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