904 /* We want to look for a literal 0000 swizzle -- this
1905 * is not encoded with all zeroes, however */
1906
1907 enum mali_channel z = MALI_CHANNEL_ZERO;
1908 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1909 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1910
1911 if (!good_swizzle)
1912 pandecode_msg("XXX: expected zero swizzle for discard\n");
1913
1914 if (!varying)
1915 pandecode_msg("XXX: cannot discard attribute\n");
1916
1917 /* If we're all good, omit the record */
1918 if (!zero && varying && good_swizzle) {
1919 pandecode_log("/* discarded varying */\n");
1920 continue;
1921 }
1922 }
1923
1924 if (attr_meta->index > max_index)
1925 max_index = attr_meta->index;
1926
1927 if (attr_meta->unknown1 != 0x2) {
1928 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1929 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1930 }
1931
1932 if (attr_meta->unknown3) {
1933 pandecode_msg("XXX: unexpected unknown3 set\n");
1934 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1935 }
1936
1937 pandecode_format_short(attr_meta->format, false);
1938 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1939
1940 if (attr_meta->src_offset)
1941 pandecode_log_cont("[%u]", attr_meta->src_offset);
1942
1943 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1944
1945 pandecode_log_cont(";\n");
1946 }
1947
1948 pandecode_log("\n");
1949
1950 return count ? (max_index + 1) : 0;
1951 }
1952
1953 /* return bits [lo, hi) of word */
1954 static u32
1955 bits(u32 word, u32 lo, u32 hi)
1956 {
1957 if (hi - lo >= 32)
1958 return word; // avoid undefined behavior with the shift
1959
1960 return (word >> lo) & ((1 << (hi - lo)) - 1);
1961 }
1962
1963 static void
1964 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1965 {
1966 pandecode_log(".prefix = {\n");
1967 pandecode_indent++;
1968
1969 /* Decode invocation_count. See the comment before the definition of
1970 * invocation_count for an explanation.
1971 */
1972
1973 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1974 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1975 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1976 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1977 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1978 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1979
1980 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1981 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1982 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1983
1984 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1985 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1986 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1987
1988 /* Even though we have this decoded, we want to ensure that the
1989 * representation is "unique" so we don't lose anything by printing only
1990 * the final result. More specifically, we need to check that we were
1991 * passed something in canonical form, since the definition per the
1992 * hardware is inherently not unique. How? Well, take the resulting
1993 * decode and pack it ourselves! If it is bit exact with what we
1994 * decoded, we're good to go. */
1995
1996 struct mali_vertex_tiler_prefix ref;
1997 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1998
1999 bool canonical =
2000 (p->invocation_count == ref.invocation_count) &&
2001 (p->invocation_shifts == ref.invocation_shifts);
2002
2003 if (!canonical) {
2004 pandecode_msg("XXX: non-canonical workgroups packing\n");
2005 pandecode_msg("expected: %X, %X",
2006 ref.invocation_count,
2007 ref.invocation_shifts);
2008
2009 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
2010 pandecode_prop("size_y_shift = %d", size_y_shift);
2011 pandecode_prop("size_z_shift = %d", size_z_shift);
2012 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
2013 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
2014 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
2015 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
2016 }
2017
2018 /* Regardless, print the decode */
2019 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
2020 size_x, size_y, size_z,
2021 groups_x, groups_y, groups_z);
2022
2023 /* TODO: Decode */
2024 if (p->unknown_draw)
2025 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
2026
2027 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
2028
2029 if (p->draw_mode != MALI_DRAW_NONE)
2030 pandecode_prop("draw_mode = %s", pandecode_draw_mode(p->draw_mode));
2031
2032 /* Index count only exists for tiler jobs anyway */
2033
2034 if (p->index_count)
2035 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
2036
2037
2038 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
2039 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
2040
2041 /* Validate an index buffer is present if we need one. TODO: verify
2042 * relationship between invocation_count and index_count */
2043
2044 if (p->indices) {
2045 unsigned count = p->index_count;
2046
2047 /* Grab the size */
2048 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
2049
2050 /* Ensure we got a size, and if so, validate the index buffer
2051 * is large enough to hold a full set of indices of the given
2052 * size */
2053
2054 if (!index_raw_size)
2055 pandecode_msg("XXX: index size missing\n");
2056 else
2057 pandecode_validate_buffer(p->indices, count * size);
2058 } else if (index_raw_size)
2059 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
2060
2061 if (p->offset_bias_correction)
2062 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
2063
2064 /* TODO: Figure out what this is. It's not zero */
2065 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
2066
2067 pandecode_indent--;
2068 pandecode_log("},\n");
2069 }
2070
2071 static void
2072 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
2073 {
2074 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
2075 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
2076
2077 for (int i = 0; i < ubufs_count; i++) {
