panfrost: XMLify draw_mode
[mesa.git] / src / panfrost / lib / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <midgard_pack.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
43
44 #define MEMORY_PROP(obj, p) {\
45 if (obj->p) { \
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
48 free(a); \
49 } \
50 }
51
52 #define MEMORY_PROP_DIR(obj, p) {\
53 if (obj.p) { \
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
56 free(a); \
57 } \
58 }
59
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, 0); \
65 }
66
67 #define DUMP_ADDR(title, T, addr, indent) {\
68 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
69 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
70 DUMP_CL(title, T, cl, indent); \
71 }
72
73 FILE *pandecode_dump_stream;
74
75 /* Semantic logging type.
76 *
77 * Raw: for raw messages to be printed as is.
78 * Message: for helpful information to be commented out in replays.
79 * Property: for properties of a struct
80 *
81 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
82 */
83
84 enum pandecode_log_type {
85 PANDECODE_RAW,
86 PANDECODE_MESSAGE,
87 PANDECODE_PROPERTY
88 };
89
90 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
91 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
92 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
93
94 unsigned pandecode_indent = 0;
95
96 static void
97 pandecode_make_indent(void)
98 {
99 for (unsigned i = 0; i < pandecode_indent; ++i)
100 fprintf(pandecode_dump_stream, " ");
101 }
102
103 static void
104 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
105 {
106 va_list ap;
107
108 pandecode_make_indent();
109
110 if (type == PANDECODE_MESSAGE)
111 fprintf(pandecode_dump_stream, "// ");
112 else if (type == PANDECODE_PROPERTY)
113 fprintf(pandecode_dump_stream, ".");
114
115 va_start(ap, format);
116 vfprintf(pandecode_dump_stream, format, ap);
117 va_end(ap);
118
119 if (type == PANDECODE_PROPERTY)
120 fprintf(pandecode_dump_stream, ",\n");
121 }
122
123 static void
124 pandecode_log_cont(const char *format, ...)
125 {
126 va_list ap;
127
128 va_start(ap, format);
129 vfprintf(pandecode_dump_stream, format, ap);
130 va_end(ap);
131 }
132
133 /* To check for memory safety issues, validates that the given pointer in GPU
134 * memory is valid, containing at least sz bytes. The goal is to eliminate
135 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
136 * overruns) by statically validating pointers.
137 */
138
139 static void
140 pandecode_validate_buffer(mali_ptr addr, size_t sz)
141 {
142 if (!addr) {
143 pandecode_msg("XXX: null pointer deref");
144 return;
145 }
146
147 /* Find a BO */
148
149 struct pandecode_mapped_memory *bo =
150 pandecode_find_mapped_gpu_mem_containing(addr);
151
152 if (!bo) {
153 pandecode_msg("XXX: invalid memory dereference\n");
154 return;
155 }
156
157 /* Bounds check */
158
159 unsigned offset = addr - bo->gpu_va;
160 unsigned total = offset + sz;
161
162 if (total > bo->length) {
163 pandecode_msg("XXX: buffer overrun. "
164 "Chunk of size %zu at offset %d in buffer of size %zu. "
165 "Overrun by %zu bytes. \n",
166 sz, offset, bo->length, total - bo->length);
167 return;
168 }
169 }
170
171 struct pandecode_flag_info {
172 u64 flag;
173 const char *name;
174 };
175
176 static void
177 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
178 u64 flags)
179 {
180 bool decodable_flags_found = false;
181
182 for (int i = 0; flag_info[i].name; i++) {
183 if ((flags & flag_info[i].flag) != flag_info[i].flag)
184 continue;
185
186 if (!decodable_flags_found) {
187 decodable_flags_found = true;
188 } else {
189 pandecode_log_cont(" | ");
190 }
191
192 pandecode_log_cont("%s", flag_info[i].name);
193
194 flags &= ~flag_info[i].flag;
195 }
196
197 if (decodable_flags_found) {
198 if (flags)
199 pandecode_log_cont(" | 0x%" PRIx64, flags);
200 } else {
201 pandecode_log_cont("0x%" PRIx64, flags);
202 }
203 }
204
205 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
206 static const struct pandecode_flag_info gl_enable_flag_info[] = {
207 FLAG_INFO(OCCLUSION_QUERY),
208 FLAG_INFO(OCCLUSION_PRECISE),
209 FLAG_INFO(FRONT_CCW_TOP),
210 FLAG_INFO(CULL_FACE_FRONT),
211 FLAG_INFO(CULL_FACE_BACK),
212 {}
213 };
214 #undef FLAG_INFO
215
216 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
217 static const struct pandecode_flag_info clear_flag_info[] = {
218 FLAG_INFO(FAST),
219 FLAG_INFO(SLOW),
220 FLAG_INFO(SLOW_STENCIL),
221 {}
222 };
223 #undef FLAG_INFO
224
225 #define FLAG_INFO(flag) { MALI_MASK_##flag, "MALI_MASK_" #flag }
226 static const struct pandecode_flag_info mask_flag_info[] = {
227 FLAG_INFO(R),
228 FLAG_INFO(G),
229 FLAG_INFO(B),
230 FLAG_INFO(A),
231 {}
232 };
233 #undef FLAG_INFO
234
235 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
236 static const struct pandecode_flag_info u3_flag_info[] = {
237 FLAG_INFO(HAS_MSAA),
238 FLAG_INFO(PER_SAMPLE),
239 FLAG_INFO(CAN_DISCARD),
240 FLAG_INFO(HAS_BLEND_SHADER),
241 FLAG_INFO(DEPTH_WRITEMASK),
242 FLAG_INFO(DEPTH_CLIP_NEAR),
243 FLAG_INFO(DEPTH_CLIP_FAR),
244 {}
245 };
246
247 static const struct pandecode_flag_info u4_flag_info[] = {
248 FLAG_INFO(NO_MSAA),
249 FLAG_INFO(NO_DITHER),
250 FLAG_INFO(DEPTH_RANGE_A),
251 FLAG_INFO(DEPTH_RANGE_B),
252 FLAG_INFO(STENCIL_TEST),
253 FLAG_INFO(ALPHA_TO_COVERAGE),
254 {}
255 };
256 #undef FLAG_INFO
257
258 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
259 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
260 FLAG_INFO(SRGB),
261 {}
262 };
263 #undef FLAG_INFO
264
265 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
266 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
267 FLAG_INFO(YTR),
268 {}
269 };
270 #undef FLAG_INFO
271
272 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
273 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
274 FLAG_INFO(PRESENT),
275 {}
276 };
277 #undef FLAG_INFO
278
279 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
280 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
281 FLAG_INFO(ZS),
282 {}
283 };
284 #undef FLAG_INFO
285
286 #define FLAG_INFO(flag) { MALI_##flag, "MALI_" #flag }
287 static const struct pandecode_flag_info shader_midgard1_flag_lo_info [] = {
288 FLAG_INFO(WRITES_Z),
289 FLAG_INFO(EARLY_Z),
290 FLAG_INFO(READS_TILEBUFFER),
291 FLAG_INFO(WRITES_GLOBAL),
292 FLAG_INFO(READS_ZS),
293 {}
294 };
295
296 static const struct pandecode_flag_info shader_midgard1_flag_hi_info [] = {
297 FLAG_INFO(WRITES_S),
298 FLAG_INFO(SUPPRESS_INF_NAN),
299 {}
300 };
301 #undef FLAG_INFO
302
303 #define FLAG_INFO(flag) { MALI_BIFROST_##flag, "MALI_BIFROST_" #flag }
304 static const struct pandecode_flag_info shader_bifrost_info [] = {
305 FLAG_INFO(FULL_THREAD),
306 FLAG_INFO(EARLY_Z),
307 FLAG_INFO(FIRST_ATEST),
308 {}
309 };
310
311 #undef FLAG_INFO
312
313 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
314 static const struct pandecode_flag_info mfbd_flag_info [] = {
315 FLAG_INFO(DEPTH_WRITE),
316 FLAG_INFO(EXTRA),
317 {}
318 };
319 #undef FLAG_INFO
320
321 #define FLAG_INFO(flag) { MALI_SAMP_##flag, "MALI_SAMP_" #flag }
322 static const struct pandecode_flag_info sampler_flag_info [] = {
323 FLAG_INFO(MAG_NEAREST),
324 FLAG_INFO(MIN_NEAREST),
325 FLAG_INFO(MIP_LINEAR_1),
326 FLAG_INFO(MIP_LINEAR_2),
327 FLAG_INFO(NORM_COORDS),
328 {}
329 };
330 #undef FLAG_INFO
331
332 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
333 static const struct pandecode_flag_info sfbd_unk1_info [] = {
334 FLAG_INFO(MSAA_8),
335 FLAG_INFO(MSAA_A),
336 {}
337 };
338 #undef FLAG_INFO
339
340 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
341 static const struct pandecode_flag_info sfbd_unk2_info [] = {
342 FLAG_INFO(MSAA_B),
343 FLAG_INFO(SRGB),
344 {}
345 };
346 #undef FLAG_INFO
347
348 #define DEFINE_CASE(name) case MALI_## name: return "MALI_" #name
349 static char *pandecode_format(enum mali_format format)
350 {
351 static char unk_format_str[10];
352
353 switch (format) {
354 DEFINE_CASE(ETC2_RGB8);
355 DEFINE_CASE(ETC2_R11_UNORM);
356 DEFINE_CASE(ETC2_RGBA8);
357 DEFINE_CASE(ETC2_RG11_UNORM);
358 DEFINE_CASE(ETC2_R11_SNORM);
359 DEFINE_CASE(ETC2_RG11_SNORM);
360 DEFINE_CASE(ETC2_RGB8A1);
361 DEFINE_CASE(NXR);
362 DEFINE_CASE(BC1_UNORM);
363 DEFINE_CASE(BC2_UNORM);
364 DEFINE_CASE(BC3_UNORM);
365 DEFINE_CASE(BC4_UNORM);
366 DEFINE_CASE(BC4_SNORM);
367 DEFINE_CASE(BC5_UNORM);
368 DEFINE_CASE(BC5_SNORM);
369 DEFINE_CASE(BC6H_UF16);
370 DEFINE_CASE(BC6H_SF16);
371 DEFINE_CASE(BC7_UNORM);
372 DEFINE_CASE(ASTC_3D_LDR);
373 DEFINE_CASE(ASTC_3D_HDR);
374 DEFINE_CASE(ASTC_2D_LDR);
375 DEFINE_CASE(ASTC_2D_HDR);
376 DEFINE_CASE(RGB565);
377 DEFINE_CASE(RGB5_X1_UNORM);
378 DEFINE_CASE(RGB5_A1_UNORM);
379 DEFINE_CASE(RGB10_A2_UNORM);
380 DEFINE_CASE(RGB10_A2_SNORM);
381 DEFINE_CASE(RGB10_A2UI);
382 DEFINE_CASE(RGB10_A2I);
383 DEFINE_CASE(RGB332_UNORM);
384 DEFINE_CASE(RGB233_UNORM);
385 DEFINE_CASE(Z24X8_UNORM);
386 DEFINE_CASE(R32_FIXED);
387 DEFINE_CASE(RG32_FIXED);
388 DEFINE_CASE(RGB32_FIXED);
389 DEFINE_CASE(RGBA32_FIXED);
390 DEFINE_CASE(R11F_G11F_B10F);
391 DEFINE_CASE(R9F_G9F_B9F_E5F);
392 DEFINE_CASE(VARYING_POS);
393 DEFINE_CASE(VARYING_DISCARD);
394
395 DEFINE_CASE(R8_SNORM);
396 DEFINE_CASE(R16_SNORM);
397 DEFINE_CASE(R32_SNORM);
398 DEFINE_CASE(RG8_SNORM);
399 DEFINE_CASE(RG16_SNORM);
400 DEFINE_CASE(RG32_SNORM);
401 DEFINE_CASE(RGB8_SNORM);
402 DEFINE_CASE(RGB16_SNORM);
403 DEFINE_CASE(RGB32_SNORM);
404 DEFINE_CASE(RGBA8_SNORM);
405 DEFINE_CASE(RGBA16_SNORM);
406 DEFINE_CASE(RGBA32_SNORM);
407
408 DEFINE_CASE(R8UI);
409 DEFINE_CASE(R16UI);
410 DEFINE_CASE(R32UI);
411 DEFINE_CASE(RG8UI);
412 DEFINE_CASE(RG16UI);
413 DEFINE_CASE(RG32UI);
414 DEFINE_CASE(RGB8UI);
415 DEFINE_CASE(RGB16UI);
416 DEFINE_CASE(RGB32UI);
417 DEFINE_CASE(RGBA8UI);
418 DEFINE_CASE(RGBA16UI);
419 DEFINE_CASE(RGBA32UI);
420
421 DEFINE_CASE(R8_UNORM);
422 DEFINE_CASE(R16_UNORM);
423 DEFINE_CASE(R32_UNORM);
424 DEFINE_CASE(R32F);
425 DEFINE_CASE(RG8_UNORM);
426 DEFINE_CASE(RG16_UNORM);
427 DEFINE_CASE(RG32_UNORM);
428 DEFINE_CASE(RG32F);
429 DEFINE_CASE(RGB8_UNORM);
430 DEFINE_CASE(RGB16_UNORM);
431 DEFINE_CASE(RGB32_UNORM);
432 DEFINE_CASE(RGB32F);
433 DEFINE_CASE(RGBA4_UNORM);
434 DEFINE_CASE(RGBA8_UNORM);
435 DEFINE_CASE(RGBA16_UNORM);
436 DEFINE_CASE(RGBA32_UNORM);
437 DEFINE_CASE(RGBA32F);
438
439 DEFINE_CASE(R8I);
440 DEFINE_CASE(R16I);
441 DEFINE_CASE(R32I);
442 DEFINE_CASE(RG8I);
443 DEFINE_CASE(R16F);
444 DEFINE_CASE(RG16I);
445 DEFINE_CASE(RG32I);
446 DEFINE_CASE(RG16F);
447 DEFINE_CASE(RGB8I);
448 DEFINE_CASE(RGB16I);
449 DEFINE_CASE(RGB32I);
450 DEFINE_CASE(RGB16F);
451 DEFINE_CASE(RGBA8I);
452 DEFINE_CASE(RGBA16I);
453 DEFINE_CASE(RGBA32I);
454 DEFINE_CASE(RGBA16F);
455
456 DEFINE_CASE(RGBA4);
457 DEFINE_CASE(RGBA8_2);
458 DEFINE_CASE(RGB10_A2_2);
459 default:
460 snprintf(unk_format_str, sizeof(unk_format_str), "MALI_0x%02x", format);
461 return unk_format_str;
462 }
463 }
464
465 #undef DEFINE_CASE
466
467 #define DEFINE_CASE(name) case MALI_FUNC_ ## name: return "MALI_FUNC_" #name
468 static char *
469 pandecode_func(enum mali_func mode)
470 {
471 switch (mode) {
472 DEFINE_CASE(NEVER);
473 DEFINE_CASE(LESS);
474 DEFINE_CASE(EQUAL);
475 DEFINE_CASE(LEQUAL);
476 DEFINE_CASE(GREATER);
477 DEFINE_CASE(NOTEQUAL);
478 DEFINE_CASE(GEQUAL);
479 DEFINE_CASE(ALWAYS);
480
481 default:
482 pandecode_msg("XXX: invalid func %X\n", mode);
483 return "";
484 }
485 }
486 #undef DEFINE_CASE
487
488 #define DEFINE_CASE(name) case MALI_MSAA_ ## name: return "MALI_MSAA_" #name
489 static char *
490 pandecode_msaa_mode(enum mali_msaa_mode mode)
491 {
492 switch (mode) {
493 DEFINE_CASE(SINGLE);
494 DEFINE_CASE(AVERAGE);
495 DEFINE_CASE(MULTIPLE);
496 DEFINE_CASE(LAYERED);
497 default:
498 unreachable("Impossible");
499 return "";
500 }
501 }
502 #undef DEFINE_CASE
503
504 #define DEFINE_CASE(name) case MALI_STENCIL_ ## name: return "MALI_STENCIL_" #name
505 static char *
506 pandecode_stencil_op(enum mali_stencil_op op)
507 {
508 switch (op) {
509 DEFINE_CASE(KEEP);
510 DEFINE_CASE(REPLACE);
511 DEFINE_CASE(ZERO);
512 DEFINE_CASE(INVERT);
513 DEFINE_CASE(INCR_WRAP);
514 DEFINE_CASE(DECR_WRAP);
515 DEFINE_CASE(INCR);
516 DEFINE_CASE(DECR);
517
518 default:
519 pandecode_msg("XXX: invalid stencil op %X\n", op);
520 return "";
521 }
522 }
523
524 #undef DEFINE_CASE
525
526 static char *pandecode_attr_mode_short(enum mali_attr_mode mode)
527 {
528 switch(mode) {
529 /* TODO: Combine to just "instanced" once this can be done
530 * unambiguously in all known cases */
531 case MALI_ATTR_POT_DIVIDE:
532 return "instanced_pot";
533 case MALI_ATTR_MODULO:
534 return "instanced_mod";
535 case MALI_ATTR_NPOT_DIVIDE:
536 return "instanced_npot";
537 case MALI_ATTR_IMAGE:
538 return "image";
539 default:
540 pandecode_msg("XXX: invalid attribute mode %X\n", mode);
541 return "";
542 }
543 }
544
545 static const char *
546 pandecode_special_record(uint64_t v, bool* attribute)
547 {
548 switch(v) {
549 case MALI_ATTR_VERTEXID:
550 *attribute = true;
551 return "gl_VertexID";
552 case MALI_ATTR_INSTANCEID:
553 *attribute = true;
554 return "gl_InstanceID";
