pan/decode: Use unpack for vertex_tiler_postfix_pre
[mesa.git] / src / panfrost / lib / decode.c
1 /*
2 * Copyright (C) 2017-2019 Alyssa Rosenzweig
3 * Copyright (C) 2017-2019 Connor Abbott
4 * Copyright (C) 2019 Collabora, Ltd.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 */
25
26 #include <midgard_pack.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <memory.h>
30 #include <stdbool.h>
31 #include <stdarg.h>
32 #include <ctype.h>
33 #include "decode.h"
34 #include "util/macros.h"
35 #include "util/u_math.h"
36
37 #include "midgard/disassemble.h"
38 #include "bifrost/disassemble.h"
39
40 #include "pan_encoder.h"
41
42 static void pandecode_swizzle(unsigned swizzle, enum mali_format format);
43
44 #define MEMORY_PROP(obj, p) {\
45 if (obj->p) { \
46 char *a = pointer_as_memory_reference(obj->p); \
47 pandecode_prop("%s = %s", #p, a); \
48 free(a); \
49 } \
50 }
51
52 #define MEMORY_PROP_DIR(obj, p) {\
53 if (obj.p) { \
54 char *a = pointer_as_memory_reference(obj.p); \
55 pandecode_prop("%s = %s", #p, a); \
56 free(a); \
57 } \
58 }
59
60 #define DUMP_CL(title, T, cl, indent) {\
61 fprintf(pandecode_dump_stream, "%s\n", title); \
62 struct MALI_ ## T temp; \
63 MALI_ ## T ## _unpack((const uint8_t *) cl, &temp); \
64 MALI_ ## T ## _print(pandecode_dump_stream, &temp, indent * 2); \
65 }
66
67 #define MAP_ADDR(T, addr, cl) \
68 const uint8_t *cl = 0; \
69 { \
70 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(addr); \
71 cl = pandecode_fetch_gpu_mem(mapped_mem, addr, MALI_ ## T ## _LENGTH); \
72 }
73
74 #define DUMP_ADDR(title, T, addr, indent) {\
75 MAP_ADDR(T, addr, cl) \
76 DUMP_CL(title, T, cl, indent); \
77 }
78
79 FILE *pandecode_dump_stream;
80
81 /* Semantic logging type.
82 *
83 * Raw: for raw messages to be printed as is.
84 * Message: for helpful information to be commented out in replays.
85 * Property: for properties of a struct
86 *
87 * Use one of pandecode_log, pandecode_msg, or pandecode_prop as syntax sugar.
88 */
89
90 enum pandecode_log_type {
91 PANDECODE_RAW,
92 PANDECODE_MESSAGE,
93 PANDECODE_PROPERTY
94 };
95
96 #define pandecode_log(...) pandecode_log_typed(PANDECODE_RAW, __VA_ARGS__)
97 #define pandecode_msg(...) pandecode_log_typed(PANDECODE_MESSAGE, __VA_ARGS__)
98 #define pandecode_prop(...) pandecode_log_typed(PANDECODE_PROPERTY, __VA_ARGS__)
99
100 unsigned pandecode_indent = 0;
101
102 static void
103 pandecode_make_indent(void)
104 {
105 for (unsigned i = 0; i < pandecode_indent; ++i)
106 fprintf(pandecode_dump_stream, " ");
107 }
108
109 static void
110 pandecode_log_typed(enum pandecode_log_type type, const char *format, ...)
111 {
112 va_list ap;
113
114 pandecode_make_indent();
115
116 if (type == PANDECODE_MESSAGE)
117 fprintf(pandecode_dump_stream, "// ");
118 else if (type == PANDECODE_PROPERTY)
119 fprintf(pandecode_dump_stream, ".");
120
121 va_start(ap, format);
122 vfprintf(pandecode_dump_stream, format, ap);
123 va_end(ap);
124
125 if (type == PANDECODE_PROPERTY)
126 fprintf(pandecode_dump_stream, ",\n");
127 }
128
129 static void
130 pandecode_log_cont(const char *format, ...)
131 {
132 va_list ap;
133
134 va_start(ap, format);
135 vfprintf(pandecode_dump_stream, format, ap);
136 va_end(ap);
137 }
138
139 /* To check for memory safety issues, validates that the given pointer in GPU
140 * memory is valid, containing at least sz bytes. The goal is to eliminate
141 * GPU-side memory bugs (NULL pointer dereferences, buffer overflows, or buffer
142 * overruns) by statically validating pointers.
143 */
144
145 static void
146 pandecode_validate_buffer(mali_ptr addr, size_t sz)
147 {
148 if (!addr) {
149 pandecode_msg("XXX: null pointer deref");
150 return;
151 }
152
153 /* Find a BO */
154
155 struct pandecode_mapped_memory *bo =
156 pandecode_find_mapped_gpu_mem_containing(addr);
157
158 if (!bo) {
159 pandecode_msg("XXX: invalid memory dereference\n");
160 return;
161 }
162
163 /* Bounds check */
164
165 unsigned offset = addr - bo->gpu_va;
166 unsigned total = offset + sz;
167
168 if (total > bo->length) {
169 pandecode_msg("XXX: buffer overrun. "
170 "Chunk of size %zu at offset %d in buffer of size %zu. "
171 "Overrun by %zu bytes. \n",
172 sz, offset, bo->length, total - bo->length);
173 return;
174 }
175 }
176
177 struct pandecode_flag_info {
178 u64 flag;
179 const char *name;
180 };
181
182 static void
183 pandecode_log_decoded_flags(const struct pandecode_flag_info *flag_info,
184 u64 flags)
185 {
186 bool decodable_flags_found = false;
187
188 for (int i = 0; flag_info[i].name; i++) {
189 if ((flags & flag_info[i].flag) != flag_info[i].flag)
190 continue;
191
192 if (!decodable_flags_found) {
193 decodable_flags_found = true;
194 } else {
195 pandecode_log_cont(" | ");
196 }
197
198 pandecode_log_cont("%s", flag_info[i].name);
199
200 flags &= ~flag_info[i].flag;
201 }
202
203 if (decodable_flags_found) {
204 if (flags)
205 pandecode_log_cont(" | 0x%" PRIx64, flags);
206 } else {
207 pandecode_log_cont("0x%" PRIx64, flags);
208 }
209 }
210
211 #define FLAG_INFO(flag) { MALI_CLEAR_##flag, "MALI_CLEAR_" #flag }
212 static const struct pandecode_flag_info clear_flag_info[] = {
213 FLAG_INFO(FAST),
214 FLAG_INFO(SLOW),
215 FLAG_INFO(SLOW_STENCIL),
216 {}
217 };
218 #undef FLAG_INFO
219
220 #define FLAG_INFO(flag) { MALI_MFBD_FORMAT_##flag, "MALI_MFBD_FORMAT_" #flag }
221 static const struct pandecode_flag_info mfbd_fmt_flag_info[] = {
222 FLAG_INFO(SRGB),
223 {}
224 };
225 #undef FLAG_INFO
226
227 #define FLAG_INFO(flag) { MALI_AFBC_##flag, "MALI_AFBC_" #flag }
228 static const struct pandecode_flag_info afbc_fmt_flag_info[] = {
229 FLAG_INFO(YTR),
230 {}
231 };
232 #undef FLAG_INFO
233
234 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
235 static const struct pandecode_flag_info mfbd_extra_flag_hi_info[] = {
236 FLAG_INFO(PRESENT),
237 {}
238 };
239 #undef FLAG_INFO
240
241 #define FLAG_INFO(flag) { MALI_EXTRA_##flag, "MALI_EXTRA_" #flag }
242 static const struct pandecode_flag_info mfbd_extra_flag_lo_info[] = {
243 FLAG_INFO(ZS),
244 {}
245 };
246 #undef FLAG_INFO
247
248 #define FLAG_INFO(flag) { MALI_MFBD_##flag, "MALI_MFBD_" #flag }
249 static const struct pandecode_flag_info mfbd_flag_info [] = {
250 FLAG_INFO(DEPTH_WRITE),
251 FLAG_INFO(EXTRA),
252 {}
253 };
254 #undef FLAG_INFO
255
256 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
257 static const struct pandecode_flag_info sfbd_unk1_info [] = {
258 FLAG_INFO(MSAA_8),
259 FLAG_INFO(MSAA_A),
260 {}
261 };
262 #undef FLAG_INFO
263
264 #define FLAG_INFO(flag) { MALI_SFBD_FORMAT_##flag, "MALI_SFBD_FORMAT_" #flag }
265 static const struct pandecode_flag_info sfbd_unk2_info [] = {
266 FLAG_INFO(MSAA_B),
267 FLAG_INFO(SRGB),
268 {}
269 };
270 #undef FLAG_INFO
271
272 /* Midgard's tiler descriptor is embedded within the
273 * larger FBD */
274
275 static void
276 pandecode_midgard_tiler_descriptor(
277 const struct midgard_tiler_descriptor *t,
278 unsigned width,
279 unsigned height,
280 bool is_fragment,
281 bool has_hierarchy)
282 {
283 pandecode_log(".tiler = {\n");
284 pandecode_indent++;
285
286 if (t->hierarchy_mask == MALI_TILER_DISABLED)
287 pandecode_prop("hierarchy_mask = MALI_TILER_DISABLED");
288 else
289 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
290
291 /* We know this name from the kernel, but we never see it nonzero */
292
293 if (t->flags)
294 pandecode_msg("XXX: unexpected tiler flags 0x%" PRIx16, t->flags);
295
296 MEMORY_PROP(t, polygon_list);
297
298 /* The body is offset from the base of the polygon list */
299 //assert(t->polygon_list_body > t->polygon_list);
300 unsigned body_offset = t->polygon_list_body - t->polygon_list;
301
302 /* It needs to fit inside the reported size */
303 //assert(t->polygon_list_size >= body_offset);
304
305 /* Now that we've sanity checked, we'll try to calculate the sizes
306 * ourselves for comparison */
307
308 unsigned ref_header = panfrost_tiler_header_size(width, height, t->hierarchy_mask, has_hierarchy);
309 unsigned ref_size = panfrost_tiler_full_size(width, height, t->hierarchy_mask, has_hierarchy);
310
311 if (!((ref_header == body_offset) && (ref_size == t->polygon_list_size))) {
312 pandecode_msg("XXX: bad polygon list size (expected %d / 0x%x)\n",
313 ref_header, ref_size);
314 pandecode_prop("polygon_list_size = 0x%x", t->polygon_list_size);
315 pandecode_msg("body offset %d\n", body_offset);
316 }
317
318 /* The tiler heap has a start and end specified -- it should be
319 * identical to what we have in the BO. The exception is if tiling is
320 * disabled. */
321
322 MEMORY_PROP(t, heap_start);
323 assert(t->heap_end >= t->heap_start);
324
325 unsigned heap_size = t->heap_end - t->heap_start;
326
327 /* Tiling is enabled with a special flag */
328 unsigned hierarchy_mask = t->hierarchy_mask & MALI_HIERARCHY_MASK;
329 unsigned tiler_flags = t->hierarchy_mask ^ hierarchy_mask;
330
331 bool tiling_enabled = hierarchy_mask;
332
333 if (tiling_enabled) {
334 /* We should also have no other flags */
335 if (tiler_flags)
336 pandecode_msg("XXX: unexpected tiler %X\n", tiler_flags);
337 } else {
338 /* When tiling is disabled, we should have that flag and no others */