2078 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
2079 mali_ptr addr = (ubufs[i] >> 10) << 2;
2080
2081 pandecode_validate_buffer(addr, size);
2082
2083 char *ptr = pointer_as_memory_reference(addr);
2084 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
2085 free(ptr);
2086 }
2087
2088 pandecode_log("\n");
2089 }
2090
2091 static void
2092 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
2093 {
2094 pandecode_validate_buffer(uniforms, uniform_count * 16);
2095
2096 char *ptr = pointer_as_memory_reference(uniforms);
2097 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
2098 free(ptr);
2099 }
2100
2101 static const char *
2102 shader_type_for_job(unsigned type)
2103 {
2104 switch (type) {
2105 case JOB_TYPE_VERTEX: return "VERTEX";
2106 case JOB_TYPE_TILER: return "FRAGMENT";
2107 case JOB_TYPE_COMPUTE: return "COMPUTE";
2108 default:
2109 return "UNKNOWN";
2110 }
2111 }
2112
2113 static unsigned shader_id = 0;
2114
2115 static struct midgard_disasm_stats
2116 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
2117 bool is_bifrost, unsigned gpu_id)
2118 {
2119 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
2120 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
2121
2122 /* Compute maximum possible size */
2123 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
2124
2125 /* Print some boilerplate to clearly denote the assembly (which doesn't
2126 * obey indentation rules), and actually do the disassembly! */
2127
2128 pandecode_log_cont("\n\n");
2129
2130 struct midgard_disasm_stats stats;
2131
2132 if (is_bifrost) {
2133 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
2134
2135 /* TODO: Extend stats to Bifrost */
2136 stats.texture_count = -128;
2137 stats.sampler_count = -128;
2138 stats.attribute_count = -128;
2139 stats.varying_count = -128;
2140 stats.uniform_count = -128;
2141 stats.uniform_buffer_count = -128;
2142 stats.work_count = -128;
2143
2144 stats.instruction_count = 0;
2145 stats.bundle_count = 0;
2146 stats.quadword_count = 0;
2147 stats.helper_invocations = false;
2148 } else {
2149 stats = disassemble_midgard(pandecode_dump_stream,
2150 code, sz, gpu_id,
2151 type == JOB_TYPE_TILER ?
2152 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
2153 }
2154
2155 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2156 * driver-internal purposes with the blob and interfere */
2157
2158 bool should_shaderdb = type != JOB_TYPE_COMPUTE;
2159
2160 if (should_shaderdb) {
2161 unsigned nr_threads =
2162 (stats.work_count <= 4) ? 4 :
2163 (stats.work_count <= 8) ? 2 :
2164 1;
2165
2166 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2167 "%u inst, %u bundles, %u quadwords, "
2168 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2169 shader_id++,
2170 shader_type_for_job(type),
2171 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2172 stats.work_count, nr_threads);
2173 }
2174
2175
2176 return stats;
2177 }
2178
2179 static void
2180 pandecode_texture_payload(mali_ptr payload,
2181 enum mali_texture_type type,
2182 enum mali_texture_layout layout,
2183 bool manual_stride,
2184 uint8_t levels,
2185 uint16_t depth,
2186 uint16_t array_size,
2187 struct pandecode_mapped_memory *tmem)
2188 {
2189 pandecode_log(".payload = {\n");
2190 pandecode_indent++;
2191
2192 /* A bunch of bitmap pointers follow.
2193 * We work out the correct number,
2194 * based on the mipmap/cubemap
2195 * properties, but dump extra
2196 * possibilities to futureproof */
2197
2198 int bitmap_count = levels + 1;
2199
2200 /* Miptree for each face */
2201 if (type == MALI_TEX_CUBE)
2202 bitmap_count *= 6;
2203
2204 /* Array of layers */
2205 bitmap_count *= (depth + 1);
2206
2207 /* Array of textures */
2208 bitmap_count *= (array_size + 1);
2209
2210 /* Stride for each element */
2211 if (manual_stride)
2212 bitmap_count *= 2;
2213
2214 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2215 payload, sizeof(mali_ptr) * bitmap_count);
2216 for (int i = 0; i < bitmap_count; ++i) {
2217 /* How we dump depends if this is a stride or a pointer */
2218
2219 if (manual_stride && (i & 1)) {
2220 /* signed 32-bit snuck in as a 64-bit pointer */
2221 uint64_t stride_set = pointers_and_strides[i];
2222 uint32_t clamped_stride = stride_set;
2223 int32_t stride = clamped_stride;
2224 assert(stride_set == clamped_stride);
2225 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2226 } else {
2227 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2228 pandecode_log("%s, \n", a);
2229 free(a);
2230 }
2231 }
2232
2233 pandecode_indent--;
2234 pandecode_log("},\n");
2235 }
2236
2237 static void
2238 pandecode_texture(mali_ptr u,
2239 struct pandecode_mapped_memory *tmem,
2240 unsigned job_no, unsigned tex)
2241 {
2242 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2243
2244 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2245 pandecode_indent++;
2246
2247 pandecode_prop("width = %" PRId32, t->width);
2248 pandecode_prop("height = %" PRId32, t->height);
2249 pandecode_prop("depth = %" PRId32, t->depth);
2250 pandecode_prop("array_size = %" PRId32, t->array_size);
2251
2252 pandecode_log("\n");
2253 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2254 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2255 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2256 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2257 pandecode_prop("f.type = %" PRId32, t->format.type);
2258 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2259 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2260 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2261 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2262 pandecode_log("\n");
2263
2264 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2265 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2266 pandecode_prop("levels = %" PRId32, t->levels);
2267 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2268 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2269