555 case MALI_VARYING_FRAG_COORD:
556 return "gl_FragCoord";
557 case MALI_VARYING_FRONT_FACING:
558 return "gl_FrontFacing";
559 case MALI_VARYING_POINT_COORD:
560 return "gl_PointCoord";
561 default:
562 pandecode_msg("XXX: invalid special record %" PRIx64 "\n", v);
563 return "";
564 }
565 }
566
567 #define DEFINE_CASE(name) case MALI_WRAP_## name: return "MALI_WRAP_" #name
568 static char *
569 pandecode_wrap_mode(enum mali_wrap_mode op)
570 {
571 switch (op) {
572 DEFINE_CASE(REPEAT);
573 DEFINE_CASE(CLAMP_TO_EDGE);
574 DEFINE_CASE(CLAMP);
575 DEFINE_CASE(CLAMP_TO_BORDER);
576 DEFINE_CASE(MIRRORED_REPEAT);
577 DEFINE_CASE(MIRRORED_CLAMP_TO_EDGE);
578 DEFINE_CASE(MIRRORED_CLAMP);
579 DEFINE_CASE(MIRRORED_CLAMP_TO_BORDER);
580
581 default:
582 pandecode_msg("XXX: invalid wrap mode %X\n", op);
583 return "";
584 }
585 }
586 #undef DEFINE_CASE
587
588 #define DEFINE_CASE(name) case MALI_BLOCK_## name: return "MALI_BLOCK_" #name
589 static char *
590 pandecode_block_format(enum mali_block_format fmt)
591 {
592 switch (fmt) {
593 DEFINE_CASE(TILED);
594 DEFINE_CASE(UNKNOWN);
595 DEFINE_CASE(LINEAR);
596 DEFINE_CASE(AFBC);
597
598 default:
599 unreachable("Invalid case");
600 }
601 }
602 #undef DEFINE_CASE
603
604 #define DEFINE_CASE(name) case MALI_EXCEPTION_ACCESS_## name: return ""#name
605 static char *
606 pandecode_exception_access(unsigned access)
607 {
608 switch (access) {
609 DEFINE_CASE(NONE);
610 DEFINE_CASE(EXECUTE);
611 DEFINE_CASE(READ);
612 DEFINE_CASE(WRITE);
613
614 default:
615 unreachable("Invalid case");
616 }
617 }
618 #undef DEFINE_CASE
619
620 /* Midgard's tiler descriptor is embedded within the
621 * larger FBD */
622
623 static void
624 pandecode_midgard_tiler_descriptor(
625 const struct midgard_tiler_descriptor *t,
626 unsigned width,
627 unsigned height,
628 bool is_fragment,
629 bool has_hierarchy)
630 {
631 pandecode_log(".tiler = {\n");
632 pandecode_indent++;
633
634 if (t->hierarchy_mask == MALI_TILER_DISABLED)
635 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
636 else
637 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
638
639 /* We know this name from the kernel, but we never see it nonzero */
640
641 if (t->flags)
642 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
643
644 MEMORY_PROP(t, polygon_list);
645
646 /* The body is offset from the base of the polygon list */
647 //assert(t->polygon_list_body > t->polygon_list);
648 unsigned body_offset = t->polygon_list_body - t->polygon_list;
649
650 /* It needs to fit inside the reported size */
651 //assert(t->polygon_list_size >= body_offset);
652
653 /* Now that we've sanity checked, we'll try to calculate the sizes
654 * ourselves for comparison */
655
656 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
657 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
658
659 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
660 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
661 ref_header, ref_size);
662 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
663 pandecode_msg("body offset %d\n", body_offset);
664 }
665
666 /* The tiler heap has a start and end specified -- it should be
667 * identical to what we have in the BO. The exception is if tiling is
668 * disabled. */
669
670 MEMORY_PROP(t, heap_start);
671 assert(t->heap_end >= t->heap_start);
672
673 struct pandecode_mapped_memory *heap =
674 pandecode_find_mapped_gpu_mem_containing(t->heap_start);
675
676 unsigned heap_size = t->heap_end - t->heap_start;
677
678 /* Tiling is enabled with a special flag */
679 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
680 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
681
682 bool tiling_enabled = hierarchy_mask;
683
684 if (tiling_enabled) {
685 /* When tiling is enabled, the heap should be a tight fit */
686 unsigned heap_offset = t->heap_start - heap->gpu_va;
687 if ((heap_offset + heap_size) != heap->length) {
688 pandecode_msg("XXX: heap size %u (expected %zu)\n",
689 heap_size, heap->length - heap_offset);
690 }
691
692 /* We should also have no other flags */
693 if (tiler_flags)
694 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
695 } else {
696 /* When tiling is disabled, we should have that flag and no others */
697
698 if (tiler_flags != MALI_TILER_DISABLED) {
699 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
700 tiler_flags);
701 }
702
703 /* We should also have an empty heap */
704 if (heap_size) {
705 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
706 heap_size);
707 }
708
709 /* Disabled tiling is used only for clear-only jobs, which are
710 * purely FRAGMENT, so we should never see this for
711 * non-FRAGMENT descriptors. */
712
713 if (!is_fragment)
714 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
715 }
716
717 /* We've never seen weights used in practice, but we know from the
718 * kernel these fields is there */
719
720 bool nonzero_weights = false;
721
722 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
723 nonzero_weights |= t->weights[w] != 0x0;
724 }
725
726 if (nonzero_weights) {
727 pandecode_log(".weights = { ");
728
729 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
730 pandecode_log_cont("%d, ", t->weights[w]);
731 }
732
733 pandecode_log("},");
734 }
735
736 pandecode_indent--;
737 pandecode_log("}\n");
738 }
739
740 /* TODO: The Bifrost tiler is not understood at all yet */
741
742 static void
743 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
744 {
745 pandecode_log(".tiler = {\n");
746 pandecode_indent++;
747
748 MEMORY_PROP(fb, tiler_meta);
749
750 for (int i = 0; i < 16; i++) {
751 if (fb->zeros[i] != 0) {
752 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
753 i, fb->zeros[i]);
754 }
755 }
756
757 pandecode_log("},\n");
758
759 pandecode_indent--;
760 pandecode_log("}\n");
761
762 }
763
764 /* Information about the framebuffer passed back for
765 * additional analysis */
766
767 struct pandecode_fbd {
768 unsigned width;
769 unsigned height;
770 unsigned rt_count;
771 bool has_extra;
772 };
773
774 static void
775 pandecode_sfbd_format(struct mali_sfbd_format format)
776 {
777 pandecode_log(".format = {\n");
778 pandecode_indent++;
779
780 pandecode_log(".unk1 = ");
781 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
782 pandecode_log_cont(",\n");
783
784 /* TODO: Map formats so we can check swizzles and print nicely */
785 pandecode_log("swizzle");
786 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
787 pandecode_log_cont(",\n");
788
789 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
790 (format.nr_channels + 1));
791
792 pandecode_log(".unk2 = ");
793 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
794 pandecode_log_cont(",\n");
795
796 pandecode_prop("block = %s", pandecode_block_format(format.block));
797
798 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
799
800 pandecode_indent--;
801 pandecode_log("},\n");
802 }
803
804 static void
805 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
806 {
807 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
808
809 if (desc->unk0)
810 pandecode_prop("unk0 = 0x%x", desc->unk0);
811
812 if (desc->shared_workgroup_count != 0x1F) {
813 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
814 if (!is_compute)
815 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
816 }
817
818 if (desc->shared_unk1 || desc->shared_shift) {
819 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
820 pandecode_prop("shared_shift = %X", desc->shared_shift);
821
822 if (!is_compute)
823 pandecode_msg("XXX: shared memory configured in noncompute shader");
824 }
825
826 if (desc->shared_zero) {
827 pandecode_msg("XXX: shared memory zero tripped\n");
828 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
829 }
830
831 if (desc->shared_memory && !is_compute)
832 pandecode_msg("XXX: shared memory used in noncompute shader\n");
833
834 MEMORY_PROP(desc, scratchpad);
835 MEMORY_PROP(desc, shared_memory);
836 MEMORY_PROP(desc, unknown1);
837
838 if (desc->scratchpad) {
839 struct pandecode_mapped_memory *smem =
840 pandecode_find_mapped_gpu_mem_containing(desc->scratchpad);
841
842 pandecode_msg("scratchpad size %u\n", smem->length);
843 }
844
845 }
846
847 static struct pandecode_fbd
848 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
849 {
850 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
851 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
852
853 struct pandecode_fbd info = {
854 .has_extra = false,
855 .rt_count = 1
856 };
857
858 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
859 pandecode_indent++;
860
861 pandecode_log(".shared_memory = {\n");
862 pandecode_indent++;
863 pandecode_shared_memory(&s->shared_memory, false);
864 pandecode_indent--;
865 pandecode_log("},\n");
866
867 pandecode_sfbd_format(s->format);
868
869 info.width = s->width + 1;
870 info.height = s->height + 1;
871
872 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
873 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
874
875 MEMORY_PROP(s, checksum);
876
877 if (s->checksum_stride)
878 pandecode_prop("checksum_stride = %d", s->checksum_stride);
879
880 MEMORY_PROP(s, framebuffer);
881 pandecode_prop("stride = %d", s->stride);
882
883 /* Earlier in the actual commandstream -- right before width -- but we
884 * delay to flow nicer */
885
886 pandecode_log(".clear_flags = ");
887 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
888 pandecode_log_cont(",\n");
889
890 if (s->depth_buffer) {
891 MEMORY_PROP(s, depth_buffer);
892 pandecode_prop("depth_stride = %d", s->depth_stride);
893 }
894
895 if (s->stencil_buffer) {
896 MEMORY_PROP(s, stencil_buffer);
897 pandecode_prop("stencil_stride = %d", s->stencil_stride);
898 }
899
900 if (s->depth_stride_zero ||
901 s->stencil_stride_zero ||
902 s->zero7 || s->zero8) {
903 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
904 pandecode_prop("depth_stride_zero = 0x%x",
905 s->depth_stride_zero);
906 pandecode_prop("stencil_stride_zero = 0x%x",
907 s->stencil_stride_zero);
908 pandecode_prop("zero7 = 0x%" PRIx32,
909 s->zero7);
910 pandecode_prop("zero8 = 0x%" PRIx32,
911 s->zero8);
912 }
913
914 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
915 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
916 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
917 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
918 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
919 }
920
921 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
922 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
923 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
924 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
925 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
926 }
927
928 if (s->clear_stencil) {
929 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
930 }
931
932 const struct midgard_tiler_descriptor t = s->tiler;
933
934 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
935 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
936
937 pandecode_indent--;
938 pandecode_log("};\n");
939
940 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
941 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
942 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
943
944 pandecode_log_cont(".zero3 = {");
945
946 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
947 pandecode_log_cont("%X, ", s->zero3[i]);
948
949 pandecode_log_cont("},\n");
950
951 pandecode_log_cont(".zero6 = {");
952
953 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
954 pandecode_log_cont("%X, ", s->zero6[i]);
955
956 pandecode_log_cont("},\n");
957
958 return info;
959 }
960
961 static void
962 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
963 {
964 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
965 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
966
967 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
968 pandecode_indent++;
969 pandecode_shared_memory(s, true);
970 pandecode_indent--;
971 pandecode_log("},\n");
972 }
973
974 /* Extracts the number of components associated with a Mali format */
975
976 static unsigned
977 pandecode_format_component_count(enum mali_format fmt)
978 {
979 /* Mask out the format class */
980 unsigned top = fmt & 0b11100000;
981
982 switch (top) {
983 case MALI_FORMAT_SNORM:
984 case MALI_FORMAT_UINT:
985 case MALI_FORMAT_UNORM:
986 case MALI_FORMAT_SINT:
987 return ((fmt >> 3) & 3) + 1;
988 default:
989 /* TODO: Validate */
990 return 4;
991 }
992 }
993
994 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
995
996 static unsigned
997 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
998 {
999 unsigned mask = 0;
1000 assert(MALI_CHANNEL_RED == 0);
1001
1002 for (unsigned c = 0; c < 4; ++c) {
1003 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
1004
1005 if (chan <= MALI_CHANNEL_ALPHA)
1006 mask |= (1 << chan);
1007 }
1008
1009 return mask;
1010 }
1011
1012 /* Validates that a (format, swizzle) pair is valid, in the sense that the
1013 * swizzle doesn't access any components that are undefined in the format.