339
340 if (tiler_flags != MALI_TILER_DISABLED) {
341 pandecode_msg("XXX: unexpected tiler flag %X, expected MALI_TILER_DISABLED\n",
342 tiler_flags);
343 }
344
345 /* We should also have an empty heap */
346 if (heap_size) {
347 pandecode_msg("XXX: tiler heap size %d given, expected empty\n",
348 heap_size);
349 }
350
351 /* Disabled tiling is used only for clear-only jobs, which are
352 * purely FRAGMENT, so we should never see this for
353 * non-FRAGMENT descriptors. */
354
355 if (!is_fragment)
356 pandecode_msg("XXX: tiler disabled for non-FRAGMENT job\n");
357 }
358
359 /* We've never seen weights used in practice, but we know from the
360 * kernel these fields is there */
361
362 bool nonzero_weights = false;
363
364 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
365 nonzero_weights |= t->weights[w] != 0x0;
366 }
367
368 if (nonzero_weights) {
369 pandecode_log(".weights = { ");
370
371 for (unsigned w = 0; w < ARRAY_SIZE(t->weights); ++w) {
372 pandecode_log_cont("%d, ", t->weights[w]);
373 }
374
375 pandecode_log("},");
376 }
377
378 pandecode_indent--;
379 pandecode_log("}\n");
380 }
381
382 /* TODO: The Bifrost tiler is not understood at all yet */
383
384 static void
385 pandecode_bifrost_tiler_descriptor(const struct mali_framebuffer *fb)
386 {
387 pandecode_log(".tiler = {\n");
388 pandecode_indent++;
389
390 MEMORY_PROP(fb, tiler_meta);
391
392 for (int i = 0; i < 16; i++) {
393 if (fb->zeros[i] != 0) {
394 pandecode_msg("XXX: tiler descriptor zero %d tripped, value %x\n",
395 i, fb->zeros[i]);
396 }
397 }
398
399 pandecode_log("},\n");
400
401 pandecode_indent--;
402 pandecode_log("}\n");
403
404 }
405
406 /* Information about the framebuffer passed back for
407 * additional analysis */
408
409 struct pandecode_fbd {
410 unsigned width;
411 unsigned height;
412 unsigned rt_count;
413 bool has_extra;
414 };
415
416 static void
417 pandecode_sfbd_format(struct mali_sfbd_format format)
418 {
419 pandecode_log(".format = {\n");
420 pandecode_indent++;
421
422 pandecode_log(".unk1 = ");
423 pandecode_log_decoded_flags(sfbd_unk1_info, format.unk1);
424 pandecode_log_cont(",\n");
425
426 /* TODO: Map formats so we can check swizzles and print nicely */
427 pandecode_log("swizzle");
428 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
429 pandecode_log_cont(",\n");
430
431 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
432 (format.nr_channels + 1));
433
434 pandecode_log(".unk2 = ");
435 pandecode_log_decoded_flags(sfbd_unk2_info, format.unk2);
436 pandecode_log_cont(",\n");
437
438 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
439
440 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
441
442 pandecode_indent--;
443 pandecode_log("},\n");
444 }
445
446 static void
447 pandecode_shared_memory(const struct mali_shared_memory *desc, bool is_compute)
448 {
449 pandecode_prop("stack_shift = 0x%x", desc->stack_shift);
450
451 if (desc->unk0)
452 pandecode_prop("unk0 = 0x%x", desc->unk0);
453
454 if (desc->shared_workgroup_count != 0x1F) {
455 pandecode_prop("shared_workgroup_count = %d", desc->shared_workgroup_count);
456 if (!is_compute)
457 pandecode_msg("XXX: wrong workgroup count for noncompute\n");
458 }
459
460 if (desc->shared_unk1 || desc->shared_shift) {
461 pandecode_prop("shared_unk1 = %X", desc->shared_unk1);
462 pandecode_prop("shared_shift = %X", desc->shared_shift);
463
464 if (!is_compute)
465 pandecode_msg("XXX: shared memory configured in noncompute shader");
466 }
467
468 if (desc->shared_zero) {
469 pandecode_msg("XXX: shared memory zero tripped\n");
470 pandecode_prop("shared_zero = 0x%" PRIx32, desc->shared_zero);
471 }
472
473 if (desc->shared_memory && !is_compute)
474 pandecode_msg("XXX: shared memory used in noncompute shader\n");
475
476 MEMORY_PROP(desc, scratchpad);
477 MEMORY_PROP(desc, shared_memory);
478 MEMORY_PROP(desc, unknown1);
479 }
480
481 static struct pandecode_fbd
482 pandecode_sfbd(uint64_t gpu_va, int job_no, bool is_fragment, unsigned gpu_id)
483 {
484 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
485 const struct mali_single_framebuffer *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
486
487 struct pandecode_fbd info = {
488 .has_extra = false,
489 .rt_count = 1
490 };
491
492 pandecode_log("struct mali_single_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
493 pandecode_indent++;
494
495 pandecode_log(".shared_memory = {\n");
496 pandecode_indent++;
497 pandecode_shared_memory(&s->shared_memory, false);
498 pandecode_indent--;
499 pandecode_log("},\n");
500
501 pandecode_sfbd_format(s->format);
502
503 info.width = s->width + 1;
504 info.height = s->height + 1;
505
506 pandecode_prop("width = MALI_POSITIVE(%" PRId16 ")", info.width);
507 pandecode_prop("height = MALI_POSITIVE(%" PRId16 ")", info.height);
508
509 MEMORY_PROP(s, checksum);
510
511 if (s->checksum_stride)
512 pandecode_prop("checksum_stride = %d", s->checksum_stride);
513
514 MEMORY_PROP(s, framebuffer);
515 pandecode_prop("stride = %d", s->stride);
516
517 /* Earlier in the actual commandstream -- right before width -- but we
518 * delay to flow nicer */
519
520 pandecode_log(".clear_flags = ");
521 pandecode_log_decoded_flags(clear_flag_info, s->clear_flags);
522 pandecode_log_cont(",\n");
523
524 if (s->depth_buffer) {
525 MEMORY_PROP(s, depth_buffer);
526 pandecode_prop("depth_stride = %d", s->depth_stride);
527 }
528
529 if (s->stencil_buffer) {
530 MEMORY_PROP(s, stencil_buffer);
531 pandecode_prop("stencil_stride = %d", s->stencil_stride);
532 }
533
534 if (s->depth_stride_zero ||
535 s->stencil_stride_zero ||
536 s->zero7 || s->zero8) {
537 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
538 pandecode_prop("depth_stride_zero = 0x%x",
539 s->depth_stride_zero);
540 pandecode_prop("stencil_stride_zero = 0x%x",
541 s->stencil_stride_zero);
542 pandecode_prop("zero7 = 0x%" PRIx32,
543 s->zero7);
544 pandecode_prop("zero8 = 0x%" PRIx32,
545 s->zero8);
546 }
547
548 if (s->clear_color_1 | s->clear_color_2 | s->clear_color_3 | s->clear_color_4) {
549 pandecode_prop("clear_color_1 = 0x%" PRIx32, s->clear_color_1);
550 pandecode_prop("clear_color_2 = 0x%" PRIx32, s->clear_color_2);
551 pandecode_prop("clear_color_3 = 0x%" PRIx32, s->clear_color_3);
552 pandecode_prop("clear_color_4 = 0x%" PRIx32, s->clear_color_4);
553 }
554
555 if (s->clear_depth_1 != 0 || s->clear_depth_2 != 0 || s->clear_depth_3 != 0 || s->clear_depth_4 != 0) {
556 pandecode_prop("clear_depth_1 = %f", s->clear_depth_1);
557 pandecode_prop("clear_depth_2 = %f", s->clear_depth_2);
558 pandecode_prop("clear_depth_3 = %f", s->clear_depth_3);
559 pandecode_prop("clear_depth_4 = %f", s->clear_depth_4);
560 }
561
562 if (s->clear_stencil) {
563 pandecode_prop("clear_stencil = 0x%x", s->clear_stencil);
564 }
565
566 const struct midgard_tiler_descriptor t = s->tiler;
567
568 bool has_hierarchy = !(gpu_id == 0x0720 || gpu_id == 0x0820 || gpu_id == 0x0830);
569 pandecode_midgard_tiler_descriptor(&t, s->width + 1, s->height + 1, is_fragment, has_hierarchy);
570
571 pandecode_indent--;
572 pandecode_log("};\n");
573
574 pandecode_prop("zero2 = 0x%" PRIx32, s->zero2);
575 pandecode_prop("zero4 = 0x%" PRIx32, s->zero4);
576 pandecode_prop("zero5 = 0x%" PRIx32, s->zero5);
577
578 pandecode_log_cont(".zero3 = {");
579
580 for (int i = 0; i < sizeof(s->zero3) / sizeof(s->zero3[0]); ++i)
581 pandecode_log_cont("%X, ", s->zero3[i]);
582
583 pandecode_log_cont("},\n");
584
585 pandecode_log_cont(".zero6 = {");
586
587 for (int i = 0; i < sizeof(s->zero6) / sizeof(s->zero6[0]); ++i)
588 pandecode_log_cont("%X, ", s->zero6[i]);
589
590 pandecode_log_cont("},\n");
591
592 return info;
593 }
594
595 static void
596 pandecode_compute_fbd(uint64_t gpu_va, int job_no)
597 {
598 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
599 const struct mali_shared_memory *PANDECODE_PTR_VAR(s, mem, (mali_ptr) gpu_va);
600
601 pandecode_log("struct mali_shared_memory shared_%"PRIx64"_%d = {\n", gpu_va, job_no);
602 pandecode_indent++;
603 pandecode_shared_memory(s, true);
604 pandecode_indent--;
605 pandecode_log("},\n");
606 }
607
608 /* Extracts the number of components associated with a Mali format */
609
610 static unsigned
611 pandecode_format_component_count(enum mali_format fmt)
612 {
613 /* Mask out the format class */
614 unsigned top = fmt & 0b11100000;
615
616 switch (top) {
617 case MALI_FORMAT_SNORM:
618 case MALI_FORMAT_UINT:
619 case MALI_FORMAT_UNORM:
620 case MALI_FORMAT_SINT:
621 return ((fmt >> 3) & 3) + 1;
622 default:
623 /* TODO: Validate */
624 return 4;
625 }
626 }
627
628 /* Extracts a mask of accessed components from a 12-bit Mali swizzle */
629
630 static unsigned
631 pandecode_access_mask_from_channel_swizzle(unsigned swizzle)
632 {
633 unsigned mask = 0;
634 assert(MALI_CHANNEL_R == 0);
635
636 for (unsigned c = 0; c < 4; ++c) {
637 enum mali_channel chan = (swizzle >> (3*c)) & 0x7;
638
639 if (chan <= MALI_CHANNEL_A)
640 mask |= (1 << chan);
641 }
642
643 return mask;
644 }
645
646 /* Validates that a (format, swizzle) pair is valid, in the sense that the
647 * swizzle doesn't access any components that are undefined in the format.