2270 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2271 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2272 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2273 pandecode_log("\n");
2274
2275 struct mali_texture_format f = t->format;
2276
2277 /* See the definiton of enum mali_texture_type */
2278
2279 bool is_cube = f.type == MALI_TEX_CUBE;
2280 unsigned dimension = is_cube ? 2 : f.type;
2281
2282 pandecode_make_indent();
2283
2284 /* Print the layout. Default is linear; a modifier can denote AFBC or
2285 * u-interleaved/tiled modes */
2286
2287 if (f.layout == MALI_TEXTURE_AFBC)
2288 pandecode_log_cont("afbc");
2289 else if (f.layout == MALI_TEXTURE_TILED)
2290 pandecode_log_cont("tiled");
2291 else if (f.layout == MALI_TEXTURE_LINEAR)
2292 pandecode_log_cont("linear");
2293 else
2294 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2295
2296 pandecode_swizzle(t->swizzle, f.format);
2297 pandecode_log_cont(" ");
2298
2299 /* Distinguish cube/2D with modifier */
2300
2301 if (is_cube)
2302 pandecode_log_cont("cube ");
2303
2304 pandecode_format_short(f.format, f.srgb);
2305 pandecode_swizzle(f.swizzle, f.format);
2306
2307 /* All four width/height/depth/array_size dimensions are present
2308 * regardless of the type of texture, but it is an error to have
2309 * non-zero dimensions for unused dimensions. Verify this. array_size
2310 * can always be set, as can width. Depth used for MSAA. */
2311
2312 if (t->height && dimension < 2)
2313 pandecode_msg("XXX: nonzero height for <2D texture\n");
2314
2315 /* Print only the dimensions that are actually there */
2316
2317 pandecode_log_cont(": %d", t->width + 1);
2318
2319 if (t->height || t->depth)
2320 pandecode_log_cont("x%u", t->height + 1);
2321
2322 if (t->depth)
2323 pandecode_log_cont("x%u", t->depth + 1);
2324
2325 if (t->array_size)
2326 pandecode_log_cont("[%u]", t->array_size + 1);
2327
2328 if (t->levels)
2329 pandecode_log_cont(" mip %u", t->levels);
2330
2331 pandecode_log_cont("\n");
2332
2333 if (f.unknown1 | f.zero) {
2334 pandecode_msg("XXX: texture format zero tripped\n");
2335 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2336 pandecode_prop("zero = %" PRId32, f.zero);
2337 }
2338
2339 if (!f.unknown2) {
2340 pandecode_msg("XXX: expected unknown texture bit set\n");
2341 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2342 }
2343
2344 if (t->swizzle_zero) {
2345 pandecode_msg("XXX: swizzle zero tripped\n");
2346 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2347 }
2348
2349 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2350 pandecode_msg("XXX: texture zero tripped\n");
2351 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2352 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2353 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2354 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2355 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2356 }
2357
2358 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2359
2360 pandecode_indent--;
2361 pandecode_log("};\n");
2362 }
2363
2364 static void
2365 pandecode_bifrost_texture(
2366 const struct bifrost_texture_descriptor *t,
2367 unsigned job_no,
2368 unsigned tex)
2369 {
2370 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2371 pandecode_indent++;
2372
2373 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2374 pandecode_prop("type = %" PRId32, t->type);
2375
2376 if (t->zero) {
2377 pandecode_msg("XXX: zero tripped\n");
2378 pandecode_prop("zero = 0x%" PRIx32, t->zero);
2379 }
2380
2381 pandecode_prop("format_swizzle = 0x%" PRIx32, t->format_swizzle);
2382 pandecode_prop("format = 0x%" PRIx32, t->format);
2383 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2384 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2385 pandecode_prop("width = %" PRId32, t->width);
2386 pandecode_prop("height = %" PRId32, t->height);
2387 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2388 pandecode_prop("levels = %" PRId32, t->levels);
2389 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2390 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2391 pandecode_prop("level_2 = %" PRId32, t->level_2);
2392 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2393 pandecode_prop("array_size = %" PRId32, t->array_size);
2394 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2395 pandecode_prop("depth = %" PRId32, t->depth);
2396 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2397 pandecode_log("\n");
2398
2399 /* See the definiton of enum mali_texture_type */
2400
2401 bool is_cube = t->type == MALI_TEX_CUBE;
2402 unsigned dimension = is_cube ? 2 : t->type;
2403
2404 /* Print the layout. Default is linear; a modifier can denote AFBC or
2405 * u-interleaved/tiled modes */
2406
2407 if (t->layout == MALI_TEXTURE_AFBC)
2408 pandecode_log_cont("afbc");
2409 else if (t->layout == MALI_TEXTURE_TILED)
2410 pandecode_log_cont("tiled");
2411 else if (t->layout == MALI_TEXTURE_LINEAR)
2412 pandecode_log_cont("linear");
2413 else
2414 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2415
2416 pandecode_swizzle(t->swizzle, t->format);
2417 pandecode_log_cont(" ");
2418
2419 /* Distinguish cube/2D with modifier */
2420
2421 if (is_cube)
2422 pandecode_log_cont("cube ");
2423
2424 pandecode_format_short(t->format, t->srgb);
2425
2426 /* All four width/height/depth/array_size dimensions are present
2427 * regardless of the type of texture, but it is an error to have
2428 * non-zero dimensions for unused dimensions. Verify this. array_size
2429 * can always be set, as can width. */
2430
2431 if (t->height && dimension < 2)
2432 pandecode_msg("XXX: nonzero height for <2D texture\n");
2433
2434 if (t->depth && dimension < 3)
2435 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2436
2437 /* Print only the dimensions that are actually there */
2438
2439 pandecode_log_cont(": %d", t->width + 1);
2440
2441 if (dimension >= 2)
2442 pandecode_log_cont("x%u", t->height + 1);
2443
2444 if (dimension >= 3)
2445 pandecode_log_cont("x%u", t->depth + 1);