1014 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
1015 * omitted */
1016
1017 static bool
1018 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
1019 {
1020 unsigned nr_comp = pandecode_format_component_count(fmt);
1021 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
1022 unsigned valid_mask = (1 << nr_comp) - 1;
1023 unsigned invalid_mask = ~valid_mask;
1024
1025 if (access_mask & invalid_mask) {
1026 pandecode_msg("XXX: invalid components accessed\n");
1027 return false;
1028 }
1029
1030 /* Check for the default non-swizzling swizzle so we can suppress
1031 * useless printing for the defaults */
1032
1033 unsigned default_swizzles[4] = {
1034 MALI_CHANNEL_RED | (MALI_CHANNEL_ZERO << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1035 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_ZERO << 6) | (MALI_CHANNEL_ONE << 9),
1036 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ONE << 9),
1037 MALI_CHANNEL_RED | (MALI_CHANNEL_GREEN << 3) | (MALI_CHANNEL_BLUE << 6) | (MALI_CHANNEL_ALPHA << 9)
1038 };
1039
1040 return (swizzle == default_swizzles[nr_comp - 1]);
1041 }
1042
1043 /* Maps MALI_RGBA32F to rgba32f, etc */
1044
1045 static void
1046 pandecode_format_short(enum mali_format fmt, bool srgb)
1047 {
1048 /* We want a type-like format, so cut off the initial MALI_ */
1049 char *format = pandecode_format(fmt);
1050 format += strlen("MALI_");
1051
1052 unsigned len = strlen(format);
1053 char *lower_format = calloc(1, len + 1);
1054
1055 for (unsigned i = 0; i < len; ++i)
1056 lower_format[i] = tolower(format[i]);
1057
1058 /* Sanity check sRGB flag is applied to RGB, per the name */
1059 if (srgb && lower_format[0] != 'r')
1060 pandecode_msg("XXX: sRGB applied to non-colour format\n");
1061
1062 /* Just prefix with an s, so you get formats like srgba8_unorm */
1063 if (srgb)
1064 pandecode_log_cont("s");
1065
1066 pandecode_log_cont("%s", lower_format);
1067 free(lower_format);
1068 }
1069
1070 static void
1071 pandecode_swizzle(unsigned swizzle, enum mali_format format)
1072 {
1073 /* First, do some validation */
1074 bool trivial_swizzle = pandecode_validate_format_swizzle(
1075 format, swizzle);
1076
1077 if (trivial_swizzle)
1078 return;
1079
1080 /* Next, print the swizzle */
1081 pandecode_log_cont(".");
1082
1083 static const char components[] = "rgba01";
1084
1085 for (unsigned c = 0; c < 4; ++c) {
1086 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
1087
1088 if (chan >= MALI_CHANNEL_RESERVED_0) {
1089 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
1090 continue;
1091 }
1092 pandecode_log_cont("%c", components[chan]);
1093 }
1094 }
1095
1096 static void
1097 pandecode_rt_format(struct mali_rt_format format)
1098 {
1099 pandecode_log(".format = {\n");
1100 pandecode_indent++;
1101
1102 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
1103 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
1104 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
1105 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
1106
1107 pandecode_prop("block = %s", pandecode_block_format(format.block));
1108
1109 /* TODO: Map formats so we can check swizzles and print nicely */
1110 pandecode_log("swizzle");
1111 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
1112 pandecode_log_cont(",\n");
1113
1114 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
1115 (format.nr_channels + 1));
1116
1117 pandecode_log(".flags = ");
1118 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
1119 pandecode_log_cont(",\n");
1120
1121 pandecode_prop("msaa = %s", pandecode_msaa_mode(format.msaa));
1122
1123 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
1124 * which is a faster hardware-based alternative to the wallpaper method
1125 * to preserve framebuffer contents across frames. In practice, MFBD
1126 * preload is buggy on Midgard, and so this is a chicken bit. If this
1127 * bit isn't set, most likely something broke unrelated to preload */
1128
1129 if (!format.no_preload) {
1130 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
1131 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
1132 }
1133
1134 if (format.zero)
1135 pandecode_prop("zero = 0x%" PRIx32, format.zero);
1136
1137 pandecode_indent--;
1138 pandecode_log("},\n");
1139 }
1140
1141 static void
1142 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
1143 {
1144 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
1145 pandecode_indent++;
1146
1147 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
1148 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
1149 struct pandecode_mapped_memory *mem =
1150 pandecode_find_mapped_gpu_mem_containing(rt_va);
1151 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
1152
1153 pandecode_log("{\n");
1154 pandecode_indent++;
1155
1156 pandecode_rt_format(rt->format);
1157
1158 if (rt->format.block == MALI_BLOCK_AFBC) {
1159 pandecode_log(".afbc = {\n");
1160 pandecode_indent++;
1161
1162 char *a = pointer_as_memory_reference(rt->afbc.metadata);
1163 pandecode_prop("metadata = %s", a);
1164 free(a);
1165
1166 pandecode_prop("stride = %d", rt->afbc.stride);
1167
1168 pandecode_log(".flags = ");
1169 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
1170 pandecode_log_cont(",\n");
1171
1172 pandecode_indent--;
1173 pandecode_log("},\n");
1174 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
1175 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
1176 rt->afbc.metadata,
1177 rt->afbc.stride,
1178 rt->afbc.flags);
1179 }
1180
1181 MEMORY_PROP(rt, framebuffer);
1182 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
1183
1184 if (rt->layer_stride)
1185 pandecode_prop("layer_stride = %d", rt->layer_stride);
1186
1187 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
1188 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
1189 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
1190 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
1191 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
1192 }
1193
1194 if (rt->zero1 || rt->zero2) {
1195 pandecode_msg("XXX: render target zeros tripped\n");
1196 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
1197 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
1198 }
1199
1200 pandecode_indent--;
1201 pandecode_log("},\n");
1202 }
1203
1204 pandecode_indent--;
1205 pandecode_log("};\n");
1206 }
1207
1208 static struct pandecode_fbd
1209 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
1210 {
1211 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1212 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
1213
1214 struct pandecode_fbd info;
1215
1216 if (is_bifrost && fb->msaa.sample_locations) {
1217 /* The blob stores all possible sample locations in a single buffer
1218 * allocated on startup, and just switches the pointer when switching
1219 * MSAA state. For now, we just put the data into the cmdstream, but we
1220 * should do something like what the blob does with a real driver.
1221 *
1222 * There seem to be 32 slots for sample locations, followed by another
1223 * 16. The second 16 is just the center location followed by 15 zeros
1224 * in all the cases I've identified (maybe shader vs. depth/color
1225 * samples?).
1226 */
1227
1228 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
1229
1230 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
1231
1232 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
1233 pandecode_indent++;
1234
1235 for (int i = 0; i < 32 + 16; i++) {
1236 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
1237 }
1238
1239 pandecode_indent--;
1240 pandecode_log("};\n");
1241 }
1242
1243 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
1244 pandecode_indent++;
1245
1246 if (is_bifrost) {
1247 pandecode_log(".msaa = {\n");
1248 pandecode_indent++;
1249
1250 if (fb->msaa.sample_locations)
1251 pandecode_prop("sample_locations = sample_locations_%d", job_no);
1252 else
1253 pandecode_msg("XXX: sample_locations missing\n");
1254
1255 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
1256 pandecode_msg("XXX: multisampling zero tripped\n");
1257 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
1258 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
1259 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
1260 }
1261
1262 pandecode_indent--;
1263 pandecode_log("},\n");
1264 } else {
1265 pandecode_log(".shared_memory = {\n");
1266 pandecode_indent++;
1267 pandecode_shared_memory(&fb->shared_memory, is_compute);
1268 pandecode_indent--;
1269 pandecode_log("},\n");
1270 }
1271
1272 info.width = fb->width1 + 1;
1273 info.height = fb->height1 + 1;
1274 info.rt_count = fb->rt_count_1 + 1;
1275
1276 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
1277 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
1278 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
1279 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
1280
1281 pandecode_prop("unk1 = 0x%x", fb->unk1);
1282 pandecode_prop("unk2 = 0x%x", fb->unk2);
1283 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
1284 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
1285
1286 pandecode_log(".mfbd_flags = ");
1287 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
1288 pandecode_log_cont(",\n");
1289
1290 if (fb->clear_stencil)
1291 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
1292
1293 if (fb->clear_depth)
1294 pandecode_prop("clear_depth = %f", fb->clear_depth);
1295
1296 if (!is_compute)
1297 if (is_bifrost)
1298 pandecode_bifrost_tiler_descriptor(fb);
1299 else {
1300 const struct midgard_tiler_descriptor t = fb->tiler;
1301 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
1302 }
1303 else
1304 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
1305
1306 if (fb->zero3 || fb->zero4) {
1307 pandecode_msg("XXX: framebuffer zeros tripped\n");
1308 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
1309 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
1310 }
1311
1312 pandecode_indent--;
1313 pandecode_log("};\n");
1314
1315 gpu_va += sizeof(struct mali_framebuffer);
1316
1317 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
1318
1319 if (info.has_extra) {
1320 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1321 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
1322
1323 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
1324 pandecode_indent++;
1325
1326 MEMORY_PROP(fbx, checksum);
1327
1328 if (fbx->checksum_stride)
1329 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
1330
1331 pandecode_log(".flags_hi = ");
1332 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
1333 pandecode_log_cont(",\n");
1334
1335 pandecode_log(".flags_lo = ");
1336 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
1337 pandecode_log_cont(",\n");
1338
1339 pandecode_prop("zs_block = %s", pandecode_block_format(fbx->zs_block));
1340 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
1341
1342 if (fbx->zs_block == MALI_BLOCK_AFBC) {
1343 pandecode_log(".ds_afbc = {\n");
1344 pandecode_indent++;
1345
1346 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
1347 pandecode_prop("depth_stencil_afbc_stride = %d",
1348 fbx->ds_afbc.depth_stencil_afbc_stride);
1349 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
1350
1351 pandecode_log(".flags = ");
1352 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
1353 pandecode_log_cont(",\n");
1354
1355 if (fbx->ds_afbc.padding) {
1356 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
1357 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
1358 }
1359
1360 pandecode_indent--;
1361 pandecode_log("},\n");
1362 } else {
1363 pandecode_log(".ds_linear = {\n");
1364 pandecode_indent++;
1365
1366 if (fbx->ds_linear.depth) {
1367 MEMORY_PROP_DIR(fbx->ds_linear, depth);
1368 pandecode_prop("depth_stride = %d",
1369 fbx->ds_linear.depth_stride);
1370 pandecode_prop("depth_layer_stride = %d",
1371 fbx->ds_linear.depth_layer_stride);
1372 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
1373 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
1374 }
1375
1376 if (fbx->ds_linear.stencil) {
1377 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
1378 pandecode_prop("stencil_stride = %d",
1379 fbx->ds_linear.stencil_stride);
1380 pandecode_prop("stencil_layer_stride = %d",
1381 fbx->ds_linear.stencil_layer_stride);
1382 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
1383 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
1384 }
1385
1386 if (fbx->ds_linear.depth_stride_zero ||
1387 fbx->ds_linear.stencil_stride_zero) {
1388 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
1389 pandecode_prop("depth_stride_zero = 0x%x",
1390 fbx->ds_linear.depth_stride_zero);
1391 pandecode_prop("stencil_stride_zero = 0x%x",
1392 fbx->ds_linear.stencil_stride_zero);
1393 }
1394
1395 pandecode_indent--;
1396 pandecode_log("},\n");
1397 }
1398
1399 if (fbx->clear_color_1 | fbx->clear_color_2) {
1400 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1401 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1402 }
1403
1404 if (fbx->zero3) {
1405 pandecode_msg("XXX: fb_extra zeros tripped\n");
1406 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1407 }
1408
1409 pandecode_indent--;
1410 pandecode_log("};\n");
1411
1412 gpu_va += sizeof(struct mali_framebuffer_extra);
1413 }
1414
1415 if (is_fragment)
1416 pandecode_render_target(gpu_va, job_no, fb);
1417
1418 return info;
1419 }
1420
1421 /* Just add a comment decoding the shift/odd fields forming the padded vertices
1422 * count */
1423
1424 static void
1425 pandecode_padded_vertices(unsigned shift, unsigned k)
1426 {
1427 unsigned odd = 2*k + 1;
1428 unsigned pot = 1 << shift;
1429 pandecode_msg("padded_num_vertices = %d\n", odd * pot);
1430 }
1431
1432 /* Given a magic divisor, recover what we were trying to divide by.