648 * Returns whether the swizzle is trivial (doesn't do any swizzling) and can be
649 * omitted */
650
651 static bool
652 pandecode_validate_format_swizzle(enum mali_format fmt, unsigned swizzle)
653 {
654 unsigned nr_comp = pandecode_format_component_count(fmt);
655 unsigned access_mask = pandecode_access_mask_from_channel_swizzle(swizzle);
656 unsigned valid_mask = (1 << nr_comp) - 1;
657 unsigned invalid_mask = ~valid_mask;
658
659 if (access_mask & invalid_mask) {
660 pandecode_msg("XXX: invalid components accessed\n");
661 return false;
662 }
663
664 /* Check for the default non-swizzling swizzle so we can suppress
665 * useless printing for the defaults */
666
667 unsigned default_swizzles[4] = {
668 MALI_CHANNEL_R | (MALI_CHANNEL_0 << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
669 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_0 << 6) | (MALI_CHANNEL_1 << 9),
670 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_1 << 9),
671 MALI_CHANNEL_R | (MALI_CHANNEL_G << 3) | (MALI_CHANNEL_B << 6) | (MALI_CHANNEL_A << 9)
672 };
673
674 return (swizzle == default_swizzles[nr_comp - 1]);
675 }
676
677 static void
678 pandecode_swizzle(unsigned swizzle, enum mali_format format)
679 {
680 /* First, do some validation */
681 bool trivial_swizzle = pandecode_validate_format_swizzle(
682 format, swizzle);
683
684 if (trivial_swizzle)
685 return;
686
687 /* Next, print the swizzle */
688 pandecode_log_cont(".");
689
690 static const char components[] = "rgba01";
691
692 for (unsigned c = 0; c < 4; ++c) {
693 enum mali_channel chan = (swizzle >> (3 * c)) & 0x7;
694
695 if (chan > MALI_CHANNEL_1) {
696 pandecode_log("XXX: invalid swizzle channel %d\n", chan);
697 continue;
698 }
699 pandecode_log_cont("%c", components[chan]);
700 }
701 }
702
703 static void
704 pandecode_rt_format(struct mali_rt_format format)
705 {
706 pandecode_log(".format = {\n");
707 pandecode_indent++;
708
709 pandecode_prop("unk1 = 0x%" PRIx32, format.unk1);
710 pandecode_prop("unk2 = 0x%" PRIx32, format.unk2);
711 pandecode_prop("unk3 = 0x%" PRIx32, format.unk3);
712 pandecode_prop("unk4 = 0x%" PRIx32, format.unk4);
713
714 pandecode_prop("block = %s", mali_block_format_as_str(format.block));
715
716 /* TODO: Map formats so we can check swizzles and print nicely */
717 pandecode_log("swizzle");
718 pandecode_swizzle(format.swizzle, MALI_RGBA8_UNORM);
719 pandecode_log_cont(",\n");
720
721 pandecode_prop("nr_channels = MALI_POSITIVE(%d)",
722 (format.nr_channels + 1));
723
724 pandecode_log(".flags = ");
725 pandecode_log_decoded_flags(mfbd_fmt_flag_info, format.flags);
726 pandecode_log_cont(",\n");
727
728 pandecode_prop("msaa = %s", mali_msaa_as_str(format.msaa));
729
730 /* In theory, the no_preload bit can be cleared to enable MFBD preload,
731 * which is a faster hardware-based alternative to the wallpaper method
732 * to preserve framebuffer contents across frames. In practice, MFBD
733 * preload is buggy on Midgard, and so this is a chicken bit. If this
734 * bit isn't set, most likely something broke unrelated to preload */
735
736 if (!format.no_preload) {
737 pandecode_msg("XXX: buggy MFBD preload enabled - chicken bit should be clear\n");
738 pandecode_prop("no_preload = 0x%" PRIx32, format.no_preload);
739 }
740
741 if (format.zero)
742 pandecode_prop("zero = 0x%" PRIx32, format.zero);
743
744 pandecode_indent--;
745 pandecode_log("},\n");
746 }
747
748 static void
749 pandecode_render_target(uint64_t gpu_va, unsigned job_no, const struct mali_framebuffer *fb)
750 {
751 pandecode_log("struct mali_render_target rts_list_%"PRIx64"_%d[] = {\n", gpu_va, job_no);
752 pandecode_indent++;
753
754 for (int i = 0; i < (fb->rt_count_1 + 1); i++) {
755 mali_ptr rt_va = gpu_va + i * sizeof(struct mali_render_target);
756 struct pandecode_mapped_memory *mem =
757 pandecode_find_mapped_gpu_mem_containing(rt_va);
758 const struct mali_render_target *PANDECODE_PTR_VAR(rt, mem, (mali_ptr) rt_va);
759
760 pandecode_log("{\n");
761 pandecode_indent++;
762
763 pandecode_rt_format(rt->format);
764
765 if (rt->format.block == MALI_BLOCK_FORMAT_AFBC) {
766 pandecode_log(".afbc = {\n");
767 pandecode_indent++;
768
769 char *a = pointer_as_memory_reference(rt->afbc.metadata);
770 pandecode_prop("metadata = %s", a);
771 free(a);
772
773 pandecode_prop("stride = %d", rt->afbc.stride);
774
775 pandecode_log(".flags = ");
776 pandecode_log_decoded_flags(afbc_fmt_flag_info, rt->afbc.flags);
777 pandecode_log_cont(",\n");
778
779 pandecode_indent--;
780 pandecode_log("},\n");
781 } else if (rt->afbc.metadata || rt->afbc.stride || rt->afbc.flags) {
782 pandecode_msg("XXX: AFBC disabled but AFBC field set (0x%lX, 0x%x, 0x%x)\n",
783 rt->afbc.metadata,
784 rt->afbc.stride,
785 rt->afbc.flags);
786 }
787
788 MEMORY_PROP(rt, framebuffer);
789 pandecode_prop("framebuffer_stride = %d", rt->framebuffer_stride);
790
791 if (rt->layer_stride)
792 pandecode_prop("layer_stride = %d", rt->layer_stride);
793
794 if (rt->clear_color_1 | rt->clear_color_2 | rt->clear_color_3 | rt->clear_color_4) {
795 pandecode_prop("clear_color_1 = 0x%" PRIx32, rt->clear_color_1);
796 pandecode_prop("clear_color_2 = 0x%" PRIx32, rt->clear_color_2);
797 pandecode_prop("clear_color_3 = 0x%" PRIx32, rt->clear_color_3);
798 pandecode_prop("clear_color_4 = 0x%" PRIx32, rt->clear_color_4);
799 }
800
801 if (rt->zero1 || rt->zero2) {
802 pandecode_msg("XXX: render target zeros tripped\n");
803 pandecode_prop("zero1 = 0x%" PRIx64, rt->zero1);
804 pandecode_prop("zero2 = 0x%" PRIx32, rt->zero2);
805 }
806
807 pandecode_indent--;
808 pandecode_log("},\n");
809 }
810
811 pandecode_indent--;
812 pandecode_log("};\n");
813 }
814
815 static struct pandecode_fbd
816 pandecode_mfbd_bfr(uint64_t gpu_va, int job_no, bool is_fragment, bool is_compute, bool is_bifrost)
817 {
818 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
819 const struct mali_framebuffer *PANDECODE_PTR_VAR(fb, mem, (mali_ptr) gpu_va);
820
821 struct pandecode_fbd info;
822
823 if (is_bifrost && fb->msaa.sample_locations) {
824 /* The blob stores all possible sample locations in a single buffer
825 * allocated on startup, and just switches the pointer when switching
826 * MSAA state. For now, we just put the data into the cmdstream, but we
827 * should do something like what the blob does with a real driver.
828 *
829 * There seem to be 32 slots for sample locations, followed by another
830 * 16. The second 16 is just the center location followed by 15 zeros
831 * in all the cases I've identified (maybe shader vs. depth/color
832 * samples?).