2446
2447 if (t->array_size)
2448 pandecode_log_cont("[%u]", t->array_size + 1);
2449
2450 if (t->levels)
2451 pandecode_log_cont(" mip %u", t->levels);
2452
2453 pandecode_log_cont("\n");
2454
2455 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2456 if (t->payload) {
2457 pandecode_texture_payload(t->payload, t->type, t->layout,
2458 true, t->levels, t->depth,
2459 t->array_size, tmem);
2460 }
2461
2462 pandecode_indent--;
2463 pandecode_log("};\n");
2464 }
2465
2466 /* 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 */
2467
2468 static void
2469 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2470 {
2471 /* Nothing to do */
2472 if (claim == truth)
2473 return;
2474
2475 if (fuzzy && (truth < 0))
2476 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2477
2478 if ((truth >= 0) && !fuzzy) {
2479 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2480 (truth < claim) ? "warn" : "XXX",
2481 name, truth, claim);
2482 } else if ((claim > -truth) && !fuzzy) {
2483 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2484 name, -truth, claim);
2485 } else if (fuzzy && (claim < truth))
2486 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2487 name, truth, claim);
2488
2489 pandecode_log(".%s = %" PRId16, name, claim);
2490
2491 if (fuzzy)
2492 pandecode_log_cont(" /* %u used */", truth);
2493
2494 pandecode_log_cont(",\n");
2495 }
2496
2497 static void
2498 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2499 bool is_bifrost, unsigned gpu_id)
2500 {
2501 struct midgard_disasm_stats stats =
2502 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2503
2504 bool has_texture = (stats.texture_count > 0);
2505 bool has_sampler = (stats.sampler_count > 0);
2506 bool has_attribute = (stats.attribute_count > 0);
2507 bool has_varying = (stats.varying_count > 0);
2508 bool has_uniform = (stats.uniform_count > 0);
2509 bool has_ubo = (stats.uniform_buffer_count > 0);
2510
2511 if (has_texture || has_sampler)
2512 pandecode_msg("XXX: blend shader accessing textures\n");
2513
2514 if (has_attribute || has_varying)
2515 pandecode_msg("XXX: blend shader accessing interstage\n");
2516
2517 if (has_uniform || has_ubo)
2518 pandecode_msg("XXX: blend shader accessing uniforms\n");
2519 }
2520
2521 static void
2522 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2523 {
2524 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2525
2526 if (!mmem)
2527 return;
2528
2529 if (is_bifrost) {
2530 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2531
2532 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2533 pandecode_indent++;
2534
2535 for (unsigned tex = 0; tex < texture_count; ++tex)
2536 pandecode_bifrost_texture(&t[tex], job_no, tex);
2537
2538 pandecode_indent--;
2539 pandecode_log("};\n");
2540 } else {
2541 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2542
2543 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2544 pandecode_indent++;
2545
2546 for (int tex = 0; tex < texture_count; ++tex) {
2547 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2548 char *a = pointer_as_memory_reference(*u);
2549 pandecode_log("%s,\n", a);
2550 free(a);
2551 }
2552
2553 pandecode_indent--;
2554 pandecode_log("};\n");
2555
2556 /* Now, finally, descend down into the texture descriptor */
2557 for (unsigned tex = 0; tex < texture_count; ++tex) {
2558 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2559 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2560 if (tmem)
2561 pandecode_texture(*u, tmem, job_no, tex);
2562 }
2563 }
2564 }
2565
2566 static void
2567 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2568 {
2569 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2570
2571 if (!smem)
2572 return;
2573
2574 if (is_bifrost) {
2575 struct bifrost_sampler_descriptor *s;
2576
2577 for (int i = 0; i < sampler_count; ++i) {
2578 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2579
2580 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2581 pandecode_indent++;
2582
2583 if (s->unk1 != 1) {
2584 pandecode_msg("XXX: unk1 tripped\n");
2585 pandecode_prop("unk1 = 0x%x", s->unk1);
2586 }
2587
2588 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2589 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2590 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2591
2592 if (s->unk8 != 0x8) {
2593 pandecode_msg("XXX: unk8 tripped\n");
2594 pandecode_prop("unk8 = 0x%x", s->unk8);
2595 }
2596
2597 pandecode_prop("unk2 = 0x%x", s->unk2);
2598 pandecode_prop("unk3 = 0x%x", s->unk3);
2599 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2600 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2601 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2602 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2603 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2604
2605 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2606 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2607
2608 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2609 pandecode_msg("XXX: sampler zero tripped\n");
2610 pandecode_prop("zero = 0x%" PRIx8 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2611 }
2612
2613 pandecode_indent--;
2614 pandecode_log("};\n");
2615 }
2616 } else {
2617 struct mali_sampler_descriptor *s;
2618
2619 for (int i = 0; i < sampler_count; ++i) {
2620 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2621
2622 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2623 pandecode_indent++;
2624
2625 pandecode_log(".filter_mode = ");
2626 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2627 pandecode_log_cont(",\n");
2628
2629 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2630 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2631
2632 if (s->lod_bias)
2633 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2634
2635 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2636 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2637 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2638