1433 *
1434 * Let m represent the magic divisor. By definition, m is an element on Z, whre
1435 * 0 <= m < 2^N, for N bits in m.
1436 *
1437 * Let q represent the number we would like to divide by.
1438 *
1439 * By definition of a magic divisor for N-bit unsigned integers (a number you
1440 * multiply by to magically get division), m is a number such that:
1441 *
1442 * (m * x) & (2^N - 1) = floor(x/q).
1443 * for all x on Z where 0 <= x < 2^N
1444 *
1445 * Ignore the case where any of the above values equals zero; it is irrelevant
1446 * for our purposes (instanced arrays).
1447 *
1448 * Choose x = q. Then:
1449 *
1450 * (m * x) & (2^N - 1) = floor(x/q).
1451 * (m * q) & (2^N - 1) = floor(q/q).
1452 *
1453 * floor(q/q) = floor(1) = 1, therefore:
1454 *
1455 * (m * q) & (2^N - 1) = 1
1456 *
1457 * Recall the identity that the bitwise AND of one less than a power-of-two
1458 * equals the modulo with that power of two, i.e. for all x:
1459 *
1460 * x & (2^N - 1) = x % N
1461 *
1462 * Therefore:
1463 *
1464 * mq % (2^N) = 1
1465 *
1466 * By definition, a modular multiplicative inverse of a number m is the number
1467 * q such that with respect to a modulos M:
1468 *
1469 * mq % M = 1
1470 *
1471 * Therefore, q is the modular multiplicative inverse of m with modulus 2^N.
1472 *
1473 */
1474
1475 static void
1476 pandecode_magic_divisor(uint32_t magic, unsigned shift, unsigned orig_divisor, unsigned extra)
1477 {
1478 #if 0
1479 /* Compute the modular inverse of `magic` with respect to 2^(32 -
1480 * shift) the most lame way possible... just repeatedly add.
1481 * Asymptoptically slow but nobody cares in practice, unless you have
1482 * massive numbers of vertices or high divisors. */
1483
1484 unsigned inverse = 0;
1485
1486 /* Magic implicitly has the highest bit set */
1487 magic |= (1 << 31);
1488
1489 /* Depending on rounding direction */
1490 if (extra)
1491 magic++;
1492
1493 for (;;) {
1494 uint32_t product = magic * inverse;
1495
1496 if (shift) {
1497 product >>= shift;
1498 }
1499
1500 if (product == 1)
1501 break;
1502
1503 ++inverse;
1504 }
1505
1506 pandecode_msg("dividing by %d (maybe off by two)\n", inverse);
1507
1508 /* Recall we're supposed to divide by (gl_level_divisor *
1509 * padded_num_vertices) */
1510
1511 unsigned padded_num_vertices = inverse / orig_divisor;
1512
1513 pandecode_msg("padded_num_vertices = %d\n", padded_num_vertices);
1514 #endif
1515 }
1516
1517 static void
1518 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1519 mali_ptr addr, int job_no, char *suffix,
1520 int count, bool varying, enum mali_job_type job_type)
1521 {
1522 char *prefix = varying ? "varying" : "attribute";
1523 assert(addr);
1524
1525 if (!count) {
1526 pandecode_msg("warn: No %s records\n", prefix);
1527 return;
1528 }
1529
1530 union mali_attr *attr = pandecode_fetch_gpu_mem(mem, addr, sizeof(union mali_attr) * count);
1531
1532 for (int i = 0; i < count; ++i) {
1533 /* First, check for special records */
1534 if (attr[i].elements < MALI_RECORD_SPECIAL) {
1535 if (attr[i].size)
1536 pandecode_msg("XXX: tripped size=%d\n", attr[i].size);
1537
1538 if (attr[i].stride) {
1539 /* gl_InstanceID passes a magic divisor in the
1540 * stride field to divide by the padded vertex
1541 * count. No other records should do so, so
1542 * stride should otherwise be zero. Note that
1543 * stride in the usual attribute sense doesn't
1544 * apply to special records. */
1545
1546 bool has_divisor = attr[i].elements == MALI_ATTR_INSTANCEID;
1547
1548 pandecode_log_cont("/* %smagic divisor = %X */ ",
1549 has_divisor ? "" : "XXX: ", attr[i].stride);
1550 }
1551
1552 if (attr[i].shift || attr[i].extra_flags) {
1553 /* Attributes use these fields for
1554 * instancing/padding/etc type issues, but
1555 * varyings don't */
1556
1557 pandecode_log_cont("/* %sshift=%d, extra=%d */ ",
1558 varying ? "XXX: " : "",
1559 attr[i].shift, attr[i].extra_flags);
1560 }
1561
1562 /* Print the special record name */
1563 bool attribute = false;
1564 pandecode_log("%s_%d = %s;\n", prefix, i, pandecode_special_record(attr[i].elements, &attribute));
1565
1566 /* Sanity check */
1567 if (attribute == varying)
1568 pandecode_msg("XXX: mismatched special record\n");
1569
1570 continue;
1571 }
1572
1573 enum mali_attr_mode mode = attr[i].elements & 7;
1574
1575 if (mode == MALI_ATTR_UNUSED)
1576 pandecode_msg("XXX: unused attribute record\n");
1577
1578 /* For non-linear records, we need to print the type of record */
1579 if (mode != MALI_ATTR_LINEAR)
1580 pandecode_log_cont("%s ", pandecode_attr_mode_short(mode));
1581
1582 /* Print the name to link with attr_meta */
1583 pandecode_log_cont("%s_%d", prefix, i);
1584
1585 /* Print the stride and size */
1586 pandecode_log_cont("<%u>[%u]", attr[i].stride, attr[i].size);
1587
1588 /* TODO: Sanity check the quotient itself. It must be equal to
1589 * (or be greater than, if the driver added padding) the padded
1590 * vertex count. */
1591
1592 /* Finally, print the pointer */
1593 mali_ptr raw_elements = attr[i].elements & ~7;
1594 char *a = pointer_as_memory_reference(raw_elements);
1595 pandecode_log_cont(" = (%s);\n", a);
1596 free(a);
1597
1598 /* Check the pointer */
1599 pandecode_validate_buffer(raw_elements, attr[i].size);
1600
1601 /* shift/extra_flags exist only for instanced */
1602 if (attr[i].shift | attr[i].extra_flags) {
1603 /* These are set to random values by the blob for
1604 * varyings, most likely a symptom of uninitialized
1605 * memory where the hardware masked the bug. As such we
1606 * put this at a warning, not an error. */
1607
1608 if (mode == MALI_ATTR_LINEAR)
1609 pandecode_msg("warn: instancing fields set for linear\n");
1610
1611 pandecode_prop("shift = %d", attr[i].shift);
1612 pandecode_prop("extra_flags = %d", attr[i].extra_flags);
1613 }
1614
1615 /* Decode further where possible */
1616
1617 if (mode == MALI_ATTR_MODULO) {
1618 pandecode_padded_vertices(
1619 attr[i].shift,
1620 attr[i].extra_flags);
1621 }
1622
1623 if (mode == MALI_ATTR_NPOT_DIVIDE) {
1624 i++;
1625 pandecode_log("{\n");
1626 pandecode_indent++;
1627 pandecode_prop("unk = 0x%x", attr[i].unk);
1628 pandecode_prop("magic_divisor = 0x%08x", attr[i].magic_divisor);
1629 if (attr[i].zero != 0)
1630 pandecode_prop("XXX: zero tripped (0x%x)\n", attr[i].zero);
1631 pandecode_prop("divisor = %d", attr[i].divisor);
1632 pandecode_magic_divisor(attr[i].magic_divisor, attr[i - 1].shift, attr[i].divisor, attr[i - 1].extra_flags);
1633 pandecode_indent--;
1634 pandecode_log("}, \n");
1635 }
1636
1637 }
1638
1639 pandecode_log("\n");
1640 }
1641
1642 static mali_ptr
1643 pandecode_shader_address(const char *name, mali_ptr ptr)
1644 {
1645 /* TODO: Decode flags */
1646 mali_ptr shader_ptr = ptr & ~15;
1647
1648 char *a = pointer_as_memory_reference(shader_ptr);
1649 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1650 free(a);
1651
1652 return shader_ptr;
1653 }
1654
1655 static void
1656 pandecode_stencil(const char *name, const struct mali_stencil_test *stencil)
1657 {
1658 unsigned any_nonzero =
1659 stencil->ref | stencil->mask | stencil->func |
1660 stencil->sfail | stencil->dpfail | stencil->dppass;
1661
1662 if (any_nonzero == 0)
1663 return;
1664
1665 const char *func = pandecode_func(stencil->func);
1666 const char *sfail = pandecode_stencil_op(stencil->sfail);
1667 const char *dpfail = pandecode_stencil_op(stencil->dpfail);
1668 const char *dppass = pandecode_stencil_op(stencil->dppass);
1669
1670 if (stencil->zero)
1671 pandecode_msg("XXX: stencil zero tripped: %X\n", stencil->zero);
1672
1673 pandecode_log(".stencil_%s = {\n", name);
1674 pandecode_indent++;
1675 pandecode_prop("ref = %d", stencil->ref);
1676 pandecode_prop("mask = 0x%02X", stencil->mask);
1677 pandecode_prop("func = %s", func);
1678 pandecode_prop("sfail = %s", sfail);
1679 pandecode_prop("dpfail = %s", dpfail);
1680 pandecode_prop("dppass = %s", dppass);
1681 pandecode_indent--;
1682 pandecode_log("},\n");
1683 }
1684
1685 static void
1686 pandecode_blend_equation(const struct mali_blend_equation *blend)
1687 {
1688 if (blend->zero1)
1689 pandecode_msg("XXX: blend zero tripped: %X\n", blend->zero1);
1690
1691 pandecode_log(".equation = {\n");
1692 pandecode_indent++;
1693
1694 pandecode_prop("rgb_mode = 0x%X", blend->rgb_mode);
1695 pandecode_prop("alpha_mode = 0x%X", blend->alpha_mode);
1696
1697 pandecode_log(".color_mask = ");
1698 pandecode_log_decoded_flags(mask_flag_info, blend->color_mask);
1699 pandecode_log_cont(",\n");
1700
1701 pandecode_indent--;
1702 pandecode_log("},\n");
1703 }
1704
1705 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1706
1707 static unsigned
1708 decode_bifrost_constant(u16 constant)
1709 {
1710 float lo = (float) (constant & 0xFF);
1711 float hi = (float) (constant >> 8);
1712
1713 return (hi / 255.0) + (lo / 65535.0);
1714 }
1715
1716 static mali_ptr
1717 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1718 {
1719 struct bifrost_blend_rt *b =
1720 ((struct bifrost_blend_rt *) descs) + rt_no;
1721
1722 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1723 pandecode_indent++;
1724
1725 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1726 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1727 b->constant, decode_bifrost_constant(b->constant));
1728
1729 /* TODO figure out blend shader enable bit */
1730 pandecode_blend_equation(&b->equation);
1731
1732 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1733 pandecode_prop("index = 0x%" PRIx16, b->index);
1734
1735 pandecode_log(".format = ");
1736 pandecode_format_short(b->format, false);
1737 pandecode_swizzle(b->swizzle, b->format);
1738 pandecode_log_cont(",\n");
1739
1740 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1741 pandecode_prop("format = 0x%" PRIx32, b->format);
1742
1743 if (b->zero1) {
1744 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1745 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1746 }
1747
1748 pandecode_log(".shader_type = ");
1749 switch(b->shader_type) {
1750 case BIFROST_BLEND_F16:
1751 pandecode_log_cont("BIFROST_BLEND_F16");
1752 break;
1753 case BIFROST_BLEND_F32:
1754 pandecode_log_cont("BIFROST_BLEND_F32");
1755 break;
1756 case BIFROST_BLEND_I32:
1757 pandecode_log_cont("BIFROST_BLEND_I32");
1758 break;
1759 case BIFROST_BLEND_U32:
1760 pandecode_log_cont("BIFROST_BLEND_U32");
1761 break;
1762 case BIFROST_BLEND_I16:
1763 pandecode_log_cont("BIFROST_BLEND_I16");
1764 break;
1765 case BIFROST_BLEND_U16:
1766 pandecode_log_cont("BIFROST_BLEND_U16");
1767 break;
1768 }
1769 pandecode_log_cont(",\n");
1770
1771 if (b->zero2) {
1772 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1773 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1774 }
1775
1776 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1777
1778 pandecode_indent--;
1779 pandecode_log("},\n");
1780
1781 return 0;
1782 }
1783
1784 static mali_ptr
1785 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1786 {
1787 /* constant/equation is in a union */
1788 if (!blend->shader)
1789 return 0;
1790
1791 pandecode_log(".blend = {\n");
1792 pandecode_indent++;
1793
1794 if (is_shader) {
1795 pandecode_shader_address("shader", blend->shader);
1796 } else {
1797 pandecode_blend_equation(&blend->equation);
1798 pandecode_prop("constant = %f", blend->constant);
1799 }
1800
1801 pandecode_indent--;
1802 pandecode_log("},\n");
1803
1804 /* Return blend shader to disassemble if present */
1805 return is_shader ? (blend->shader & ~0xF) : 0;
1806 }
1807
1808 static mali_ptr
1809 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1810 {
1811 struct midgard_blend_rt *b =
1812 ((struct midgard_blend_rt *) descs) + rt_no;
1813
1814 /* Flags determine presence of blend shader */
1815 bool is_shader = (b->flags & 0xF) >= 0x2;
1816
1817 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1818 pandecode_indent++;
1819
1820 pandecode_prop("flags = 0x%" PRIx64, b->flags);
1821
1822 union midgard_blend blend = b->blend;
1823 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1824
1825 pandecode_indent--;
1826 pandecode_log("};\n");
1827
1828 return shader;
1829 }
1830
1831 /* Attributes and varyings have descriptor records, which contain information
1832 * about their format and ordering with the attribute/varying buffers. We'll
1833 * want to validate that the combinations specified are self-consistent.