833 */
834
835 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(fb->msaa.sample_locations);
836
837 const u16 *PANDECODE_PTR_VAR(samples, smem, fb->msaa.sample_locations);
838
839 pandecode_log("uint16_t sample_locations_%d[] = {\n", job_no);
840 pandecode_indent++;
841
842 for (int i = 0; i < 32 + 16; i++) {
843 pandecode_log("%d, %d,\n", samples[2 * i], samples[2 * i + 1]);
844 }
845
846 pandecode_indent--;
847 pandecode_log("};\n");
848 }
849
850 pandecode_log("struct mali_framebuffer framebuffer_%"PRIx64"_%d = {\n", gpu_va, job_no);
851 pandecode_indent++;
852
853 if (is_bifrost) {
854 pandecode_log(".msaa = {\n");
855 pandecode_indent++;
856
857 if (fb->msaa.sample_locations)
858 pandecode_prop("sample_locations = sample_locations_%d", job_no);
859 else
860 pandecode_msg("XXX: sample_locations missing\n");
861
862 if (fb->msaa.zero1 || fb->msaa.zero2 || fb->msaa.zero4) {
863 pandecode_msg("XXX: multisampling zero tripped\n");
864 pandecode_prop("zero1 = %" PRIx64, fb->msaa.zero1);
865 pandecode_prop("zero2 = %" PRIx64, fb->msaa.zero2);
866 pandecode_prop("zero4 = %" PRIx64, fb->msaa.zero4);
867 }
868
869 pandecode_indent--;
870 pandecode_log("},\n");
871 } else {
872 pandecode_log(".shared_memory = {\n");
873 pandecode_indent++;
874 pandecode_shared_memory(&fb->shared_memory, is_compute);
875 pandecode_indent--;
876 pandecode_log("},\n");
877 }
878
879 info.width = fb->width1 + 1;
880 info.height = fb->height1 + 1;
881 info.rt_count = fb->rt_count_1 + 1;
882
883 pandecode_prop("width1 = MALI_POSITIVE(%d)", fb->width1 + 1);
884 pandecode_prop("height1 = MALI_POSITIVE(%d)", fb->height1 + 1);
885 pandecode_prop("width2 = MALI_POSITIVE(%d)", fb->width2 + 1);
886 pandecode_prop("height2 = MALI_POSITIVE(%d)", fb->height2 + 1);
887
888 pandecode_prop("unk1 = 0x%x", fb->unk1);
889 pandecode_prop("unk2 = 0x%x", fb->unk2);
890 pandecode_prop("rt_count_1 = MALI_POSITIVE(%d)", fb->rt_count_1 + 1);
891 pandecode_prop("rt_count_2 = %d", fb->rt_count_2);
892
893 pandecode_log(".mfbd_flags = ");
894 pandecode_log_decoded_flags(mfbd_flag_info, fb->mfbd_flags);
895 pandecode_log_cont(",\n");
896
897 if (fb->clear_stencil)
898 pandecode_prop("clear_stencil = 0x%x", fb->clear_stencil);
899
900 if (fb->clear_depth)
901 pandecode_prop("clear_depth = %f", fb->clear_depth);
902
903 if (!is_compute)
904 if (is_bifrost)
905 pandecode_bifrost_tiler_descriptor(fb);
906 else {
907 const struct midgard_tiler_descriptor t = fb->tiler;
908 pandecode_midgard_tiler_descriptor(&t, fb->width1 + 1, fb->height1 + 1, is_fragment, true);
909 }
910 else
911 pandecode_msg("XXX: skipping compute MFBD, fixme\n");
912
913 if (fb->zero3 || fb->zero4) {
914 pandecode_msg("XXX: framebuffer zeros tripped\n");
915 pandecode_prop("zero3 = 0x%" PRIx32, fb->zero3);
916 pandecode_prop("zero4 = 0x%" PRIx32, fb->zero4);
917 }
918
919 pandecode_indent--;
920 pandecode_log("};\n");
921
922 gpu_va += sizeof(struct mali_framebuffer);
923
924 info.has_extra = (fb->mfbd_flags & MALI_MFBD_EXTRA) && is_fragment;
925
926 if (info.has_extra) {
927 mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
928 const struct mali_framebuffer_extra *PANDECODE_PTR_VAR(fbx, mem, (mali_ptr) gpu_va);
929
930 pandecode_log("struct mali_framebuffer_extra fb_extra_%"PRIx64"_%d = {\n", gpu_va, job_no);
931 pandecode_indent++;
932
933 MEMORY_PROP(fbx, checksum);
934
935 if (fbx->checksum_stride)
936 pandecode_prop("checksum_stride = %d", fbx->checksum_stride);
937
938 pandecode_log(".flags_hi = ");
939 pandecode_log_decoded_flags(mfbd_extra_flag_hi_info, fbx->flags_hi);
940 pandecode_log_cont(",\n");
941
942 pandecode_log(".flags_lo = ");
943 pandecode_log_decoded_flags(mfbd_extra_flag_lo_info, fbx->flags_lo);
944 pandecode_log_cont(",\n");
945
946 pandecode_prop("zs_block = %s", mali_block_format_as_str(fbx->zs_block));
947 pandecode_prop("zs_samples = MALI_POSITIVE(%u)", fbx->zs_samples + 1);
948
949 if (fbx->zs_block == MALI_BLOCK_FORMAT_AFBC) {
950 pandecode_log(".ds_afbc = {\n");
951 pandecode_indent++;
952
953 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil_afbc_metadata);
954 pandecode_prop("depth_stencil_afbc_stride = %d",
955 fbx->ds_afbc.depth_stencil_afbc_stride);
956 MEMORY_PROP_DIR(fbx->ds_afbc, depth_stencil);
957
958 pandecode_log(".flags = ");
959 pandecode_log_decoded_flags(afbc_fmt_flag_info, fbx->ds_afbc.flags);
960 pandecode_log_cont(",\n");
961
962 if (fbx->ds_afbc.padding) {
963 pandecode_msg("XXX: Depth/stencil AFBC zeros tripped\n");
964 pandecode_prop("padding = 0x%" PRIx64, fbx->ds_afbc.padding);
965 }
966
967 pandecode_indent--;
968 pandecode_log("},\n");
969 } else {
970 pandecode_log(".ds_linear = {\n");
971 pandecode_indent++;
972
973 if (fbx->ds_linear.depth) {
974 MEMORY_PROP_DIR(fbx->ds_linear, depth);
975 pandecode_prop("depth_stride = %d",
976 fbx->ds_linear.depth_stride);
977 pandecode_prop("depth_layer_stride = %d",
978 fbx->ds_linear.depth_layer_stride);
979 } else if (fbx->ds_linear.depth_stride || fbx->ds_linear.depth_layer_stride) {
980 pandecode_msg("XXX: depth stride zero tripped %d %d\n", fbx->ds_linear.depth_stride, fbx->ds_linear.depth_layer_stride);
981 }
982
983 if (fbx->ds_linear.stencil) {
984 MEMORY_PROP_DIR(fbx->ds_linear, stencil);
985 pandecode_prop("stencil_stride = %d",
986 fbx->ds_linear.stencil_stride);
987 pandecode_prop("stencil_layer_stride = %d",
988 fbx->ds_linear.stencil_layer_stride);
989 } else if (fbx->ds_linear.stencil_stride || fbx->ds_linear.stencil_layer_stride) {
990 pandecode_msg("XXX: stencil stride zero tripped %d %d\n", fbx->ds_linear.stencil_stride, fbx->ds_linear.stencil_layer_stride);
991 }
992
993 if (fbx->ds_linear.depth_stride_zero ||
994 fbx->ds_linear.stencil_stride_zero) {
995 pandecode_msg("XXX: Depth/stencil zeros tripped\n");
996 pandecode_prop("depth_stride_zero = 0x%x",
997 fbx->ds_linear.depth_stride_zero);
998 pandecode_prop("stencil_stride_zero = 0x%x",
999 fbx->ds_linear.stencil_stride_zero);
1000 }
1001
1002 pandecode_indent--;
1003 pandecode_log("},\n");
1004 }
1005
1006 if (fbx->clear_color_1 | fbx->clear_color_2) {
1007 pandecode_prop("clear_color_1 = 0x%" PRIx32, fbx->clear_color_1);
1008 pandecode_prop("clear_color_2 = 0x%" PRIx32, fbx->clear_color_2);
1009 }
1010
1011 if (fbx->zero3) {
1012 pandecode_msg("XXX: fb_extra zeros tripped\n");
1013 pandecode_prop("zero3 = 0x%" PRIx64, fbx->zero3);
1014 }
1015
1016 pandecode_indent--;
1017 pandecode_log("};\n");
1018
1019 gpu_va += sizeof(struct mali_framebuffer_extra);
1020 }
1021
1022 if (is_fragment)
1023 pandecode_render_target(gpu_va, job_no, fb);
1024
1025 return info;
1026 }
1027
1028 static void
1029 pandecode_attributes(const struct pandecode_mapped_memory *mem,
1030 mali_ptr addr, int job_no, char *suffix,
1031 int count, bool varying, enum mali_job_type job_type)
1032 {
1033 char *prefix = varying ? "Varying" : "Attribute";
1034 assert(addr);
1035
1036 if (!count) {
1037 pandecode_msg("warn: No %s records\n", prefix);
1038 return;
1039 }
1040
1041 MAP_ADDR(ATTRIBUTE_BUFFER, addr, cl);
1042
1043 for (int i = 0; i < count; ++i) {
1044 fprintf(pandecode_dump_stream, "%s\n", prefix);
1045
1046 struct MALI_ATTRIBUTE_BUFFER temp;
1047 MALI_ATTRIBUTE_BUFFER_unpack(cl + i * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp);
1048 MALI_ATTRIBUTE_BUFFER_print(pandecode_dump_stream, &temp, 2);
1049
1050 if (temp.type == MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR) {
1051 struct MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT temp2;
1052 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_unpack(cl + (i + 1) * MALI_ATTRIBUTE_BUFFER_LENGTH, &temp2);
1053 MALI_ATTRIBUTE_BUFFER_CONTINUATION_NPOT_print(pandecode_dump_stream, &temp2, 2);
1054 }
1055 }
1056 }
1057
1058 static mali_ptr
1059 pandecode_shader_address(const char *name, mali_ptr ptr)
1060 {
1061 /* TODO: Decode flags */
1062 mali_ptr shader_ptr = ptr & ~15;
1063
1064 char *a = pointer_as_memory_reference(shader_ptr);
1065 pandecode_prop("%s = (%s) | %d", name, a, (int) (ptr & 15));
1066 free(a);
1067
1068 return shader_ptr;
1069 }
1070
1071 /* Decodes a Bifrost blend constant. See the notes in bifrost_blend_rt */
1072
1073 static unsigned
1074 decode_bifrost_constant(u16 constant)
1075 {
1076 float lo = (float) (constant & 0xFF);
1077 float hi = (float) (constant >> 8);
1078
1079 return (hi / 255.0) + (lo / 65535.0);
1080 }
1081
1082 static mali_ptr
1083 pandecode_bifrost_blend(void *descs, int job_no, int rt_no)
1084 {
1085 struct bifrost_blend_rt *b =
1086 ((struct bifrost_blend_rt *) descs) + rt_no;
1087
1088 pandecode_log("struct bifrost_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1089 pandecode_indent++;
1090
1091 pandecode_prop("flags = 0x%" PRIx16, b->flags);
1092 pandecode_prop("constant = 0x%" PRIx8 " /* %f */",
1093 b->constant, decode_bifrost_constant(b->constant));
1094
1095 /* TODO figure out blend shader enable bit */
1096 DUMP_CL("Equation", BLEND_EQUATION, &b->equation, 2);
1097
1098 pandecode_prop("unk2 = 0x%" PRIx16, b->unk2);
1099 pandecode_prop("index = 0x%" PRIx16, b->index);
1100
1101 pandecode_log(".format = %s", mali_format_as_str(b->format));
1102 pandecode_swizzle(b->swizzle, b->format);
1103 pandecode_log_cont(",\n");
1104
1105 pandecode_prop("swizzle = 0x%" PRIx32, b->swizzle);
1106 pandecode_prop("format = 0x%" PRIx32, b->format);
1107
1108 if (b->zero1) {
1109 pandecode_msg("XXX: pandecode_bifrost_blend zero1 tripped\n");
1110 pandecode_prop("zero1 = 0x%" PRIx32, b->zero1);
1111 }
1112
1113 pandecode_log(".shader_type = ");
1114 switch(b->shader_type) {
1115 case BIFROST_BLEND_F16:
1116 pandecode_log_cont("BIFROST_BLEND_F16");
1117 break;
1118 case BIFROST_BLEND_F32:
1119 pandecode_log_cont("BIFROST_BLEND_F32");
1120 break;
1121 case BIFROST_BLEND_I32:
1122 pandecode_log_cont("BIFROST_BLEND_I32");
1123 break;
1124 case BIFROST_BLEND_U32:
1125 pandecode_log_cont("BIFROST_BLEND_U32");
1126 break;
1127 case BIFROST_BLEND_I16:
1128 pandecode_log_cont("BIFROST_BLEND_I16");
1129 break;
1130 case BIFROST_BLEND_U16:
1131 pandecode_log_cont("BIFROST_BLEND_U16");
1132 break;
1133 }
1134 pandecode_log_cont(",\n");
1135
1136 if (b->zero2) {
1137 pandecode_msg("XXX: pandecode_bifrost_blend zero2 tripped\n");
1138 pandecode_prop("zero2 = 0x%" PRIx32, b->zero2);
1139 }
1140
1141 pandecode_prop("shader = 0x%" PRIx32, b->shader);
1142
1143 pandecode_indent--;
1144 pandecode_log("},\n");
1145
1146 return 0;
1147 }
1148
1149 static mali_ptr
1150 pandecode_midgard_blend(union midgard_blend *blend, bool is_shader)
1151 {
1152 /* constant/equation is in a union */
1153 if (!blend->shader)
1154 return 0;
1155
1156 pandecode_log(".blend = {\n");
1157 pandecode_indent++;
1158
1159 if (is_shader) {
1160 pandecode_shader_address("shader", blend->shader);
1161 } else {
1162 DUMP_CL("Equation", BLEND_EQUATION, &blend->equation, 2);
1163 pandecode_prop("constant = %f", blend->constant);
1164 }
1165
1166 pandecode_indent--;
1167 pandecode_log("},\n");
1168
1169 /* Return blend shader to disassemble if present */
1170 return is_shader ? (blend->shader & ~0xF) : 0;
1171 }
1172
1173 static mali_ptr
1174 pandecode_midgard_blend_mrt(void *descs, int job_no, int rt_no)
1175 {
1176 struct midgard_blend_rt *b =
1177 ((struct midgard_blend_rt *) descs) + rt_no;
1178
1179 /* Flags determine presence of blend shader */
1180 bool is_shader = b->flags.opaque[0] & 0x2;
1181
1182 pandecode_log("struct midgard_blend_rt blend_rt_%d_%d = {\n", job_no, rt_no);
1183 pandecode_indent++;
1184
1185 DUMP_CL("Flags", BLEND_FLAGS, &b->flags, 2);
1186
1187 union midgard_blend blend = b->blend;
1188 mali_ptr shader = pandecode_midgard_blend(&blend, is_shader);
1189
1190 pandecode_indent--;
1191 pandecode_log("};\n");
1192
1193 return shader;
1194 }
1195
1196 /* Attributes and varyings have descriptor records, which contain information
1197 * about their format and ordering with the attribute/varying buffers. We'll
1198 * want to validate that the combinations specified are self-consistent.