2639 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2640
2641 if (s->zero || s->zero2) {
2642 pandecode_msg("XXX: sampler zero tripped\n");
2643 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2644 }
2645
2646 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2647
2648 pandecode_prop("border_color = { %f, %f, %f, %f }",
2649 s->border_color[0],
2650 s->border_color[1],
2651 s->border_color[2],
2652 s->border_color[3]);
2653
2654 pandecode_indent--;
2655 pandecode_log("};\n");
2656 }
2657 }
2658 }
2659
2660 static void
2661 pandecode_vertex_tiler_postfix_pre(
2662 const struct mali_vertex_tiler_postfix *p,
2663 int job_no, enum mali_job_type job_type,
2664 char *suffix, bool is_bifrost, unsigned gpu_id)
2665 {
2666 struct pandecode_mapped_memory *attr_mem;
2667
2668 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2669 * are the only things actually needed from the FBD, vertex/tiler jobs
2670 * no longer reference the FBD -- instead, this field points to some
2671 * info about the scratchpad.
2672 */
2673
2674 struct pandecode_fbd fbd_info = {
2675 /* Default for Bifrost */
2676 .rt_count = 1
2677 };
2678
2679 if (is_bifrost) {
2680 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2681 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2682 } else if (p->shared_memory & MALI_MFBD)
2683 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == JOB_TYPE_COMPUTE, false);
2684 else if (job_type == JOB_TYPE_COMPUTE)
2685 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2686 else
2687 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2688
2689 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2690 int texture_count = 0, sampler_count = 0;
2691
2692 if (p->shader) {
2693 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2694 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2695
2696 /* Disassemble ahead-of-time to get stats. Initialize with
2697 * stats for the missing-shader case so we get validation
2698 * there, too */
2699
2700 struct midgard_disasm_stats info = {
2701 .texture_count = 0,
2702 .sampler_count = 0,
2703 .attribute_count = 0,
2704 .varying_count = 0,
2705 .work_count = 1,
2706
2707 .uniform_count = -128,
2708 .uniform_buffer_count = 0
2709 };
2710
2711 if (s->shader & ~0xF)
2712 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2713
2714 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2715 pandecode_indent++;
2716
2717 /* Save for dumps */
2718 attribute_count = s->attribute_count;
2719 varying_count = s->varying_count;
2720 texture_count = s->texture_count;
2721 sampler_count = s->sampler_count;
2722
2723 if (is_bifrost) {
2724 uniform_count = s->bifrost2.uniform_count;
2725 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2726 } else {
2727 uniform_count = s->midgard1.uniform_count;
2728 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2729 }
2730
2731 pandecode_shader_address("shader", s->shader);
2732
2733 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2734 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2735 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2736 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2737 pandecode_shader_prop("uniform_buffer_count",
2738 uniform_buffer_count,
2739 info.uniform_buffer_count, true);
2740
2741 if (!is_bifrost) {
2742 pandecode_shader_prop("uniform_count",
2743 uniform_count,
2744 info.uniform_count, false);
2745
2746 pandecode_shader_prop("work_count",
2747 s->midgard1.work_count, info.work_count, false);
2748 }
2749
2750 if (is_bifrost) {
2751 pandecode_log("bifrost1.unk1 = ");
2752 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2753 pandecode_log_cont(",\n");
2754 } else {
2755 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2756
2757 if (helpers != info.helper_invocations) {
2758 pandecode_msg("XXX: expected helpers %u but got %u\n",
2759 info.helper_invocations, helpers);
2760 }
2761
2762 pandecode_log(".midgard1.flags_lo = ");
2763 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2764 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2765 pandecode_log_cont(",\n");
2766
2767 pandecode_log(".midgard1.flags_hi = ");
2768 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2769 pandecode_log_cont(",\n");
2770 }
2771
2772 if (s->depth_units || s->depth_factor) {
2773 pandecode_prop("depth_factor = %f", s->depth_factor);
2774 pandecode_prop("depth_units = %f", s->depth_units);
2775 }
2776
2777 if (s->coverage_mask)
2778 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
2779
2780 if (s->unknown2_2)
2781 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2782
2783 if (s->unknown2_3 || s->unknown2_4) {
2784 pandecode_log(".unknown2_3 = ");
2785
2786 int unknown2_3 = s->unknown2_3;
2787 int unknown2_4 = s->unknown2_4;
2788
2789 /* We're not quite sure what these flags mean without the depth test, if anything */
2790
2791 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2792 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2793 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2794
2795 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2796 }
2797
2798 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2799 pandecode_log_cont(",\n");
2800
2801 pandecode_log(".unknown2_4 = ");
2802 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2803 pandecode_log_cont(",\n");
2804 }
2805
2806 if (s->stencil_mask_front || s->stencil_mask_back) {
2807 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2808 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2809 }
2810
2811 pandecode_stencil("front", &s->stencil_front);
2812 pandecode_stencil("back", &s->stencil_back);
2813
2814 if (is_bifrost) {
2815 pandecode_log(".bifrost2 = {\n");
2816 pandecode_indent++;
2817
2818 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2819 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2820 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2821 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2822
2823 pandecode_indent--;