1834 */
1835
1836 static int
1837 pandecode_attribute_meta(int job_no, int count, const struct mali_vertex_tiler_postfix *v, bool varying, char *suffix)
1838 {
1839 char base[128];
1840 char *prefix = varying ? "varying" : "attribute";
1841 unsigned max_index = 0;
1842 snprintf(base, sizeof(base), "%s_meta", prefix);
1843
1844 struct mali_attr_meta *attr_meta;
1845 mali_ptr p = varying ? v->varying_meta : v->attribute_meta;
1846
1847 struct pandecode_mapped_memory *attr_mem = pandecode_find_mapped_gpu_mem_containing(p);
1848
1849 for (int i = 0; i < count; ++i, p += sizeof(struct mali_attr_meta)) {
1850 attr_meta = pandecode_fetch_gpu_mem(attr_mem, p,
1851 sizeof(*attr_mem));
1852
1853 /* If the record is discard, it should be zero for everything else */
1854
1855 if (attr_meta->format == MALI_VARYING_DISCARD) {
1856 uint64_t zero =
1857 attr_meta->index |
1858 attr_meta->unknown1 |
1859 attr_meta->unknown3 |
1860 attr_meta->src_offset;
1861
1862 if (zero)
1863 pandecode_msg("XXX: expected empty record for varying discard\n");
1864
1865 /* We want to look for a literal 0000 swizzle -- this
1866 * is not encoded with all zeroes, however */
1867
1868 enum mali_channel z = MALI_CHANNEL_ZERO;
1869 unsigned zero_swizzle = z | (z << 3) | (z << 6) | (z << 9);
1870 bool good_swizzle = attr_meta->swizzle == zero_swizzle;
1871
1872 if (!good_swizzle)
1873 pandecode_msg("XXX: expected zero swizzle for discard\n");
1874
1875 if (!varying)
1876 pandecode_msg("XXX: cannot discard attribute\n");
1877
1878 /* If we're all good, omit the record */
1879 if (!zero && varying && good_swizzle) {
1880 pandecode_log("/* discarded varying */\n");
1881 continue;
1882 }
1883 }
1884
1885 if (attr_meta->index > max_index)
1886 max_index = attr_meta->index;
1887
1888 if (attr_meta->unknown1 != 0x2) {
1889 pandecode_msg("XXX: expected unknown1 = 0x2\n");
1890 pandecode_prop("unknown1 = 0x%" PRIx64, (u64) attr_meta->unknown1);
1891 }
1892
1893 if (attr_meta->unknown3) {
1894 pandecode_msg("XXX: unexpected unknown3 set\n");
1895 pandecode_prop("unknown3 = 0x%" PRIx64, (u64) attr_meta->unknown3);
1896 }
1897
1898 pandecode_format_short(attr_meta->format, false);
1899 pandecode_log_cont(" %s_%u", prefix, attr_meta->index);
1900
1901 if (attr_meta->src_offset)
1902 pandecode_log_cont("[%u]", attr_meta->src_offset);
1903
1904 pandecode_swizzle(attr_meta->swizzle, attr_meta->format);
1905
1906 pandecode_log_cont(";\n");
1907 }
1908
1909 pandecode_log("\n");
1910
1911 return count ? (max_index + 1) : 0;
1912 }
1913
1914 /* return bits [lo, hi) of word */
1915 static u32
1916 bits(u32 word, u32 lo, u32 hi)
1917 {
1918 if (hi - lo >= 32)
1919 return word; // avoid undefined behavior with the shift
1920
1921 return (word >> lo) & ((1 << (hi - lo)) - 1);
1922 }
1923
1924 static void
1925 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1926 {
1927 pandecode_log(".prefix = {\n");
1928 pandecode_indent++;
1929
1930 /* Decode invocation_count. See the comment before the definition of
1931 * invocation_count for an explanation.
1932 */
1933
1934 unsigned size_y_shift = bits(p->invocation_shifts, 0, 5);
1935 unsigned size_z_shift = bits(p->invocation_shifts, 5, 10);
1936 unsigned workgroups_x_shift = bits(p->invocation_shifts, 10, 16);
1937 unsigned workgroups_y_shift = bits(p->invocation_shifts, 16, 22);
1938 unsigned workgroups_z_shift = bits(p->invocation_shifts, 22, 28);
1939 unsigned workgroups_x_shift_2 = bits(p->invocation_shifts, 28, 32);
1940
1941 unsigned size_x = bits(p->invocation_count, 0, size_y_shift) + 1;
1942 unsigned size_y = bits(p->invocation_count, size_y_shift, size_z_shift) + 1;
1943 unsigned size_z = bits(p->invocation_count, size_z_shift, workgroups_x_shift) + 1;
1944
1945 unsigned groups_x = bits(p->invocation_count, workgroups_x_shift, workgroups_y_shift) + 1;
1946 unsigned groups_y = bits(p->invocation_count, workgroups_y_shift, workgroups_z_shift) + 1;
1947 unsigned groups_z = bits(p->invocation_count, workgroups_z_shift, 32) + 1;
1948
1949 /* Even though we have this decoded, we want to ensure that the
1950 * representation is "unique" so we don't lose anything by printing only
1951 * the final result. More specifically, we need to check that we were
1952 * passed something in canonical form, since the definition per the
1953 * hardware is inherently not unique. How? Well, take the resulting
1954 * decode and pack it ourselves! If it is bit exact with what we
1955 * decoded, we're good to go. */
1956
1957 struct mali_vertex_tiler_prefix ref;
1958 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1959
1960 bool canonical =
1961 (p->invocation_count == ref.invocation_count) &&
1962 (p->invocation_shifts == ref.invocation_shifts);
1963
1964 if (!canonical) {
1965 pandecode_msg("XXX: non-canonical workgroups packing\n");
1966 pandecode_msg("expected: %X, %X",
1967 ref.invocation_count,
1968 ref.invocation_shifts);
1969
1970 pandecode_prop("invocation_count = 0x%" PRIx32, p->invocation_count);
1971 pandecode_prop("size_y_shift = %d", size_y_shift);
1972 pandecode_prop("size_z_shift = %d", size_z_shift);
1973 pandecode_prop("workgroups_x_shift = %d", workgroups_x_shift);
1974 pandecode_prop("workgroups_y_shift = %d", workgroups_y_shift);
1975 pandecode_prop("workgroups_z_shift = %d", workgroups_z_shift);
1976 pandecode_prop("workgroups_x_shift_2 = %d", workgroups_x_shift_2);
1977 }
1978
1979 /* Regardless, print the decode */
1980 pandecode_msg("size (%d, %d, %d), count (%d, %d, %d)\n",
1981 size_x, size_y, size_z,
1982 groups_x, groups_y, groups_z);
1983
1984 /* TODO: Decode */
1985 if (p->unknown_draw)
1986 pandecode_prop("unknown_draw = 0x%" PRIx32, p->unknown_draw);
1987
1988 pandecode_prop("workgroups_x_shift_3 = 0x%" PRIx32, p->workgroups_x_shift_3);
1989
1990 if (p->draw_mode != MALI_DRAW_MODE_NONE)
1991 pandecode_prop("draw_mode = %s", mali_draw_mode_as_str(p->draw_mode));
1992
1993 /* Index count only exists for tiler jobs anyway */
1994
1995 if (p->index_count)
1996 pandecode_prop("index_count = MALI_POSITIVE(%" PRId32 ")", p->index_count + 1);
1997
1998
1999 unsigned index_raw_size = (p->unknown_draw & MALI_DRAW_INDEXED_SIZE);
2000 index_raw_size >>= MALI_DRAW_INDEXED_SHIFT;
2001
2002 /* Validate an index buffer is present if we need one. TODO: verify
2003 * relationship between invocation_count and index_count */
2004
2005 if (p->indices) {
2006 unsigned count = p->index_count;
2007
2008 /* Grab the size */
2009 unsigned size = (index_raw_size == 0x3) ? 4 : index_raw_size;
2010
2011 /* Ensure we got a size, and if so, validate the index buffer
2012 * is large enough to hold a full set of indices of the given
2013 * size */
2014
2015 if (!index_raw_size)
2016 pandecode_msg("XXX: index size missing\n");
2017 else
2018 pandecode_validate_buffer(p->indices, count * size);
2019 } else if (index_raw_size)
2020 pandecode_msg("XXX: unexpected index size %u\n", index_raw_size);
2021
2022 if (p->offset_bias_correction)
2023 pandecode_prop("offset_bias_correction = %d", p->offset_bias_correction);
2024
2025 /* TODO: Figure out what this is. It's not zero */
2026 pandecode_prop("zero1 = 0x%" PRIx32, p->zero1);
2027
2028 pandecode_indent--;
2029 pandecode_log("},\n");
2030 }
2031
2032 static void
2033 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
2034 {
2035 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
2036 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
2037
2038 for (int i = 0; i < ubufs_count; i++) {
2039 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
2040 mali_ptr addr = (ubufs[i] >> 10) << 2;
2041
2042 pandecode_validate_buffer(addr, size);
2043
2044 char *ptr = pointer_as_memory_reference(addr);
2045 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
2046 free(ptr);
2047 }
2048
2049 pandecode_log("\n");
2050 }
2051
2052 static void
2053 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
2054 {
2055 pandecode_validate_buffer(uniforms, uniform_count * 16);
2056
2057 char *ptr = pointer_as_memory_reference(uniforms);
2058 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
2059 free(ptr);
2060 }
2061
2062 static const char *
2063 shader_type_for_job(unsigned type)
2064 {
2065 switch (type) {
2066 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
2067 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
2068 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
2069 default:
2070 return "UNKNOWN";
2071 }
2072 }
2073
2074 static unsigned shader_id = 0;
2075
2076 static struct midgard_disasm_stats
2077 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
2078 bool is_bifrost, unsigned gpu_id)
2079 {
2080 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
2081 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
2082
2083 /* Compute maximum possible size */
2084 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
2085
2086 /* Print some boilerplate to clearly denote the assembly (which doesn't
2087 * obey indentation rules), and actually do the disassembly! */
2088
2089 pandecode_log_cont("\n\n");
2090
2091 struct midgard_disasm_stats stats;
2092
2093 if (is_bifrost) {
2094 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
2095
2096 /* TODO: Extend stats to Bifrost */
2097 stats.texture_count = -128;
2098 stats.sampler_count = -128;
2099 stats.attribute_count = -128;
2100 stats.varying_count = -128;
2101 stats.uniform_count = -128;
2102 stats.uniform_buffer_count = -128;
2103 stats.work_count = -128;
2104
2105 stats.instruction_count = 0;
2106 stats.bundle_count = 0;
2107 stats.quadword_count = 0;
2108 stats.helper_invocations = false;
2109 } else {
2110 stats = disassemble_midgard(pandecode_dump_stream,
2111 code, sz, gpu_id,
2112 type == MALI_JOB_TYPE_TILER ?
2113 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
2114 }
2115
2116 /* Print shader-db stats. Skip COMPUTE jobs since they are used for
2117 * driver-internal purposes with the blob and interfere */
2118
2119 bool should_shaderdb = type != MALI_JOB_TYPE_COMPUTE;
2120
2121 if (should_shaderdb) {
2122 unsigned nr_threads =
2123 (stats.work_count <= 4) ? 4 :
2124 (stats.work_count <= 8) ? 2 :
2125 1;
2126
2127 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
2128 "%u inst, %u bundles, %u quadwords, "
2129 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
2130 shader_id++,
2131 shader_type_for_job(type),
2132 stats.instruction_count, stats.bundle_count, stats.quadword_count,
2133 stats.work_count, nr_threads);
2134 }
2135
2136
2137 return stats;
2138 }
2139
2140 static void
2141 pandecode_texture_payload(mali_ptr payload,
2142 enum mali_texture_type type,
2143 enum mali_texture_layout layout,
2144 bool manual_stride,
2145 uint8_t levels,
2146 uint16_t depth,
2147 uint16_t array_size,
2148 struct pandecode_mapped_memory *tmem)
2149 {
2150 pandecode_log(".payload = {\n");
2151 pandecode_indent++;
2152
2153 /* A bunch of bitmap pointers follow.