1199 */
1200
1201 static int
1202 pandecode_attribute_meta(int count, mali_ptr attribute, bool varying, char *suffix)
1203 {
1204 const char *prefix = varying ? "Varying" : "Attribute";
1205
1206 for (int i = 0; i < count; ++i, attribute += MALI_ATTRIBUTE_LENGTH)
1207 DUMP_ADDR(prefix, ATTRIBUTE, attribute, 1);
1208
1209 return count;
1210 }
1211
1212 /* return bits [lo, hi) of word */
1213 static u32
1214 bits(u32 word, u32 lo, u32 hi)
1215 {
1216 if (hi - lo >= 32)
1217 return word; // avoid undefined behavior with the shift
1218
1219 return (word >> lo) & ((1 << (hi - lo)) - 1);
1220 }
1221
1222 static void
1223 pandecode_vertex_tiler_prefix(struct mali_vertex_tiler_prefix *p, int job_no, bool graphics)
1224 {
1225 pandecode_log(".prefix = {\n");
1226 pandecode_indent++;
1227
1228 /* Decode invocation_count. See the comment before the definition of
1229 * invocation_count for an explanation.
1230 */
1231 struct MALI_INVOCATION invocation;
1232 struct mali_invocation_packed invocation_packed = p->invocation;
1233 MALI_INVOCATION_unpack((const uint8_t *) &invocation_packed, &invocation);
1234
1235 unsigned size_x = bits(invocation.invocations, 0, invocation.size_y_shift) + 1;
1236 unsigned size_y = bits(invocation.invocations, invocation.size_y_shift, invocation.size_z_shift) + 1;
1237 unsigned size_z = bits(invocation.invocations, invocation.size_z_shift, invocation.workgroups_x_shift) + 1;
1238
1239 unsigned groups_x = bits(invocation.invocations, invocation.workgroups_x_shift, invocation.workgroups_y_shift) + 1;
1240 unsigned groups_y = bits(invocation.invocations, invocation.workgroups_y_shift, invocation.workgroups_z_shift) + 1;
1241 unsigned groups_z = bits(invocation.invocations, invocation.workgroups_z_shift, 32) + 1;
1242
1243 /* Even though we have this decoded, we want to ensure that the
1244 * representation is "unique" so we don't lose anything by printing only
1245 * the final result. More specifically, we need to check that we were
1246 * passed something in canonical form, since the definition per the
1247 * hardware is inherently not unique. How? Well, take the resulting
1248 * decode and pack it ourselves! If it is bit exact with what we
1249 * decoded, we're good to go. */
1250
1251 struct mali_invocation_packed ref;
1252 panfrost_pack_work_groups_compute(&ref, groups_x, groups_y, groups_z, size_x, size_y, size_z, graphics);
1253
1254 if (memcmp(&ref, &invocation_packed, sizeof(ref))) {
1255 pandecode_msg("XXX: non-canonical workgroups packing\n");
1256 MALI_INVOCATION_print(pandecode_dump_stream, &invocation, 1 * 2);
1257 }
1258
1259 /* Regardless, print the decode */
1260 fprintf(pandecode_dump_stream,
1261 "Invocation (%d, %d, %d) x (%d, %d, %d)\n",
1262 size_x, size_y, size_z,
1263 groups_x, groups_y, groups_z);
1264
1265 fprintf(pandecode_dump_stream, "Primitive\n");
1266 struct MALI_PRIMITIVE primitive;
1267 struct mali_primitive_packed prim_packed = p->primitive;
1268 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed, &primitive);
1269 MALI_PRIMITIVE_print(pandecode_dump_stream, &primitive, 1 * 2);
1270
1271 /* Validate an index buffer is present if we need one. TODO: verify
1272 * relationship between invocation_count and index_count */
1273
1274 if (primitive.indices) {
1275 /* Grab the size */
1276 unsigned size = (primitive.index_type == MALI_INDEX_TYPE_UINT32) ?
1277 sizeof(uint32_t) : primitive.index_type;
1278
1279 /* Ensure we got a size, and if so, validate the index buffer
1280 * is large enough to hold a full set of indices of the given
1281 * size */
1282
1283 if (!size)
1284 pandecode_msg("XXX: index size missing\n");
1285 else
1286 pandecode_validate_buffer(primitive.indices, primitive.index_count * size);
1287 } else if (primitive.index_type)
1288 pandecode_msg("XXX: unexpected index size\n");
1289
1290 pandecode_indent--;
1291 pandecode_log("},\n");
1292 }
1293
1294 static void
1295 pandecode_uniform_buffers(mali_ptr pubufs, int ubufs_count, int job_no)
1296 {
1297 struct pandecode_mapped_memory *umem = pandecode_find_mapped_gpu_mem_containing(pubufs);
1298 uint64_t *PANDECODE_PTR_VAR(ubufs, umem, pubufs);
1299
1300 for (int i = 0; i < ubufs_count; i++) {
1301 unsigned size = (ubufs[i] & ((1 << 10) - 1)) * 16;
1302 mali_ptr addr = (ubufs[i] >> 10) << 2;
1303
1304 pandecode_validate_buffer(addr, size);
1305
1306 char *ptr = pointer_as_memory_reference(addr);
1307 pandecode_log("ubuf_%d[%u] = %s;\n", i, size, ptr);
1308 free(ptr);
1309 }
1310
1311 pandecode_log("\n");
1312 }
1313
1314 static void
1315 pandecode_uniforms(mali_ptr uniforms, unsigned uniform_count)
1316 {
1317 pandecode_validate_buffer(uniforms, uniform_count * 16);
1318
1319 char *ptr = pointer_as_memory_reference(uniforms);
1320 pandecode_log("vec4 uniforms[%u] = %s;\n", uniform_count, ptr);
1321 free(ptr);
1322 }
1323
1324 static const char *
1325 shader_type_for_job(unsigned type)
1326 {
1327 switch (type) {
1328 case MALI_JOB_TYPE_VERTEX: return "VERTEX";
1329 case MALI_JOB_TYPE_TILER: return "FRAGMENT";
1330 case MALI_JOB_TYPE_COMPUTE: return "COMPUTE";
1331 default: return "UNKNOWN";
1332 }
1333 }
1334
1335 static unsigned shader_id = 0;
1336
1337 static struct midgard_disasm_stats
1338 pandecode_shader_disassemble(mali_ptr shader_ptr, int shader_no, int type,
1339 bool is_bifrost, unsigned gpu_id)
1340 {
1341 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(shader_ptr);
1342 uint8_t *PANDECODE_PTR_VAR(code, mem, shader_ptr);
1343
1344 /* Compute maximum possible size */
1345 size_t sz = mem->length - (shader_ptr - mem->gpu_va);
1346
1347 /* Print some boilerplate to clearly denote the assembly (which doesn't
1348 * obey indentation rules), and actually do the disassembly! */
1349
1350 pandecode_log_cont("\n\n");
1351
1352 struct midgard_disasm_stats stats;
1353
1354 if (is_bifrost) {
1355 disassemble_bifrost(pandecode_dump_stream, code, sz, true);
1356
1357 /* TODO: Extend stats to Bifrost */
1358 stats.texture_count = -128;
1359 stats.sampler_count = -128;
1360 stats.attribute_count = -128;
1361 stats.varying_count = -128;
1362 stats.uniform_count = -128;
1363 stats.uniform_buffer_count = -128;
1364 stats.work_count = -128;
1365
1366 stats.instruction_count = 0;
1367 stats.bundle_count = 0;
1368 stats.quadword_count = 0;
1369 stats.helper_invocations = false;
1370 } else {
1371 stats = disassemble_midgard(pandecode_dump_stream,
1372 code, sz, gpu_id,
1373 type == MALI_JOB_TYPE_TILER ?
1374 MESA_SHADER_FRAGMENT : MESA_SHADER_VERTEX);
1375 }
1376
1377 unsigned nr_threads =
1378 (stats.work_count <= 4) ? 4 :
1379 (stats.work_count <= 8) ? 2 :
1380 1;
1381
1382 pandecode_log_cont("shader%d - MESA_SHADER_%s shader: "
1383 "%u inst, %u bundles, %u quadwords, "
1384 "%u registers, %u threads, 0 loops, 0:0 spills:fills\n\n\n",
1385 shader_id++,
1386 shader_type_for_job(type),
1387 stats.instruction_count, stats.bundle_count, stats.quadword_count,
1388 stats.work_count, nr_threads);
1389
1390 return stats;
1391 }
1392
1393 static void
1394 pandecode_texture_payload(mali_ptr payload,
1395 enum mali_texture_dimension dim,
1396 enum mali_texture_layout layout,
1397 bool manual_stride,
1398 uint8_t levels,
1399 uint16_t depth,
1400 uint16_t array_size,
1401 struct pandecode_mapped_memory *tmem)
1402 {
1403 pandecode_log(".payload = {\n");
1404 pandecode_indent++;
1405
1406 /* A bunch of bitmap pointers follow.