2824 pandecode_log("},\n");
2825 } else if (s->midgard2.unknown2_7) {
2826 pandecode_log(".midgard2 = {\n");
2827 pandecode_indent++;
2828
2829 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2830 pandecode_indent--;
2831 pandecode_log("},\n");
2832 }
2833
2834 if (s->padding) {
2835 pandecode_msg("XXX: shader padding tripped\n");
2836 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2837 }
2838
2839 if (!is_bifrost) {
2840 /* TODO: Blend shaders routing/disasm */
2841 union midgard_blend blend = s->blend;
2842 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2843 if (shader & ~0xF)
2844 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2845 } else {
2846 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2847 }
2848
2849 pandecode_indent--;
2850 pandecode_log("};\n");
2851
2852 /* MRT blend fields are used whenever MFBD is used, with
2853 * per-RT descriptors */
2854
2855 if (job_type == JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2856 void* blend_base = (void *) (s + 1);
2857
2858 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2859 mali_ptr shader = 0;
2860
2861 if (is_bifrost)
2862 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2863 else
2864 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2865
2866 if (shader & ~0xF)
2867 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2868
2869 }
2870 }
2871 } else
2872 pandecode_msg("XXX: missing shader descriptor\n");
2873
2874 if (p->viewport) {
2875 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2876 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2877
2878 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2879 pandecode_indent++;
2880
2881 pandecode_prop("clip_minx = %f", f->clip_minx);
2882 pandecode_prop("clip_miny = %f", f->clip_miny);
2883 pandecode_prop("clip_minz = %f", f->clip_minz);
2884 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2885 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2886 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2887
2888 /* Only the higher coordinates are MALI_POSITIVE scaled */
2889
2890 pandecode_prop("viewport0 = { %d, %d }",
2891 f->viewport0[0], f->viewport0[1]);
2892
2893 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2894 f->viewport1[0] + 1, f->viewport1[1] + 1);
2895
2896 pandecode_indent--;
2897 pandecode_log("};\n");
2898 }
2899
2900 unsigned max_attr_index = 0;
2901
2902 if (p->attribute_meta)
2903 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2904
2905 if (p->attributes) {
2906 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2907 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2908 }
2909
2910 /* Varyings are encoded like attributes but not actually sent; we just
2911 * pass a zero buffer with the right stride/size set, (or whatever)
2912 * since the GPU will write to it itself */
2913
2914 if (p->varying_meta) {
2915 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2916 }
2917
2918 if (p->varyings) {
2919 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2920
2921 /* Number of descriptors depends on whether there are
2922 * non-internal varyings */
2923
2924 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2925 }
2926
2927 if (p->uniform_buffers) {
2928 if (uniform_buffer_count)
2929 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2930 else
2931 pandecode_msg("warn: UBOs specified but not referenced\n");
2932 } else if (uniform_buffer_count)
2933 pandecode_msg("XXX: UBOs referenced but not specified\n");
2934
2935 /* We don't want to actually dump uniforms, but we do need to validate
2936 * that the counts we were given are sane */
2937
2938 if (p->uniforms) {
2939 if (uniform_count)
2940 pandecode_uniforms(p->uniforms, uniform_count);
2941 else
2942 pandecode_msg("warn: Uniforms specified but not referenced\n");
2943 } else if (uniform_count)
2944 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2945
2946 if (p->textures)
2947 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2948
2949 if (p->sampler_descriptor)
2950 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2951 }
2952
2953 static void
2954 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2955 {
2956 pandecode_log(".gl_enables = ");
2957
2958 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2959
2960 pandecode_log_cont(",\n");
2961 }
2962
2963 static void
2964 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2965 {
2966 if (p->shader & 0xF)
2967 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2968
2969 pandecode_log(".postfix = {\n");
2970 pandecode_indent++;
2971
2972 pandecode_gl_enables(p->gl_enables, JOB_TYPE_TILER);
2973 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2974 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2975
2976 if (p->zero4) {
2977 pandecode_msg("XXX: vertex only zero tripped");
2978 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2979 }
2980
2981 pandecode_prop("offset_start = 0x%x", p->offset_start);
2982
2983 if (p->zero5) {
2984 pandecode_msg("XXX: vertex only zero tripped");
2985 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2986 }
2987
2988 MEMORY_PROP(p, position_varying);
2989 MEMORY_PROP(p, occlusion_counter);
2990
2991 pandecode_indent--;
2992 pandecode_log("},\n");
2993 }
2994
2995 static void
2996 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2997 {
2998 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2999 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
3000
3001 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
3002 pandecode_indent++;
3003
3004 if (h->zero) {
3005 pandecode_msg("XXX: tiler heap zero tripped\n");
3006 pandecode_prop("zero = 0x%x", h->zero);
3007 }
3008
3009 pandecode_prop("heap_size = 0x%x", h->heap_size);
3010 MEMORY_PROP(h, tiler_heap_start);
3011 MEMORY_PROP(h, tiler_heap_free);
3012
3013 /* this might point to the beginning of another buffer, when it's
3014 * really the end of the tiler heap buffer, so we have to be careful
3015 * here. but for zero length, we need the same pointer.