2154 * We work out the correct number,
2155 * based on the mipmap/cubemap
2156 * properties, but dump extra
2157 * possibilities to futureproof */
2158
2159 int bitmap_count = levels + 1;
2160
2161 /* Miptree for each face */
2162 if (type == MALI_TEX_CUBE)
2163 bitmap_count *= 6;
2164
2165 /* Array of layers */
2166 bitmap_count *= (depth + 1);
2167
2168 /* Array of textures */
2169 bitmap_count *= (array_size + 1);
2170
2171 /* Stride for each element */
2172 if (manual_stride)
2173 bitmap_count *= 2;
2174
2175 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
2176 payload, sizeof(mali_ptr) * bitmap_count);
2177 for (int i = 0; i < bitmap_count; ++i) {
2178 /* How we dump depends if this is a stride or a pointer */
2179
2180 if (manual_stride && (i & 1)) {
2181 /* signed 32-bit snuck in as a 64-bit pointer */
2182 uint64_t stride_set = pointers_and_strides[i];
2183 uint32_t clamped_stride = stride_set;
2184 int32_t stride = clamped_stride;
2185 assert(stride_set == clamped_stride);
2186 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
2187 } else {
2188 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
2189 pandecode_log("%s, \n", a);
2190 free(a);
2191 }
2192 }
2193
2194 pandecode_indent--;
2195 pandecode_log("},\n");
2196 }
2197
2198 static void
2199 pandecode_texture(mali_ptr u,
2200 struct pandecode_mapped_memory *tmem,
2201 unsigned job_no, unsigned tex)
2202 {
2203 struct mali_texture_descriptor *PANDECODE_PTR_VAR(t, tmem, u);
2204
2205 pandecode_log("struct mali_texture_descriptor texture_descriptor_%"PRIx64"_%d_%d = {\n", u, job_no, tex);
2206 pandecode_indent++;
2207
2208 pandecode_prop("width = %" PRId32, t->width);
2209 pandecode_prop("height = %" PRId32, t->height);
2210 pandecode_prop("depth = %" PRId32, t->depth);
2211 pandecode_prop("array_size = %" PRId32, t->array_size);
2212
2213 pandecode_log("\n");
2214 pandecode_prop("f.swizzle = 0x%" PRIx32, t->format.swizzle);
2215 pandecode_prop("f.format = 0x%" PRIx32, t->format.format);
2216 pandecode_prop("f.srgb = 0x%" PRIx32, t->format.srgb);
2217 pandecode_prop("f.unknown1 = 0x%" PRIx32, t->format.unknown1);
2218 pandecode_prop("f.type = %" PRId32, t->format.type);
2219 pandecode_prop("f.layout = %" PRId32, t->format.layout);
2220 pandecode_prop("f.unknown2 = 0x%" PRIx32, t->format.unknown2);
2221 pandecode_prop("f.manual_stride = %" PRId32, t->format.manual_stride);
2222 pandecode_prop("f.zero = 0x%" PRIx32, t->format.zero);
2223 pandecode_log("\n");
2224
2225 pandecode_prop("unknown3 = 0x%" PRIx32, t->unknown3);
2226 pandecode_prop("unknown3A = 0x%" PRIx32, t->unknown3A);
2227 pandecode_prop("levels = %" PRId32, t->levels);
2228 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2229 pandecode_prop("swizzle_zero = 0x%" PRIx32, t->swizzle_zero);
2230
2231 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2232 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2233 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2234 pandecode_log("\n");
2235
2236 struct mali_texture_format f = t->format;
2237
2238 /* See the definiton of enum mali_texture_type */
2239
2240 bool is_cube = f.type == MALI_TEX_CUBE;
2241 unsigned dimension = is_cube ? 2 : f.type;
2242
2243 pandecode_make_indent();
2244
2245 /* Print the layout. Default is linear; a modifier can denote AFBC or
2246 * u-interleaved/tiled modes */
2247
2248 if (f.layout == MALI_TEXTURE_AFBC)
2249 pandecode_log_cont("afbc");
2250 else if (f.layout == MALI_TEXTURE_TILED)
2251 pandecode_log_cont("tiled");
2252 else if (f.layout == MALI_TEXTURE_LINEAR)
2253 pandecode_log_cont("linear");
2254 else
2255 pandecode_msg("XXX: invalid texture layout 0x%X\n", f.layout);
2256
2257 pandecode_swizzle(t->swizzle, f.format);
2258 pandecode_log_cont(" ");
2259
2260 /* Distinguish cube/2D with modifier */
2261
2262 if (is_cube)
2263 pandecode_log_cont("cube ");
2264
2265 pandecode_format_short(f.format, f.srgb);
2266 pandecode_swizzle(f.swizzle, f.format);
2267
2268 /* All four width/height/depth/array_size dimensions are present
2269 * regardless of the type of texture, but it is an error to have
2270 * non-zero dimensions for unused dimensions. Verify this. array_size
2271 * can always be set, as can width. Depth used for MSAA. */
2272
2273 if (t->height && dimension < 2)
2274 pandecode_msg("XXX: nonzero height for <2D texture\n");
2275
2276 /* Print only the dimensions that are actually there */
2277
2278 pandecode_log_cont(": %d", t->width + 1);
2279
2280 if (t->height || t->depth)
2281 pandecode_log_cont("x%u", t->height + 1);
2282
2283 if (t->depth)
2284 pandecode_log_cont("x%u", t->depth + 1);
2285
2286 if (t->array_size)
2287 pandecode_log_cont("[%u]", t->array_size + 1);
2288
2289 if (t->levels)
2290 pandecode_log_cont(" mip %u", t->levels);
2291
2292 pandecode_log_cont("\n");
2293
2294 if (f.unknown1 | f.zero) {
2295 pandecode_msg("XXX: texture format zero tripped\n");
2296 pandecode_prop("unknown1 = %" PRId32, f.unknown1);
2297 pandecode_prop("zero = %" PRId32, f.zero);
2298 }
2299
2300 if (!f.unknown2) {
2301 pandecode_msg("XXX: expected unknown texture bit set\n");
2302 pandecode_prop("unknown2 = %" PRId32, f.unknown2);
2303 }
2304
2305 if (t->swizzle_zero) {
2306 pandecode_msg("XXX: swizzle zero tripped\n");
2307 pandecode_prop("swizzle_zero = %d", t->swizzle_zero);
2308 }
2309
2310 if (t->unknown3 | t->unknown3A | t->unknown5 | t->unknown6 | t->unknown7) {
2311 pandecode_msg("XXX: texture zero tripped\n");
2312 pandecode_prop("unknown3 = %" PRId16, t->unknown3);
2313 pandecode_prop("unknown3A = %" PRId8, t->unknown3A);
2314 pandecode_prop("unknown5 = 0x%" PRIx32, t->unknown5);
2315 pandecode_prop("unknown6 = 0x%" PRIx32, t->unknown6);
2316 pandecode_prop("unknown7 = 0x%" PRIx32, t->unknown7);
2317 }
2318
2319 pandecode_texture_payload(u + sizeof(*t), f.type, f.layout, f.manual_stride, t->levels, t->depth, t->array_size, tmem);
2320
2321 pandecode_indent--;
2322 pandecode_log("};\n");
2323 }
2324
2325 static void
2326 pandecode_bifrost_texture(
2327 const struct bifrost_texture_descriptor *t,
2328 unsigned job_no,
2329 unsigned tex)
2330 {
2331 pandecode_log("struct bifrost_texture_descriptor texture_descriptor_%d_%d = {\n", job_no, tex);
2332 pandecode_indent++;
2333
2334 pandecode_prop("format_unk = 0x%" PRIx32, t->format_unk);
2335 pandecode_prop("type = %" PRId32, t->type);
2336
2337 if (t->zero) {
2338 pandecode_msg("XXX: zero tripped\n");
2339 pandecode_prop("zero = 0x%" PRIx32, t->zero);
2340 }
2341
2342 pandecode_prop("format_swizzle = 0x%" PRIx32, t->format_swizzle);
2343 pandecode_prop("format = 0x%" PRIx32, t->format);
2344 pandecode_prop("srgb = 0x%" PRIx32, t->srgb);
2345 pandecode_prop("format_unk3 = 0x%" PRIx32, t->format_unk3);
2346 pandecode_prop("width = %" PRId32, t->width);
2347 pandecode_prop("height = %" PRId32, t->height);
2348 pandecode_prop("swizzle = 0x%" PRIx32, t->swizzle);
2349 pandecode_prop("levels = %" PRId32, t->levels);
2350 pandecode_prop("unk1 = 0x%" PRIx32, t->unk1);
2351 pandecode_prop("levels_unk = %" PRId32, t->levels_unk);
2352 pandecode_prop("level_2 = %" PRId32, t->level_2);
2353 pandecode_prop("payload = 0x%" PRIx64, t->payload);
2354 pandecode_prop("array_size = %" PRId32, t->array_size);
2355 pandecode_prop("unk4 = 0x%" PRIx32, t->unk4);
2356 pandecode_prop("depth = %" PRId32, t->depth);
2357 pandecode_prop("unk5 = 0x%" PRIx32, t->unk5);
2358 pandecode_log("\n");
2359
2360 /* See the definiton of enum mali_texture_type */
2361
2362 bool is_cube = t->type == MALI_TEX_CUBE;
2363 unsigned dimension = is_cube ? 2 : t->type;
2364
2365 /* Print the layout. Default is linear; a modifier can denote AFBC or
2366 * u-interleaved/tiled modes */
2367
2368 if (t->layout == MALI_TEXTURE_AFBC)
2369 pandecode_log_cont("afbc");
2370 else if (t->layout == MALI_TEXTURE_TILED)
2371 pandecode_log_cont("tiled");
2372 else if (t->layout == MALI_TEXTURE_LINEAR)
2373 pandecode_log_cont("linear");
2374 else
2375 pandecode_msg("XXX: invalid texture layout 0x%X\n", t->layout);
2376
2377 pandecode_swizzle(t->swizzle, t->format);
2378 pandecode_log_cont(" ");
2379
2380 /* Distinguish cube/2D with modifier */
2381
2382 if (is_cube)
2383 pandecode_log_cont("cube ");
2384
2385 pandecode_format_short(t->format, t->srgb);
2386
2387 /* All four width/height/depth/array_size dimensions are present
2388 * regardless of the type of texture, but it is an error to have
2389 * non-zero dimensions for unused dimensions. Verify this. array_size
2390 * can always be set, as can width. */
2391
2392 if (t->height && dimension < 2)
2393 pandecode_msg("XXX: nonzero height for <2D texture\n");
2394
2395 if (t->depth && dimension < 3)
2396 pandecode_msg("XXX: nonzero depth for <2D texture\n");
2397
2398 /* Print only the dimensions that are actually there */
2399
2400 pandecode_log_cont(": %d", t->width + 1);
2401
2402 if (dimension >= 2)
2403 pandecode_log_cont("x%u", t->height + 1);
2404
2405 if (dimension >= 3)
2406 pandecode_log_cont("x%u", t->depth + 1);
2407
2408 if (t->array_size)
2409 pandecode_log_cont("[%u]", t->array_size + 1);
2410
2411 if (t->levels)
2412 pandecode_log_cont(" mip %u", t->levels);
2413
2414 pandecode_log_cont("\n");
2415
2416 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(t->payload);
2417 if (t->payload) {
2418 pandecode_texture_payload(t->payload, t->type, t->layout,
2419 true, t->levels, t->depth,
2420 t->array_size, tmem);
2421 }
2422
2423 pandecode_indent--;
2424 pandecode_log("};\n");
2425 }
2426
2427 /* 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 */
2428
2429 static void
2430 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
2431 {
2432 /* Nothing to do */
2433 if (claim == truth)
2434 return;
2435
2436 if (fuzzy && (truth < 0))
2437 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
2438
2439 if ((truth >= 0) && !fuzzy) {
2440 pandecode_msg("%s: expected %s = %d, claimed %u\n",
2441 (truth < claim) ? "warn" : "XXX",
2442 name, truth, claim);
2443 } else if ((claim > -truth) && !fuzzy) {
2444 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
2445 name, -truth, claim);
2446 } else if (fuzzy && (claim < truth))
2447 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
2448 name, truth, claim);
2449
2450 pandecode_log(".%s = %" PRId16, name, claim);
2451
2452 if (fuzzy)
2453 pandecode_log_cont(" /* %u used */", truth);
2454
2455 pandecode_log_cont(",\n");
2456 }
2457
2458 static void
2459 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
2460 bool is_bifrost, unsigned gpu_id)
2461 {
2462 struct midgard_disasm_stats stats =
2463 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
2464
2465 bool has_texture = (stats.texture_count > 0);
2466 bool has_sampler = (stats.sampler_count > 0);
2467 bool has_attribute = (stats.attribute_count > 0);
2468 bool has_varying = (stats.varying_count > 0);
2469 bool has_uniform = (stats.uniform_count > 0);
2470 bool has_ubo = (stats.uniform_buffer_count > 0);
2471
2472 if (has_texture || has_sampler)
2473 pandecode_msg("XXX: blend shader accessing textures\n");
2474
2475 if (has_attribute || has_varying)
2476 pandecode_msg("XXX: blend shader accessing interstage\n");
2477
2478 if (has_uniform || has_ubo)
2479 pandecode_msg("XXX: blend shader accessing uniforms\n");
2480 }
2481
2482 static void
2483 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
2484 {
2485 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
2486
2487 if (!mmem)
2488 return;
2489
2490 if (is_bifrost) {
2491 const struct bifrost_texture_descriptor *PANDECODE_PTR_VAR(t, mmem, textures);
2492
2493 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2494 pandecode_indent++;
2495
2496 for (unsigned tex = 0; tex < texture_count; ++tex)
2497 pandecode_bifrost_texture(&t[tex], job_no, tex);
2498
2499 pandecode_indent--;
2500 pandecode_log("};\n");
2501 } else {
2502 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
2503
2504 pandecode_log("uint64_t textures_%"PRIx64"_%d[] = {\n", textures, job_no);
2505 pandecode_indent++;
2506
2507 for (int tex = 0; tex < texture_count; ++tex) {
2508 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2509 char *a = pointer_as_memory_reference(*u);
2510 pandecode_log("%s,\n", a);
2511 free(a);
2512 }
2513
2514 pandecode_indent--;
2515 pandecode_log("};\n");
2516
2517 /* Now, finally, descend down into the texture descriptor */
2518 for (unsigned tex = 0; tex < texture_count; ++tex) {
2519 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
2520 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
2521 if (tmem)
2522 pandecode_texture(*u, tmem, job_no, tex);
2523 }
2524 }
2525 }
2526
2527 static void
2528 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
2529 {
2530 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(samplers);
2531
2532 if (!smem)
2533 return;
2534
2535 if (is_bifrost) {
2536 struct bifrost_sampler_descriptor *s;
2537
2538 for (int i = 0; i < sampler_count; ++i) {
2539 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2540
2541 pandecode_log("struct bifrost_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2542 pandecode_indent++;
2543
2544 if (s->unk1 != 1) {
2545 pandecode_msg("XXX: unk1 tripped\n");
2546 pandecode_prop("unk1 = 0x%x", s->unk1);
2547 }
2548
2549 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2550 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2551 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2552
2553 if (s->unk8 != 0x8) {
2554 pandecode_msg("XXX: unk8 tripped\n");
2555 pandecode_prop("unk8 = 0x%x", s->unk8);
2556 }
2557
2558 pandecode_prop("unk2 = 0x%x", s->unk2);
2559 pandecode_prop("unk3 = 0x%x", s->unk3);
2560 pandecode_prop("min_filter = %s", s->min_filter ? "nearest" : "linear");
2561 pandecode_prop("norm_coords = 0x%x", s->norm_coords & 0x1);
2562 pandecode_prop("zero1 = 0x%x", s->zero1 & 0x1);
2563 pandecode_prop("mip_filter = %s", s->mip_filter ? "linear" : "nearest");
2564 pandecode_prop("mag_filter = %s", s->mag_filter ? "linear" : "nearest");
2565
2566 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2567 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2568
2569 if (s->zero1 || s->zero2 || s->zero3 || s->zero4) {
2570 pandecode_msg("XXX: sampler zero tripped\n");
2571 pandecode_prop("zero = 0x%" PRIx8 ", 0x%" PRIx64 ", 0x%" PRIx64 ", 0x%" PRIx64 "\n", s->zero1, s->zero2, s->zero3, s->zero4);
2572 }
2573
2574 pandecode_indent--;
2575 pandecode_log("};\n");
2576 }
2577 } else {
2578 struct mali_sampler_descriptor *s;
2579
2580 for (int i = 0; i < sampler_count; ++i) {
2581 s = pandecode_fetch_gpu_mem(smem, samplers + sizeof(*s) * i, sizeof(*s));
2582
2583 pandecode_log("struct mali_sampler_descriptor sampler_descriptor_%"PRIx64"_%d_%d = {\n", samplers + sizeof(*s) * i, job_no, i);
2584 pandecode_indent++;
2585
2586 pandecode_log(".filter_mode = ");
2587 pandecode_log_decoded_flags(sampler_flag_info, s->filter_mode);
2588 pandecode_log_cont(",\n");
2589
2590 pandecode_prop("min_lod = FIXED_16(%f)", DECODE_FIXED_16(s->min_lod));
2591 pandecode_prop("max_lod = FIXED_16(%f)", DECODE_FIXED_16(s->max_lod));
2592
2593 if (s->lod_bias)
2594 pandecode_prop("lod_bias = FIXED_16(%f)", DECODE_FIXED_16(s->lod_bias));
2595
2596 pandecode_prop("wrap_s = %s", pandecode_wrap_mode(s->wrap_s));
2597 pandecode_prop("wrap_t = %s", pandecode_wrap_mode(s->wrap_t));
2598 pandecode_prop("wrap_r = %s", pandecode_wrap_mode(s->wrap_r));
2599
2600 pandecode_prop("compare_func = %s", pandecode_func(s->compare_func));
2601
2602 if (s->zero || s->zero2) {
2603 pandecode_msg("XXX: sampler zero tripped\n");
2604 pandecode_prop("zero = 0x%X, 0x%X\n", s->zero, s->zero2);
2605 }
2606
2607 pandecode_prop("seamless_cube_map = %d", s->seamless_cube_map);
2608
2609 pandecode_prop("border_color = { %f, %f, %f, %f }",
2610 s->border_color[0],
2611 s->border_color[1],
2612 s->border_color[2],
2613 s->border_color[3]);
2614
2615 pandecode_indent--;
2616 pandecode_log("};\n");
2617 }
2618 }
2619 }
2620
2621 static void
2622 pandecode_vertex_tiler_postfix_pre(
2623 const struct mali_vertex_tiler_postfix *p,
2624 int job_no, enum mali_job_type job_type,
2625 char *suffix, bool is_bifrost, unsigned gpu_id)
2626 {
2627 struct pandecode_mapped_memory *attr_mem;
2628
2629 /* On Bifrost, since the tiler heap (for tiler jobs) and the scratchpad
2630 * are the only things actually needed from the FBD, vertex/tiler jobs
2631 * no longer reference the FBD -- instead, this field points to some
2632 * info about the scratchpad.