1407 * We work out the correct number,
1408 * based on the mipmap/cubemap
1409 * properties, but dump extra
1410 * possibilities to futureproof */
1411
1412 int bitmap_count = levels + 1;
1413
1414 /* Miptree for each face */
1415 if (dim == MALI_TEXTURE_DIMENSION_CUBE)
1416 bitmap_count *= 6;
1417
1418 /* Array of layers */
1419 bitmap_count *= depth;
1420
1421 /* Array of textures */
1422 bitmap_count *= array_size;
1423
1424 /* Stride for each element */
1425 if (manual_stride)
1426 bitmap_count *= 2;
1427
1428 mali_ptr *pointers_and_strides = pandecode_fetch_gpu_mem(tmem,
1429 payload, sizeof(mali_ptr) * bitmap_count);
1430 for (int i = 0; i < bitmap_count; ++i) {
1431 /* How we dump depends if this is a stride or a pointer */
1432
1433 if (manual_stride && (i & 1)) {
1434 /* signed 32-bit snuck in as a 64-bit pointer */
1435 uint64_t stride_set = pointers_and_strides[i];
1436 uint32_t clamped_stride = stride_set;
1437 int32_t stride = clamped_stride;
1438 assert(stride_set == clamped_stride);
1439 pandecode_log("(mali_ptr) %d /* stride */, \n", stride);
1440 } else {
1441 char *a = pointer_as_memory_reference(pointers_and_strides[i]);
1442 pandecode_log("%s, \n", a);
1443 free(a);
1444 }
1445 }
1446
1447 pandecode_indent--;
1448 pandecode_log("},\n");
1449 }
1450
1451 static void
1452 pandecode_texture(mali_ptr u,
1453 struct pandecode_mapped_memory *tmem,
1454 unsigned job_no, unsigned tex)
1455 {
1456 struct pandecode_mapped_memory *mapped_mem = pandecode_find_mapped_gpu_mem_containing(u);
1457 const uint8_t *cl = pandecode_fetch_gpu_mem(mapped_mem, u, MALI_MIDGARD_TEXTURE_LENGTH);
1458
1459 struct MALI_MIDGARD_TEXTURE temp;
1460 MALI_MIDGARD_TEXTURE_unpack(cl, &temp);
1461 MALI_MIDGARD_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1462
1463 pandecode_texture_payload(u + MALI_MIDGARD_TEXTURE_LENGTH,
1464 temp.dimension, temp.texel_ordering, temp.manual_stride,
1465 temp.levels, temp.depth, temp.array_size, mapped_mem);
1466 }
1467
1468 static void
1469 pandecode_bifrost_texture(
1470 const void *cl,
1471 unsigned job_no,
1472 unsigned tex)
1473 {
1474 struct MALI_BIFROST_TEXTURE temp;
1475 MALI_BIFROST_TEXTURE_unpack(cl, &temp);
1476 MALI_BIFROST_TEXTURE_print(pandecode_dump_stream, &temp, 2);
1477
1478 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(temp.surfaces);
1479 pandecode_texture_payload(temp.surfaces, temp.dimension, temp.texel_ordering,
1480 true, temp.levels, 1, 1, tmem);
1481 }
1482
1483 /* 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 */
1484
1485 static void
1486 pandecode_shader_prop(const char *name, unsigned claim, signed truth, bool fuzzy)
1487 {
1488 /* Nothing to do */
1489 if (claim == truth)
1490 return;
1491
1492 if (fuzzy && (truth < 0))
1493 pandecode_msg("XXX: fuzzy %s, claimed %d, expected %d\n", name, claim, truth);
1494
1495 if ((truth >= 0) && !fuzzy) {
1496 pandecode_msg("%s: expected %s = %d, claimed %u\n",
1497 (truth < claim) ? "warn" : "XXX",
1498 name, truth, claim);
1499 } else if ((claim > -truth) && !fuzzy) {
1500 pandecode_msg("XXX: expected %s <= %u, claimed %u\n",
1501 name, -truth, claim);
1502 } else if (fuzzy && (claim < truth))
1503 pandecode_msg("XXX: expected %s >= %u, claimed %u\n",
1504 name, truth, claim);
1505
1506 pandecode_log(".%s = %" PRId16, name, claim);
1507
1508 if (fuzzy)
1509 pandecode_log_cont(" /* %u used */", truth);
1510
1511 pandecode_log_cont(",\n");
1512 }
1513
1514 static void
1515 pandecode_blend_shader_disassemble(mali_ptr shader, int job_no, int job_type,
1516 bool is_bifrost, unsigned gpu_id)
1517 {
1518 struct midgard_disasm_stats stats =
1519 pandecode_shader_disassemble(shader, job_no, job_type, is_bifrost, gpu_id);
1520
1521 bool has_texture = (stats.texture_count > 0);
1522 bool has_sampler = (stats.sampler_count > 0);
1523 bool has_attribute = (stats.attribute_count > 0);
1524 bool has_varying = (stats.varying_count > 0);
1525 bool has_uniform = (stats.uniform_count > 0);
1526 bool has_ubo = (stats.uniform_buffer_count > 0);
1527
1528 if (has_texture || has_sampler)
1529 pandecode_msg("XXX: blend shader accessing textures\n");
1530
1531 if (has_attribute || has_varying)
1532 pandecode_msg("XXX: blend shader accessing interstage\n");
1533
1534 if (has_uniform || has_ubo)
1535 pandecode_msg("XXX: blend shader accessing uniforms\n");
1536 }
1537
1538 static void
1539 pandecode_textures(mali_ptr textures, unsigned texture_count, int job_no, bool is_bifrost)
1540 {
1541 struct pandecode_mapped_memory *mmem = pandecode_find_mapped_gpu_mem_containing(textures);
1542
1543 if (!mmem)
1544 return;
1545
1546 pandecode_log("Textures (%"PRIx64"):\n", textures);
1547
1548 if (is_bifrost) {
1549 const void *cl = pandecode_fetch_gpu_mem(mmem,
1550 textures, MALI_BIFROST_TEXTURE_LENGTH *
1551 texture_count);
1552
1553 for (unsigned tex = 0; tex < texture_count; ++tex) {
1554 pandecode_bifrost_texture(cl +
1555 MALI_BIFROST_TEXTURE_LENGTH * tex,
1556 job_no, tex);
1557 }
1558 } else {
1559 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures);
1560
1561 for (int tex = 0; tex < texture_count; ++tex) {
1562 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1563 char *a = pointer_as_memory_reference(*u);
1564 pandecode_log("%s,\n", a);
1565 free(a);
1566 }
1567
1568 /* Now, finally, descend down into the texture descriptor */
1569 for (unsigned tex = 0; tex < texture_count; ++tex) {
1570 mali_ptr *PANDECODE_PTR_VAR(u, mmem, textures + tex * sizeof(mali_ptr));
1571 struct pandecode_mapped_memory *tmem = pandecode_find_mapped_gpu_mem_containing(*u);
1572 if (tmem)
1573 pandecode_texture(*u, tmem, job_no, tex);
1574 }
1575 }
1576 }
1577
1578 static void
1579 pandecode_samplers(mali_ptr samplers, unsigned sampler_count, int job_no, bool is_bifrost)
1580 {
1581 for (int i = 0; i < sampler_count; ++i) {
1582 if (is_bifrost) {
1583 DUMP_ADDR("Sampler", BIFROST_SAMPLER, samplers + (MALI_BIFROST_SAMPLER_LENGTH * i), 1);
1584 } else {
1585 DUMP_ADDR("Sampler", MIDGARD_SAMPLER, samplers + (MALI_MIDGARD_SAMPLER_LENGTH * i), 1);
1586 }
1587 }
1588 }
1589
1590 static void
1591 pandecode_vertex_tiler_postfix_pre(
1592 const struct MALI_DRAW *p,
1593 int job_no, enum mali_job_type job_type,
1594 char *suffix, bool is_bifrost, unsigned gpu_id)
1595 {
1596 struct pandecode_mapped_memory *attr_mem;
1597
1598 struct pandecode_fbd fbd_info = {
1599 /* Default for Bifrost */
1600 .rt_count = 1
1601 };
1602
1603 if (is_bifrost)
1604 pandecode_compute_fbd(p->shared & ~1, job_no);
1605 else if (p->shared & MALI_MFBD)
1606 fbd_info = pandecode_mfbd_bfr((u64) ((uintptr_t) p->shared) & FBD_MASK, job_no, false, job_type == MALI_JOB_TYPE_COMPUTE, false);
1607 else if (job_type == MALI_JOB_TYPE_COMPUTE)
1608 pandecode_compute_fbd((u64) (uintptr_t) p->shared, job_no);
1609 else
1610 fbd_info = pandecode_sfbd((u64) (uintptr_t) p->shared, job_no, false, gpu_id);
1611
1612 int varying_count = 0, attribute_count = 0, uniform_count = 0, uniform_buffer_count = 0;
1613 int texture_count = 0, sampler_count = 0;
1614
1615 if (p->state) {
1616 struct pandecode_mapped_memory *smem = pandecode_find_mapped_gpu_mem_containing(p->state);
1617 uint32_t *cl = pandecode_fetch_gpu_mem(smem, p->state, MALI_STATE_LENGTH);
1618
1619 /* Disassemble ahead-of-time to get stats. Initialize with
1620 * stats for the missing-shader case so we get validation
1621 * there, too */
1622
1623 struct midgard_disasm_stats info = {
1624 .texture_count = 0,
1625 .sampler_count = 0,
1626 .attribute_count = 0,
1627 .varying_count = 0,
1628 .work_count = 1,
1629
1630 .uniform_count = -128,
1631 .uniform_buffer_count = 0
1632 };
1633
1634 struct MALI_STATE state;
1635 struct MALI_MIDGARD_PROPERTIES midg_props;
1636 struct MALI_BIFROST_PROPERTIES bi_props;
1637
1638 MALI_STATE_unpack((const uint8_t *) cl, &state);
1639
1640 if (state.shader.shader & ~0xF)
1641 info = pandecode_shader_disassemble(state.shader.shader & ~0xF, job_no, job_type, is_bifrost, gpu_id);
1642
1643 fprintf(pandecode_dump_stream, "State");
1644 MALI_STATE_print(pandecode_dump_stream, &state, 1 * 2);
1645
1646 /* Save for dumps */
1647 attribute_count = state.shader.attribute_count;
1648 varying_count = state.shader.varying_count;
1649 texture_count = state.shader.texture_count;
1650 sampler_count = state.shader.sampler_count;
1651
1652 fprintf(pandecode_dump_stream, " Properties\n");
1653 if (is_bifrost) {
1654 MALI_BIFROST_PROPERTIES_unpack((const uint8_t *) &state.properties, &bi_props);
1655 MALI_BIFROST_PROPERTIES_print(pandecode_dump_stream, &bi_props, 2 * 2);
1656
1657 uniform_count = state.preload.uniform_count;
1658 uniform_buffer_count = bi_props.uniform_buffer_count;
1659 } else {
1660 MALI_MIDGARD_PROPERTIES_unpack((const uint8_t *) &state.properties, &midg_props);
1661 MALI_MIDGARD_PROPERTIES_print(pandecode_dump_stream, &midg_props, 2 * 2);
1662
1663 uniform_count = midg_props.uniform_count;
1664 uniform_buffer_count = midg_props.uniform_buffer_count;