3016 */
3017
3018 if (h->tiler_heap_end == h->tiler_heap_start) {
3019 MEMORY_PROP(h, tiler_heap_start);
3020 } else {
3021 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
3022 pandecode_prop("tiler_heap_end = %s + 1", a);
3023 free(a);
3024 }
3025
3026 for (int i = 0; i < 10; i++) {
3027 if (h->zeros[i] != 0) {
3028 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
3029 i, h->zeros[i]);
3030 }
3031 }
3032
3033 if (h->unk1 != 0x1) {
3034 pandecode_msg("XXX: tiler heap unk1 tripped\n");
3035 pandecode_prop("unk1 = 0x%x", h->unk1);
3036 }
3037
3038 if (h->unk7e007e != 0x7e007e) {
3039 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
3040 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
3041 }
3042
3043 pandecode_indent--;
3044 pandecode_log("};\n");
3045 }
3046
3047 static void
3048 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
3049 {
3050 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
3051 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
3052
3053 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
3054
3055 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
3056 pandecode_indent++;
3057
3058 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
3059 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
3060
3061 if (t->hierarchy_mask != 0xa &&
3062 t->hierarchy_mask != 0x14 &&
3063 t->hierarchy_mask != 0x28 &&
3064 t->hierarchy_mask != 0x50 &&
3065 t->hierarchy_mask != 0xa0)
3066 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
3067
3068 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
3069
3070 pandecode_prop("flags = 0x%" PRIx16, t->flags);
3071
3072 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
3073 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
3074
3075 if (t->zero0) {
3076 pandecode_msg("XXX: tiler meta zero tripped\n");
3077 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
3078 }
3079
3080 for (int i = 0; i < 12; i++) {
3081 if (t->zeros[i] != 0) {
3082 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
3083 i, t->zeros[i]);
3084 }
3085 }
3086
3087 pandecode_indent--;
3088 pandecode_log("};\n");
3089 }
3090
3091 static void
3092 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
3093 {
3094 if (u.pointer == 0x0)
3095 return;
3096
3097 pandecode_log(".primitive_size = {\n");
3098 pandecode_indent++;
3099
3100 if (constant) {
3101 pandecode_prop("constant = %f", u.constant);
3102 } else {
3103 MEMORY_PROP((&u), pointer);
3104 }
3105
3106 pandecode_indent--;
3107 pandecode_log("},\n");
3108 }
3109
3110 static void
3111 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
3112 {
3113 pandecode_log_cont("{\n");
3114 pandecode_indent++;
3115
3116 /* TODO: gl_PointSize on Bifrost */
3117 pandecode_primitive_size(t->primitive_size, true);
3118
3119 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
3120 || t->zero6) {
3121 pandecode_msg("XXX: tiler only zero tripped\n");
3122 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
3123 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
3124 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
3125 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
3126 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
3127 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
3128 }
3129
3130 pandecode_indent--;
3131 pandecode_log("},\n");
3132 }
3133
3134 static int
3135 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
3136 const struct pandecode_mapped_memory *mem,
3137 mali_ptr payload, int job_no, unsigned gpu_id)
3138 {
3139 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
3140
3141 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
3142
3143 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
3144 pandecode_indent++;
3145
3146 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
3147 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
3148
3149 pandecode_indent--;
3150 pandecode_log("};\n");
3151
3152 return sizeof(*v);
3153 }
3154
3155 static int
3156 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
3157 const struct pandecode_mapped_memory *mem,
3158 mali_ptr payload, int job_no, unsigned gpu_id)
3159 {
3160 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
3161
3162 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
3163 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
3164
3165 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
3166 pandecode_indent++;
3167
3168 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
3169
3170 pandecode_log(".tiler = ");
3171 pandecode_tiler_only_bfr(&t->tiler, job_no);
3172
3173 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3174
3175 pandecode_indent--;
3176 pandecode_log("};\n");
3177
3178 return sizeof(*t);
3179 }
3180
3181 static int
3182 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3183 const struct pandecode_mapped_memory *mem,
3184 mali_ptr payload, int job_no, unsigned gpu_id)
3185 {
3186 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3187 bool is_graphics = (h->job_type == JOB_TYPE_VERTEX) || (h->job_type == JOB_TYPE_TILER);
3188
3189 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3190
3191 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3192 pandecode_indent++;
3193
3194 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3195 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3196
3197 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3198 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3199
3200 pandecode_indent--;
3201 pandecode_log("};\n");
3202
3203 return sizeof(*v);
3204 }
3205
3206 static int
3207 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3208 mali_ptr payload, int job_no,
3209 bool is_bifrost, unsigned gpu_id)
3210 {
3211 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3212
3213 bool is_mfbd = s->framebuffer & MALI_MFBD;
3214
3215 if (!is_mfbd && is_bifrost)
3216 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3217
3218 struct pandecode_fbd info;
3219
3220 if (is_mfbd)
3221 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3222 else
3223 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3224
3225 /* Compute the tag for the tagged pointer. This contains the type of
3226 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3227 * additional structures follow the MFBD header (an extra payload or