2633 */
2634
2635 struct pandecode_fbd fbd_info = {
2636 /* Default for Bifrost */
2637 .rt_count = 1
2638 };
2639
2640 if (is_bifrost) {
2641 pandecode_log_cont("\t/* %X %/\n", p->shared_memory & 1);
2642 pandecode_compute_fbd(p->shared_memory & ~1, job_no);
2643 } else if (p->shared_memory & MALI_MFBD)
2644 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared_memory) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
2645 else if (job_type == MALI_JOB_TYPE_COMPUTE)
2646 pandecode_compute_fbd((u64) (uintptr_t) p->shared_memory, job_no);
2647 else
2648 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared_memory, job_no, false, gpu_id);
2649
2650 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
2651 int texture_count = 0, sampler_count = 0;
2652
2653 if (p->shader) {
2654 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->shader);
2655 struct mali_shader_meta *PANDECODE_PTR_VAR(s, smem, p->shader);
2656
2657 /* Disassemble ahead-of-time to get stats. Initialize with
2658 * stats for the missing-shader case so we get validation
2659 * there, too */
2660
2661 struct midgard_disasm_stats info = {
2662 .texture_count = 0,
2663 .sampler_count = 0,
2664 .attribute_count = 0,
2665 .varying_count = 0,
2666 .work_count = 1,
2667
2668 .uniform_count = -128,
2669 .uniform_buffer_count = 0
2670 };
2671
2672 if (s->shader & ~0xF)
2673 info = pandecode_shader_disassemble(s->shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
2674
2675 pandecode_log("struct mali_shader_meta shader_meta_%"PRIx64"_%d%s = {\n", p->shader, job_no, suffix);
2676 pandecode_indent++;
2677
2678 /* Save for dumps */
2679 attribute_count = s->attribute_count;
2680 varying_count = s->varying_count;
2681 texture_count = s->texture_count;
2682 sampler_count = s->sampler_count;
2683
2684 if (is_bifrost) {
2685 uniform_count = s->bifrost2.uniform_count;
2686 uniform_buffer_count = s->bifrost1.uniform_buffer_count;
2687 } else {
2688 uniform_count = s->midgard1.uniform_count;
2689 uniform_buffer_count = s->midgard1.uniform_buffer_count;
2690 }
2691
2692 pandecode_shader_address("shader", s->shader);
2693
2694 pandecode_shader_prop("texture_count", s->texture_count, info.texture_count, false);
2695 pandecode_shader_prop("sampler_count", s->sampler_count, info.sampler_count, false);
2696 pandecode_shader_prop("attribute_count", s->attribute_count, info.attribute_count, false);
2697 pandecode_shader_prop("varying_count", s->varying_count, info.varying_count, false);
2698 pandecode_shader_prop("uniform_buffer_count",
2699 uniform_buffer_count,
2700 info.uniform_buffer_count, true);
2701
2702 if (!is_bifrost) {
2703 pandecode_shader_prop("uniform_count",
2704 uniform_count,
2705 info.uniform_count, false);
2706
2707 pandecode_shader_prop("work_count",
2708 s->midgard1.work_count, info.work_count, false);
2709 }
2710
2711 if (is_bifrost) {
2712 pandecode_log("bifrost1.unk1 = ");
2713 pandecode_log_decoded_flags(shader_bifrost_info, s->bifrost1.unk1);
2714 pandecode_log_cont(",\n");
2715 } else {
2716 bool helpers = s->midgard1.flags_lo & MALI_HELPER_INVOCATIONS;
2717
2718 if (helpers != info.helper_invocations) {
2719 pandecode_msg("XXX: expected helpers %u but got %u\n",
2720 info.helper_invocations, helpers);
2721 }
2722
2723 pandecode_log(".midgard1.flags_lo = ");
2724 pandecode_log_decoded_flags(shader_midgard1_flag_lo_info,
2725 s->midgard1.flags_lo & ~MALI_HELPER_INVOCATIONS);
2726 pandecode_log_cont(",\n");
2727
2728 pandecode_log(".midgard1.flags_hi = ");
2729 pandecode_log_decoded_flags(shader_midgard1_flag_hi_info, s->midgard1.flags_hi);
2730 pandecode_log_cont(",\n");
2731 }
2732
2733 if (s->depth_units || s->depth_factor) {
2734 pandecode_prop("depth_factor = %f", s->depth_factor);
2735 pandecode_prop("depth_units = %f", s->depth_units);
2736 }
2737
2738 if (s->coverage_mask)
2739 pandecode_prop("coverage_mask = 0x%X", s->coverage_mask);
2740
2741 if (s->unknown2_2)
2742 pandecode_prop(".unknown2_2 = %X", s->unknown2_2);
2743
2744 if (s->unknown2_3 || s->unknown2_4) {
2745 pandecode_log(".unknown2_3 = ");
2746
2747 int unknown2_3 = s->unknown2_3;
2748 int unknown2_4 = s->unknown2_4;
2749
2750 /* We're not quite sure what these flags mean without the depth test, if anything */
2751
2752 if (unknown2_3 & (MALI_DEPTH_WRITEMASK | MALI_DEPTH_FUNC_MASK)) {
2753 const char *func = pandecode_func(MALI_GET_DEPTH_FUNC(unknown2_3));
2754 unknown2_3 &= ~MALI_DEPTH_FUNC_MASK;
2755
2756 pandecode_log_cont("MALI_DEPTH_FUNC(%s) | ", func);
2757 }
2758
2759 pandecode_log_decoded_flags(u3_flag_info, unknown2_3);
2760 pandecode_log_cont(",\n");
2761
2762 pandecode_log(".unknown2_4 = ");
2763 pandecode_log_decoded_flags(u4_flag_info, unknown2_4);
2764 pandecode_log_cont(",\n");
2765 }
2766
2767 if (s->stencil_mask_front || s->stencil_mask_back) {
2768 pandecode_prop("stencil_mask_front = 0x%02X", s->stencil_mask_front);
2769 pandecode_prop("stencil_mask_back = 0x%02X", s->stencil_mask_back);
2770 }
2771
2772 pandecode_stencil("front", &s->stencil_front);
2773 pandecode_stencil("back", &s->stencil_back);
2774
2775 if (is_bifrost) {
2776 pandecode_log(".bifrost2 = {\n");
2777 pandecode_indent++;
2778
2779 pandecode_prop("unk3 = 0x%" PRIx32, s->bifrost2.unk3);
2780 pandecode_prop("preload_regs = 0x%" PRIx32, s->bifrost2.preload_regs);
2781 pandecode_prop("uniform_count = %" PRId32, s->bifrost2.uniform_count);
2782 pandecode_prop("unk4 = 0x%" PRIx32, s->bifrost2.unk4);
2783
2784 pandecode_indent--;
2785 pandecode_log("},\n");
2786 } else if (s->midgard2.unknown2_7) {
2787 pandecode_log(".midgard2 = {\n");
2788 pandecode_indent++;
2789
2790 pandecode_prop("unknown2_7 = 0x%" PRIx32, s->midgard2.unknown2_7);
2791 pandecode_indent--;
2792 pandecode_log("},\n");
2793 }
2794
2795 if (s->padding) {
2796 pandecode_msg("XXX: shader padding tripped\n");
2797 pandecode_prop("padding = 0x%" PRIx32, s->padding);
2798 }
2799
2800 if (!is_bifrost) {
2801 /* TODO: Blend shaders routing/disasm */
2802 union midgard_blend blend = s->blend;
2803 mali_ptr shader = pandecode_midgard_blend(&blend, s->unknown2_3 & MALI_HAS_BLEND_SHADER);
2804 if (shader & ~0xF)
2805 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2806 } else {
2807 pandecode_msg("mdg_blend = %" PRIx64 "\n", s->blend.shader);
2808 }
2809
2810 pandecode_indent--;
2811 pandecode_log("};\n");
2812
2813 /* MRT blend fields are used whenever MFBD is used, with
2814 * per-RT descriptors */
2815
2816 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared_memory & MALI_MFBD)) {
2817 void* blend_base = (void *) (s + 1);
2818
2819 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
2820 mali_ptr shader = 0;
2821
2822 if (is_bifrost)
2823 shader = pandecode_bifrost_blend(blend_base, job_no, i);
2824 else
2825 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
2826
2827 if (shader & ~0xF)
2828 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
2829
2830 }
2831 }
2832 } else
2833 pandecode_msg("XXX: missing shader descriptor\n");
2834
2835 if (p->viewport) {
2836 struct pandecode_mapped_memory *fmem = pandecode_find_mapped_gpu_mem_containing(p->viewport);
2837 struct mali_viewport *PANDECODE_PTR_VAR(f, fmem, p->viewport);
2838
2839 pandecode_log("struct mali_viewport viewport_%"PRIx64"_%d%s = {\n", p->viewport, job_no, suffix);
2840 pandecode_indent++;
2841
2842 pandecode_prop("clip_minx = %f", f->clip_minx);
2843 pandecode_prop("clip_miny = %f", f->clip_miny);
2844 pandecode_prop("clip_minz = %f", f->clip_minz);
2845 pandecode_prop("clip_maxx = %f", f->clip_maxx);
2846 pandecode_prop("clip_maxy = %f", f->clip_maxy);
2847 pandecode_prop("clip_maxz = %f", f->clip_maxz);
2848
2849 /* Only the higher coordinates are MALI_POSITIVE scaled */
2850
2851 pandecode_prop("viewport0 = { %d, %d }",
2852 f->viewport0[0], f->viewport0[1]);
2853
2854 pandecode_prop("viewport1 = { MALI_POSITIVE(%d), MALI_POSITIVE(%d) }",
2855 f->viewport1[0] + 1, f->viewport1[1] + 1);
2856
2857 pandecode_indent--;
2858 pandecode_log("};\n");
2859 }
2860
2861 unsigned max_attr_index = 0;
2862
2863 if (p->attribute_meta)
2864 max_attr_index = pandecode_attribute_meta(job_no, attribute_count, p, false, suffix);
2865
2866 if (p->attributes) {
2867 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attributes);
2868 pandecode_attributes(attr_mem, p->attributes, job_no, suffix, max_attr_index, false, job_type);
2869 }
2870
2871 /* Varyings are encoded like attributes but not actually sent; we just
2872 * pass a zero buffer with the right stride/size set, (or whatever)
2873 * since the GPU will write to it itself */
2874
2875 if (p->varying_meta) {
2876 varying_count = pandecode_attribute_meta(job_no, varying_count, p, true, suffix);
2877 }
2878
2879 if (p->varyings) {
2880 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varyings);
2881
2882 /* Number of descriptors depends on whether there are
2883 * non-internal varyings */
2884
2885 pandecode_attributes(attr_mem, p->varyings, job_no, suffix, varying_count, true, job_type);
2886 }
2887
2888 if (p->uniform_buffers) {
2889 if (uniform_buffer_count)
2890 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
2891 else
2892 pandecode_msg("warn: UBOs specified but not referenced\n");
2893 } else if (uniform_buffer_count)
2894 pandecode_msg("XXX: UBOs referenced but not specified\n");
2895
2896 /* We don't want to actually dump uniforms, but we do need to validate
2897 * that the counts we were given are sane */
2898
2899 if (p->uniforms) {
2900 if (uniform_count)
2901 pandecode_uniforms(p->uniforms, uniform_count);
2902 else
2903 pandecode_msg("warn: Uniforms specified but not referenced\n");
2904 } else if (uniform_count)
2905 pandecode_msg("XXX: Uniforms referenced but not specified\n");
2906
2907 if (p->textures)
2908 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
2909
2910 if (p->sampler_descriptor)
2911 pandecode_samplers(p->sampler_descriptor, sampler_count, job_no, is_bifrost);
2912 }
2913
2914 static void
2915 pandecode_gl_enables(uint32_t gl_enables, int job_type)
2916 {
2917 pandecode_log(".gl_enables = ");
2918
2919 pandecode_log_decoded_flags(gl_enable_flag_info, gl_enables);
2920
2921 pandecode_log_cont(",\n");
2922 }
2923
2924 static void
2925 pandecode_vertex_tiler_postfix(const struct mali_vertex_tiler_postfix *p, int job_no, bool is_bifrost)
2926 {
2927 if (p->shader & 0xF)
2928 pandecode_msg("warn: shader tagged %X\n", (unsigned) (p->shader & 0xF));
2929
2930 pandecode_log(".postfix = {\n");
2931 pandecode_indent++;
2932
2933 pandecode_gl_enables(p->gl_enables, MALI_JOB_TYPE_TILER);
2934 pandecode_prop("instance_shift = 0x%x", p->instance_shift);
2935 pandecode_prop("instance_odd = 0x%x", p->instance_odd);
2936
2937 if (p->zero4) {
2938 pandecode_msg("XXX: vertex only zero tripped");
2939 pandecode_prop("zero4 = 0x%" PRIx32, p->zero4);
2940 }
2941
2942 pandecode_prop("offset_start = 0x%x", p->offset_start);
2943
2944 if (p->zero5) {
2945 pandecode_msg("XXX: vertex only zero tripped");
2946 pandecode_prop("zero5 = 0x%" PRIx64, p->zero5);
2947 }
2948
2949 MEMORY_PROP(p, position_varying);
2950 MEMORY_PROP(p, occlusion_counter);
2951
2952 pandecode_indent--;
2953 pandecode_log("},\n");
2954 }
2955
2956 static void
2957 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
2958 {
2959 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
2960 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
2961
2962 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
2963 pandecode_indent++;
2964
2965 if (h->zero) {
2966 pandecode_msg("XXX: tiler heap zero tripped\n");
2967 pandecode_prop("zero = 0x%x", h->zero);
2968 }
2969
2970 pandecode_prop("heap_size = 0x%x", h->heap_size);
2971 MEMORY_PROP(h, tiler_heap_start);
2972 MEMORY_PROP(h, tiler_heap_free);
2973
2974 /* this might point to the beginning of another buffer, when it's
2975 * really the end of the tiler heap buffer, so we have to be careful
2976 * here. but for zero length, we need the same pointer.