1665 }
1666
1667 pandecode_shader_prop("texture_count", texture_count, info.texture_count, false);
1668 pandecode_shader_prop("sampler_count", sampler_count, info.sampler_count, false);
1669 pandecode_shader_prop("attribute_count", attribute_count, info.attribute_count, false);
1670 pandecode_shader_prop("varying_count", varying_count, info.varying_count, false);
1671
1672 if (is_bifrost) {
1673 uint32_t opaque = state.preload.uniform_count << 15
1674 | state.preload.untyped;
1675
1676 switch (job_type) {
1677 case MALI_JOB_TYPE_VERTEX:
1678 DUMP_CL("Preload", PRELOAD_VERTEX, &opaque, 2);
1679 break;
1680 case MALI_JOB_TYPE_TILER:
1681 DUMP_CL("Preload", PRELOAD_FRAGMENT, &opaque, 2);
1682 break;
1683 case MALI_JOB_TYPE_COMPUTE:
1684 DUMP_CL("Preload", PRELOAD_COMPUTE, &opaque, 2);
1685 break;
1686 default:
1687 DUMP_CL("Preload", PRELOAD, &opaque, 2);
1688 break;
1689 }
1690 }
1691
1692 if (!is_bifrost) {
1693 /* TODO: Blend shaders routing/disasm */
1694 union midgard_blend blend;
1695 memcpy(&blend, &state.sfbd_blend, sizeof(blend));
1696 mali_ptr shader = pandecode_midgard_blend(&blend, state.multisample_misc.sfbd_blend_shader);
1697 if (shader & ~0xF)
1698 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1699 }
1700
1701 /* MRT blend fields are used whenever MFBD is used, with
1702 * per-RT descriptors */
1703
1704 if (job_type == MALI_JOB_TYPE_TILER && (is_bifrost || p->shared & MALI_MFBD)) {
1705 void* blend_base = ((void *) cl) + MALI_STATE_LENGTH;
1706
1707 for (unsigned i = 0; i < fbd_info.rt_count; i++) {
1708 mali_ptr shader = 0;
1709
1710 if (is_bifrost)
1711 shader = pandecode_bifrost_blend(blend_base, job_no, i);
1712 else
1713 shader = pandecode_midgard_blend_mrt(blend_base, job_no, i);
1714
1715 if (shader & ~0xF)
1716 pandecode_blend_shader_disassemble(shader, job_no, job_type, false, gpu_id);
1717
1718 }
1719 }
1720 } else
1721 pandecode_msg("XXX: missing shader descriptor\n");
1722
1723 if (p->viewport)
1724 DUMP_ADDR("Viewport", VIEWPORT, p->viewport, 1);
1725
1726 unsigned max_attr_index = 0;
1727
1728 if (p->attributes)
1729 max_attr_index = pandecode_attribute_meta(attribute_count, p->attributes, false, suffix);
1730
1731 if (p->attribute_buffers) {
1732 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->attribute_buffers);
1733 pandecode_attributes(attr_mem, p->attribute_buffers, job_no, suffix, max_attr_index, false, job_type);
1734 }
1735
1736 if (p->varyings) {
1737 varying_count = pandecode_attribute_meta(varying_count, p->varyings, true, suffix);
1738 }
1739
1740 if (p->varying_buffers) {
1741 attr_mem = pandecode_find_mapped_gpu_mem_containing(p->varying_buffers);
1742 pandecode_attributes(attr_mem, p->varying_buffers, job_no, suffix, varying_count, true, job_type);
1743 }
1744
1745 if (p->uniform_buffers) {
1746 if (uniform_buffer_count)
1747 pandecode_uniform_buffers(p->uniform_buffers, uniform_buffer_count, job_no);
1748 else
1749 pandecode_msg("warn: UBOs specified but not referenced\n");
1750 } else if (uniform_buffer_count)
1751 pandecode_msg("XXX: UBOs referenced but not specified\n");
1752
1753 /* We don't want to actually dump uniforms, but we do need to validate
1754 * that the counts we were given are sane */
1755
1756 if (p->push_uniforms) {
1757 if (uniform_count)
1758 pandecode_uniforms(p->push_uniforms, uniform_count);
1759 else
1760 pandecode_msg("warn: Uniforms specified but not referenced\n");
1761 } else if (uniform_count)
1762 pandecode_msg("XXX: Uniforms referenced but not specified\n");
1763
1764 if (p->textures)
1765 pandecode_textures(p->textures, texture_count, job_no, is_bifrost);
1766
1767 if (p->samplers)
1768 pandecode_samplers(p->samplers, sampler_count, job_no, is_bifrost);
1769 }
1770
1771 static void
1772 pandecode_tiler_heap_meta(mali_ptr gpu_va, int job_no)
1773 {
1774 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1775 const struct bifrost_tiler_heap_meta *PANDECODE_PTR_VAR(h, mem, gpu_va);
1776
1777 pandecode_log("struct bifrost_tiler_heap_meta tiler_heap_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
1778 pandecode_indent++;
1779
1780 if (h->zero) {
1781 pandecode_msg("XXX: tiler heap zero tripped\n");
1782 pandecode_prop("zero = 0x%x", h->zero);
1783 }
1784
1785 pandecode_prop("heap_size = 0x%x", h->heap_size);
1786 MEMORY_PROP(h, tiler_heap_start);
1787 MEMORY_PROP(h, tiler_heap_free);
1788
1789 /* this might point to the beginning of another buffer, when it's
1790 * really the end of the tiler heap buffer, so we have to be careful
1791 * here. but for zero length, we need the same pointer.
1792 */
1793
1794 if (h->tiler_heap_end == h->tiler_heap_start) {
1795 MEMORY_PROP(h, tiler_heap_start);
1796 } else {
1797 char *a = pointer_as_memory_reference(h->tiler_heap_end - 1);
1798 pandecode_prop("tiler_heap_end = %s + 1", a);
1799 free(a);
1800 }
1801
1802 for (int i = 0; i < 10; i++) {
1803 if (h->zeros[i] != 0) {
1804 pandecode_msg("XXX: tiler heap zero %d tripped, value %x\n",
1805 i, h->zeros[i]);
1806 }
1807 }
1808
1809 if (h->unk1 != 0x1) {
1810 pandecode_msg("XXX: tiler heap unk1 tripped\n");
1811 pandecode_prop("unk1 = 0x%x", h->unk1);
1812 }
1813
1814 if (h->unk7e007e != 0x7e007e) {
1815 pandecode_msg("XXX: tiler heap unk7e007e tripped\n");
1816 pandecode_prop("unk7e007e = 0x%x", h->unk7e007e);
1817 }
1818
1819 pandecode_indent--;
1820 pandecode_log("};\n");
1821 }
1822
1823 static void
1824 pandecode_tiler_meta(mali_ptr gpu_va, int job_no)
1825 {
1826 struct pandecode_mapped_memory *mem = pandecode_find_mapped_gpu_mem_containing(gpu_va);
1827 const struct bifrost_tiler_meta *PANDECODE_PTR_VAR(t, mem, gpu_va);
1828
1829 pandecode_tiler_heap_meta(t->tiler_heap_meta, job_no);
1830
1831 pandecode_log("struct bifrost_tiler_meta tiler_meta_%"PRIx64"_%d = {\n", gpu_va, job_no);
1832 pandecode_indent++;
1833
1834 pandecode_prop("tiler_heap_next_start = 0x%" PRIx32, t->tiler_heap_next_start);
1835 pandecode_prop("used_hierarchy_mask = 0x%" PRIx32, t->used_hierarchy_mask);
1836
1837 if (t->hierarchy_mask != 0xa &&
1838 t->hierarchy_mask != 0x14 &&
1839 t->hierarchy_mask != 0x28 &&
1840 t->hierarchy_mask != 0x50 &&
1841 t->hierarchy_mask != 0xa0)
1842 pandecode_prop("XXX: Unexpected hierarchy_mask (not 0xa, 0x14, 0x28, 0x50 or 0xa0)!");
1843
1844 pandecode_prop("hierarchy_mask = 0x%" PRIx16, t->hierarchy_mask);
1845
1846 pandecode_prop("flags = 0x%" PRIx16, t->flags);
1847
1848 pandecode_prop("width = MALI_POSITIVE(%d)", t->width + 1);
1849 pandecode_prop("height = MALI_POSITIVE(%d)", t->height + 1);
1850
1851 if (t->zero0) {
1852 pandecode_msg("XXX: tiler meta zero tripped\n");
1853 pandecode_prop("zero0 = 0x%" PRIx64, t->zero0);
1854 }
1855
1856 for (int i = 0; i < 12; i++) {
1857 if (t->zeros[i] != 0) {
1858 pandecode_msg("XXX: tiler heap zero %d tripped, value %" PRIx64 "\n",
1859 i, t->zeros[i]);
1860 }
1861 }
1862
1863 pandecode_indent--;
1864 pandecode_log("};\n");
1865 }
1866
1867 static void
1868 pandecode_primitive_size(union midgard_primitive_size u, bool constant)
1869 {
1870 if (u.pointer == 0x0)
1871 return;
1872
1873 pandecode_log(".primitive_size = {\n");
1874 pandecode_indent++;
1875
1876 if (constant) {
1877 pandecode_prop("constant = %f", u.constant);
1878 } else {
1879 MEMORY_PROP((&u), pointer);
1880 }
1881
1882 pandecode_indent--;
1883 pandecode_log("},\n");
1884 }
1885
1886 static int
1887 pandecode_vertex_job_bfr(const struct mali_job_descriptor_header *h,
1888 const struct pandecode_mapped_memory *mem,
1889 mali_ptr payload, int job_no, unsigned gpu_id)
1890 {
1891 struct bifrost_payload_vertex *PANDECODE_PTR_VAR(v, mem, payload);
1892
1893 struct MALI_DRAW draw;
1894 struct mali_draw_packed draw_packed;
1895 memcpy(&draw_packed, &v->postfix, sizeof(draw_packed));
1896 MALI_DRAW_unpack((const uint8_t *) &draw_packed, &draw); \
1897 pandecode_vertex_tiler_postfix_pre(&draw, job_no, h->job_type, "", true, gpu_id);
1898
1899 pandecode_vertex_tiler_prefix(&v->prefix, job_no, false);
1900 DUMP_CL("Draw", DRAW, &draw_packed, 2);
1901
1902 return sizeof(*v);
1903 }
1904
1905 static int
1906 pandecode_tiler_job_bfr(const struct mali_job_descriptor_header *h,
1907 const struct pandecode_mapped_memory *mem,
1908 mali_ptr payload, int job_no, unsigned gpu_id)
1909 {
1910 struct bifrost_payload_tiler *PANDECODE_PTR_VAR(t, mem, payload);
1911
1912 struct MALI_DRAW draw;
1913 struct mali_draw_packed draw_packed;
1914 memcpy(&draw_packed, &t->postfix, sizeof(draw_packed));
1915 MALI_DRAW_unpack((const uint8_t *) &draw_packed, &draw); \
1916 pandecode_vertex_tiler_postfix_pre(&draw, job_no, h->job_type, "", true, gpu_id);
1917 pandecode_tiler_meta(t->tiler_meta, job_no);
1918
1919 pandecode_vertex_tiler_prefix(&t->prefix, job_no, false);
1920
1921 /* TODO: gl_PointSize on Bifrost */
1922 pandecode_primitive_size(t->primitive_size, true);
1923
1924 if (t->zero1 || t->zero2 || t->zero3 || t->zero4 || t->zero5
1925 || t->zero6) {