3228 * not, as well as a count of render targets) */
3229
3230 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3231
3232 if (is_mfbd) {
3233 if (info.has_extra)
3234 expected_tag |= MALI_MFBD_TAG_EXTRA;
3235
3236 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3237 }
3238
3239 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3240 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3241 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3242 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3243 }
3244
3245 /* Extract tile coordinates */
3246
3247 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3248 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3249
3250 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3251 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3252
3253 /* For the max, we also want the floored (rather than ceiled) version for checking */
3254
3255 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3256 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3257
3258 /* Validate the coordinates are well-ordered */
3259
3260 if (min_x == max_x)
3261 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3262 else if (min_x > max_x)
3263 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3264
3265 if (min_y == max_y)
3266 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3267 else if (min_y > max_y)
3268 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3269
3270 /* Validate the coordinates fit inside the framebuffer. We use floor,
3271 * rather than ceil, for the max coordinates, since the tile
3272 * coordinates for something like an 800x600 framebuffer will actually
3273 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3274
3275 if ((min_x > info.width) || (max_x_f > info.width))
3276 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3277
3278 if ((min_y > info.height) || (max_y_f > info.height))
3279 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3280
3281 /* After validation, we print */
3282
3283 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3284
3285 /* The FBD is a tagged pointer */
3286
3287 unsigned tag = (s->framebuffer & ~FBD_MASK);
3288
3289 if (tag != expected_tag)
3290 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3291
3292 return sizeof(*s);
3293 }
3294
3295 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3296 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3297 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3298 * might wish to combine this with the bifrost parameter) because some details
3299 * are model-specific even within a particular architecture. Minimal traces
3300 * *only* examine the job descriptors, skipping printing entirely if there is
3301 * no faults, and only descends into the payload if there are faults. This is
3302 * useful for looking for faults without the overhead of invasive traces. */
3303
3304 void
3305 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3306 {
3307 pandecode_dump_file_open();
3308
3309 struct mali_job_descriptor_header *h;
3310 unsigned job_descriptor_number = 0;
3311
3312 do {
3313 struct pandecode_mapped_memory *mem =
3314 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3315
3316 void *payload;
3317
3318 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3319
3320 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3321 * high 32-bits of the 64-bit pointer are reused to store
3322 * something else.
3323 */
3324 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3325 h->job_type != JOB_TYPE_FRAGMENT ? 4 : 0;
3326 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3327
3328 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3329
3330 int job_no = job_descriptor_number++;
3331
3332 /* If the job is good to go, skip it in minimal mode */
3333 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3334 continue;
3335
3336 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3337 pandecode_indent++;
3338
3339 pandecode_prop("job_type = %s", pandecode_job_type(h->job_type));
3340
3341 if (h->job_descriptor_size)
3342 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3343
3344 if (h->exception_status && h->exception_status != 0x1)
3345 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3346 h->exception_status,
3347 (h->exception_status >> 16) & 0xFFFF,
3348 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3349 h->exception_status & 0xFF);
3350
3351 if (h->first_incomplete_task)
3352 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3353
3354 if (h->fault_pointer)
3355 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3356
3357 if (h->job_barrier)
3358 pandecode_prop("job_barrier = %d", h->job_barrier);
3359
3360 pandecode_prop("job_index = %d", h->job_index);
3361
3362 if (h->unknown_flags)
3363 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3364
3365 if (h->job_dependency_index_1)
3366 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3367
3368 if (h->job_dependency_index_2)
3369 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3370
3371 pandecode_indent--;
3372 pandecode_log("};\n");
3373
3374 switch (h->job_type) {
3375 case JOB_TYPE_WRITE_VALUE: {
3376 struct mali_payload_write_value *s = payload;
3377 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3378 pandecode_indent++;
3379 MEMORY_PROP(s, address);
3380
3381 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3382 pandecode_msg("XXX: unknown value descriptor\n");
3383 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3384 }
3385
3386 if (s->reserved) {
3387 pandecode_msg("XXX: set value tripped\n");
3388 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3389 }
3390
3391 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3392 pandecode_indent--;
3393 pandecode_log("};\n");
3394
3395 break;
3396 }
3397
3398 case JOB_TYPE_TILER:
3399 case JOB_TYPE_VERTEX:
3400 case JOB_TYPE_COMPUTE:
3401 if (bifrost) {
3402 if (h->job_type == JOB_TYPE_TILER)
3403 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3404 else
3405 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3406 } else
3407 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3408
3409 break;
3410
3411 case JOB_TYPE_FRAGMENT:
3412 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3413 break;
3414
3415 default:
3416 break;
3417 }
3418 } while ((jc_gpu_va = h->next_job));
3419
3420 pandecode_map_read_write();
3421 }