2977 */
2978
2979 if (h->tiler_heap_end == h->tiler_heap_start) {
2980 MEMORY_PROP(h, tiler_heap_start);
2981 } else {
2982 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
2983 pandecode_prop("tiler_heap_end = %s + 1", a);
2984 free(a);
2985 }
2986
2987 for (int i = 0; i < 10; i++) {
2988 if (h->zeros[i] != 0) {
2989 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
2990 i, h->zeros[i]);
2991 }
2992 }
2993
2994 if (h->unk1 != 0x1) {
2995 pandecode_msg("XXX: tiler heap unk1 tripped\n");
2996 pandecode_prop("unk1 = 0x%x", h->unk1);
2997 }
2998
2999 if (h->unk7e007e != 0x7e007e) {
3000 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
3001 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
3002 }
3003
3004 pandecode_indent--;
3005 pandecode_log("};\n");
3006 }
3007
3008 static void
3009 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
3010 {
3011 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
3012 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
3013
3014 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
3015
3016 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
3017 pandecode_indent++;
3018
3019 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
3020 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
3021
3022 if (t->hierarchy_mask != 0xa &&
3023 t->hierarchy_mask != 0x14 &&
3024 t->hierarchy_mask != 0x28 &&
3025 t->hierarchy_mask != 0x50 &&
3026 t->hierarchy_mask != 0xa0)
3027 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
3028
3029 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
3030
3031 pandecode_prop("flags = 0x%" PRIx16, t->flags);
3032
3033 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
3034 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
3035
3036 if (t->zero0) {
3037 pandecode_msg("XXX: tiler meta zero tripped\n");
3038 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
3039 }
3040
3041 for (int i = 0; i < 12; i++) {
3042 if (t->zeros[i] != 0) {
3043 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
3044 i, t->zeros[i]);
3045 }
3046 }
3047
3048 pandecode_indent--;
3049 pandecode_log("};\n");
3050 }
3051
3052 static void
3053 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
3054 {
3055 if (u.pointer == 0x0)
3056 return;
3057
3058 pandecode_log(".primitive_size = {\n");
3059 pandecode_indent++;
3060
3061 if (constant) {
3062 pandecode_prop("constant = %f", u.constant);
3063 } else {
3064 MEMORY_PROP((&u), pointer);
3065 }
3066
3067 pandecode_indent--;
3068 pandecode_log("},\n");
3069 }
3070
3071 static void
3072 pandecode_tiler_only_bfr(const struct bifrost_tiler_only *t, int job_no)
3073 {
3074 pandecode_log_cont("{\n");
3075 pandecode_indent++;
3076
3077 /* TODO: gl_PointSize on Bifrost */
3078 pandecode_primitive_size(t->primitive_size, true);
3079
3080 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
3081 || t->zero6) {
3082 pandecode_msg("XXX: tiler only zero tripped\n");
3083 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
3084 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
3085 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
3086 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
3087 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
3088 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
3089 }
3090
3091 pandecode_indent--;
3092 pandecode_log("},\n");
3093 }
3094
3095 static int
3096 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
3097 const struct pandecode_mapped_memory *mem,
3098 mali_ptr payload, int job_no, unsigned gpu_id)
3099 {
3100 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
3101
3102 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", true, gpu_id);
3103
3104 pandecode_log("struct bifrost_payload_vertex payload_%"PRIx64"_%d = {\n", payload, job_no);
3105 pandecode_indent++;
3106
3107 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
3108 pandecode_vertex_tiler_postfix(&v->postfix, job_no, true);
3109
3110 pandecode_indent--;
3111 pandecode_log("};\n");
3112
3113 return sizeof(*v);
3114 }
3115
3116 static int
3117 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
3118 const struct pandecode_mapped_memory *mem,
3119 mali_ptr payload, int job_no, unsigned gpu_id)
3120 {
3121 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
3122
3123 pandecode_vertex_tiler_postfix_pre(&t->postfix, job_no, h->job_type, "", true, gpu_id);
3124 pandecode_tiler_meta(t->tiler.tiler_meta, job_no);
3125
3126 pandecode_log("struct bifrost_payload_tiler payload_%"PRIx64"_%d = {\n", payload, job_no);
3127 pandecode_indent++;
3128
3129 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
3130
3131 pandecode_log(".tiler = ");
3132 pandecode_tiler_only_bfr(&t->tiler, job_no);
3133
3134 pandecode_vertex_tiler_postfix(&t->postfix, job_no, true);
3135
3136 pandecode_indent--;
3137 pandecode_log("};\n");
3138
3139 return sizeof(*t);
3140 }
3141
3142 static int
3143 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
3144 const struct pandecode_mapped_memory *mem,
3145 mali_ptr payload, int job_no, unsigned gpu_id)
3146 {
3147 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
3148 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
3149
3150 pandecode_vertex_tiler_postfix_pre(&v->postfix, job_no, h->job_type, "", false, gpu_id);
3151
3152 pandecode_log("struct midgard_payload_vertex_tiler payload_%d = {\n", job_no);
3153 pandecode_indent++;
3154
3155 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
3156 pandecode_vertex_tiler_postfix(&v->postfix, job_no, false);
3157
3158 bool has_primitive_pointer = v->prefix.unknown_draw & MALI_DRAW_VARYING_SIZE;
3159 pandecode_primitive_size(v->primitive_size, !has_primitive_pointer);
3160
3161 pandecode_indent--;
3162 pandecode_log("};\n");
3163
3164 return sizeof(*v);
3165 }
3166
3167 static int
3168 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
3169 mali_ptr payload, int job_no,
3170 bool is_bifrost, unsigned gpu_id)
3171 {
3172 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
3173
3174 bool is_mfbd = s->framebuffer & MALI_MFBD;
3175
3176 if (!is_mfbd && is_bifrost)
3177 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
3178
3179 struct pandecode_fbd info;
3180
3181 if (is_mfbd)
3182 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
3183 else
3184 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
3185
3186 /* Compute the tag for the tagged pointer. This contains the type of
3187 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
3188 * additional structures follow the MFBD header (an extra payload or
3189 * not, as well as a count of render targets) */
3190
3191 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
3192
3193 if (is_mfbd) {
3194 if (info.has_extra)
3195 expected_tag |= MALI_MFBD_TAG_EXTRA;
3196
3197 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
3198 }
3199
3200 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
3201 pandecode_msg("XXX: unexpected tile coordinate bits\n");
3202 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
3203 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
3204 }
3205
3206 /* Extract tile coordinates */
3207
3208 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
3209 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
3210
3211 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3212 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
3213
3214 /* For the max, we also want the floored (rather than ceiled) version for checking */
3215
3216 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
3217 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
3218
3219 /* Validate the coordinates are well-ordered */
3220
3221 if (min_x == max_x)
3222 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3223 else if (min_x > max_x)
3224 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3225
3226 if (min_y == max_y)
3227 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
3228 else if (min_y > max_y)
3229 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
3230
3231 /* Validate the coordinates fit inside the framebuffer. We use floor,
3232 * rather than ceil, for the max coordinates, since the tile
3233 * coordinates for something like an 800x600 framebuffer will actually
3234 * resolve to 800x608, which would otherwise trigger a Y-overflow */
3235
3236 if ((min_x > info.width) || (max_x_f > info.width))
3237 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
3238
3239 if ((min_y > info.height) || (max_y_f > info.height))
3240 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
3241
3242 /* After validation, we print */
3243
3244 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
3245
3246 /* The FBD is a tagged pointer */
3247
3248 unsigned tag = (s->framebuffer & ~FBD_MASK);
3249
3250 if (tag != expected_tag)
3251 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
3252
3253 return sizeof(*s);
3254 }
3255
3256 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
3257 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
3258 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
3259 * might wish to combine this with the bifrost parameter) because some details
3260 * are model-specific even within a particular architecture. Minimal traces
3261 * *only* examine the job descriptors, skipping printing entirely if there is
3262 * no faults, and only descends into the payload if there are faults. This is
3263 * useful for looking for faults without the overhead of invasive traces. */
3264
3265 void
3266 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
3267 {
3268 pandecode_dump_file_open();
3269
3270 struct mali_job_descriptor_header *h;
3271 unsigned job_descriptor_number = 0;
3272
3273 do {
3274 struct pandecode_mapped_memory *mem =
3275 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
3276
3277 void *payload;
3278
3279 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
3280
3281 /* On Midgard, for 32-bit jobs except for fragment jobs, the
3282 * high 32-bits of the 64-bit pointer are reused to store
3283 * something else.
3284 */
3285 int offset = h->job_descriptor_size == MALI_JOB_32 &&
3286 h->job_type != MALI_JOB_TYPE_FRAGMENT ? 4 : 0;
3287 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h) - offset;
3288
3289 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 256);
3290
3291 int job_no = job_descriptor_number++;
3292
3293 /* If the job is good to go, skip it in minimal mode */
3294 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
3295 continue;
3296
3297 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
3298 pandecode_indent++;
3299
3300 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
3301
3302 if (h->job_descriptor_size)
3303 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
3304
3305 if (h->exception_status && h->exception_status != 0x1)
3306 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
3307 h->exception_status,
3308 (h->exception_status >> 16) & 0xFFFF,
3309 pandecode_exception_access((h->exception_status >> 8) & 0x3),
3310 h->exception_status & 0xFF);
3311
3312 if (h->first_incomplete_task)
3313 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
3314
3315 if (h->fault_pointer)
3316 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
3317
3318 if (h->job_barrier)
3319 pandecode_prop("job_barrier = %d", h->job_barrier);
3320
3321 pandecode_prop("job_index = %d", h->job_index);
3322
3323 if (h->unknown_flags)
3324 pandecode_prop("unknown_flags = %d", h->unknown_flags);
3325
3326 if (h->job_dependency_index_1)
3327 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
3328
3329 if (h->job_dependency_index_2)
3330 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
3331
3332 pandecode_indent--;
3333 pandecode_log("};\n");
3334
3335 switch (h->job_type) {
3336 case MALI_JOB_TYPE_WRITE_VALUE: {
3337 struct mali_payload_write_value *s = payload;
3338 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
3339 pandecode_indent++;
3340 MEMORY_PROP(s, address);
3341
3342 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
3343 pandecode_msg("XXX: unknown value descriptor\n");
3344 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
3345 }
3346
3347 if (s->reserved) {
3348 pandecode_msg("XXX: set value tripped\n");
3349 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
3350 }
3351
3352 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
3353 pandecode_indent--;
3354 pandecode_log("};\n");
3355
3356 break;
3357 }
3358
3359 case MALI_JOB_TYPE_TILER:
3360 case MALI_JOB_TYPE_VERTEX:
3361 case MALI_JOB_TYPE_COMPUTE:
3362 if (bifrost) {
3363 if (h->job_type == MALI_JOB_TYPE_TILER)
3364 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3365 else
3366 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
3367 } else
3368 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
3369
3370 break;
3371
3372 case MALI_JOB_TYPE_FRAGMENT:
3373 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
3374 break;
3375
3376 default:
3377 break;
3378 }
3379 } while ((jc_gpu_va = h->next_job));
3380
3381 pandecode_map_read_write();
3382 }