1926 pandecode_msg("XXX: tiler only zero tripped\n");
1927 pandecode_prop("zero1 = 0x%" PRIx64, t->zero1);
1928 pandecode_prop("zero2 = 0x%" PRIx64, t->zero2);
1929 pandecode_prop("zero3 = 0x%" PRIx64, t->zero3);
1930 pandecode_prop("zero4 = 0x%" PRIx64, t->zero4);
1931 pandecode_prop("zero5 = 0x%" PRIx64, t->zero5);
1932 pandecode_prop("zero6 = 0x%" PRIx64, t->zero6);
1933 }
1934
1935 DUMP_CL("Draw", DRAW, &draw_packed, 2);
1936
1937 return sizeof(*t);
1938 }
1939
1940 static int
1941 pandecode_vertex_or_tiler_job_mdg(const struct mali_job_descriptor_header *h,
1942 const struct pandecode_mapped_memory *mem,
1943 mali_ptr payload, int job_no, unsigned gpu_id)
1944 {
1945 struct midgard_payload_vertex_tiler *PANDECODE_PTR_VAR(v, mem, payload);
1946 bool is_graphics = (h->job_type == MALI_JOB_TYPE_VERTEX) || (h->job_type == MALI_JOB_TYPE_TILER);
1947
1948 struct MALI_DRAW draw;
1949 struct mali_draw_packed draw_packed;
1950 memcpy(&draw_packed, &v->postfix, sizeof(draw_packed));
1951 MALI_DRAW_unpack((const uint8_t *) &draw_packed, &draw); \
1952 pandecode_vertex_tiler_postfix_pre(&draw, job_no, h->job_type, "", false, gpu_id);
1953
1954 pandecode_vertex_tiler_prefix(&v->prefix, job_no, is_graphics);
1955 DUMP_CL("Draw", DRAW, &draw_packed, 2);
1956
1957 struct MALI_PRIMITIVE primitive;
1958 struct mali_primitive_packed prim_packed = v->prefix.primitive;
1959 MALI_PRIMITIVE_unpack((const uint8_t *) &prim_packed, &primitive);
1960
1961 pandecode_primitive_size(v->primitive_size, primitive.point_size_array == 0);
1962
1963 return sizeof(*v);
1964 }
1965
1966 static int
1967 pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
1968 mali_ptr payload, int job_no,
1969 bool is_bifrost, unsigned gpu_id)
1970 {
1971 const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
1972
1973 bool is_mfbd = s->framebuffer & MALI_MFBD;
1974
1975 if (!is_mfbd && is_bifrost)
1976 pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
1977
1978 struct pandecode_fbd info;
1979
1980 if (is_mfbd)
1981 info = pandecode_mfbd_bfr(s->framebuffer & FBD_MASK, job_no, true, false, is_bifrost);
1982 else
1983 info = pandecode_sfbd(s->framebuffer & FBD_MASK, job_no, true, gpu_id);
1984
1985 /* Compute the tag for the tagged pointer. This contains the type of
1986 * FBD (MFBD/SFBD), and in the case of an MFBD, information about which
1987 * additional structures follow the MFBD header (an extra payload or
1988 * not, as well as a count of render targets) */
1989
1990 unsigned expected_tag = is_mfbd ? MALI_MFBD : 0;
1991
1992 if (is_mfbd) {
1993 if (info.has_extra)
1994 expected_tag |= MALI_MFBD_TAG_EXTRA;
1995
1996 expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
1997 }
1998
1999 if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
2000 pandecode_msg("XXX: unexpected tile coordinate bits\n");
2001 pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
2002 pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
2003 }
2004
2005 /* Extract tile coordinates */
2006
2007 unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
2008 unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
2009
2010 unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2011 unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
2012
2013 /* For the max, we also want the floored (rather than ceiled) version for checking */
2014
2015 unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
2016 unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
2017
2018 /* Validate the coordinates are well-ordered */
2019
2020 if (min_x == max_x)
2021 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2022 else if (min_x > max_x)
2023 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2024
2025 if (min_y == max_y)
2026 pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
2027 else if (min_y > max_y)
2028 pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
2029
2030 /* Validate the coordinates fit inside the framebuffer. We use floor,
2031 * rather than ceil, for the max coordinates, since the tile
2032 * coordinates for something like an 800x600 framebuffer will actually
2033 * resolve to 800x608, which would otherwise trigger a Y-overflow */
2034
2035 if ((min_x > info.width) || (max_x_f > info.width))
2036 pandecode_msg("XXX: tile coordinates overflow in X direction\n");
2037
2038 if ((min_y > info.height) || (max_y_f > info.height))
2039 pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
2040
2041 /* After validation, we print */
2042
2043 pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
2044
2045 /* The FBD is a tagged pointer */
2046
2047 unsigned tag = (s->framebuffer & ~FBD_MASK);
2048
2049 if (tag != expected_tag)
2050 pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
2051
2052 return sizeof(*s);
2053 }
2054
2055 /* Entrypoint to start tracing. jc_gpu_va is the GPU address for the first job
2056 * in the chain; later jobs are found by walking the chain. Bifrost is, well,
2057 * if it's bifrost or not. GPU ID is the more finegrained ID (at some point, we
2058 * might wish to combine this with the bifrost parameter) because some details
2059 * are model-specific even within a particular architecture. Minimal traces
2060 * *only* examine the job descriptors, skipping printing entirely if there is
2061 * no faults, and only descends into the payload if there are faults. This is
2062 * useful for looking for faults without the overhead of invasive traces. */
2063
2064 void
2065 pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
2066 {
2067 pandecode_dump_file_open();
2068
2069 struct mali_job_descriptor_header *h;
2070 unsigned job_descriptor_number = 0;
2071
2072 do {
2073 struct pandecode_mapped_memory *mem =
2074 pandecode_find_mapped_gpu_mem_containing(jc_gpu_va);
2075
2076 void *payload;
2077
2078 h = PANDECODE_PTR(mem, jc_gpu_va, struct mali_job_descriptor_header);
2079
2080 mali_ptr payload_ptr = jc_gpu_va + sizeof(*h);
2081 payload = pandecode_fetch_gpu_mem(mem, payload_ptr, 64);
2082
2083 int job_no = job_descriptor_number++;
2084
2085 /* If the job is good to go, skip it in minimal mode */
2086 if (minimal && (h->exception_status == 0x0 || h->exception_status == 0x1))
2087 continue;
2088
2089 pandecode_log("struct mali_job_descriptor_header job_%"PRIx64"_%d = {\n", jc_gpu_va, job_no);
2090 pandecode_indent++;
2091
2092 pandecode_prop("job_type = %s", mali_job_type_as_str(h->job_type));
2093
2094 if (h->job_descriptor_size)
2095 pandecode_prop("job_descriptor_size = %d", h->job_descriptor_size);
2096
2097 if (h->exception_status && h->exception_status != 0x1)
2098 pandecode_prop("exception_status = %x (source ID: 0x%x access: %s exception: 0x%x)",
2099 h->exception_status,
2100 (h->exception_status >> 16) & 0xFFFF,
2101 mali_exception_access_as_str((h->exception_status >> 8) & 0x3),
2102 h->exception_status & 0xFF);
2103
2104 if (h->first_incomplete_task)
2105 pandecode_prop("first_incomplete_task = %d", h->first_incomplete_task);
2106
2107 if (h->fault_pointer)
2108 pandecode_prop("fault_pointer = 0x%" PRIx64, h->fault_pointer);
2109
2110 if (h->job_barrier)
2111 pandecode_prop("job_barrier = %d", h->job_barrier);
2112
2113 pandecode_prop("job_index = %d", h->job_index);
2114
2115 if (h->unknown_flags)
2116 pandecode_prop("unknown_flags = %d", h->unknown_flags);
2117
2118 if (h->job_dependency_index_1)
2119 pandecode_prop("job_dependency_index_1 = %d", h->job_dependency_index_1);
2120
2121 if (h->job_dependency_index_2)
2122 pandecode_prop("job_dependency_index_2 = %d", h->job_dependency_index_2);
2123
2124 pandecode_indent--;
2125 pandecode_log("};\n");
2126
2127 switch (h->job_type) {
2128 case MALI_JOB_TYPE_WRITE_VALUE: {
2129 struct mali_payload_write_value *s = payload;
2130 pandecode_log("struct mali_payload_write_value payload_%"PRIx64"_%d = {\n", payload_ptr, job_no);
2131 pandecode_indent++;
2132 MEMORY_PROP(s, address);
2133
2134 if (s->value_descriptor != MALI_WRITE_VALUE_ZERO) {
2135 pandecode_msg("XXX: unknown value descriptor\n");
2136 pandecode_prop("value_descriptor = 0x%" PRIX32, s->value_descriptor);
2137 }
2138
2139 if (s->reserved) {
2140 pandecode_msg("XXX: set value tripped\n");
2141 pandecode_prop("reserved = 0x%" PRIX32, s->reserved);
2142 }
2143
2144 pandecode_prop("immediate = 0x%" PRIX64, s->immediate);
2145 pandecode_indent--;
2146 pandecode_log("};\n");
2147
2148 break;
2149 }
2150
2151 case MALI_JOB_TYPE_TILER:
2152 case MALI_JOB_TYPE_VERTEX:
2153 case MALI_JOB_TYPE_COMPUTE:
2154 if (bifrost) {
2155 if (h->job_type == MALI_JOB_TYPE_TILER)
2156 pandecode_tiler_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2157 else
2158 pandecode_vertex_job_bfr(h, mem, payload_ptr, job_no, gpu_id);
2159 } else
2160 pandecode_vertex_or_tiler_job_mdg(h, mem, payload_ptr, job_no, gpu_id);
2161
2162 break;
2163
2164 case MALI_JOB_TYPE_FRAGMENT:
2165 pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
2166 break;
2167
2168 default:
2169 break;
2170 }
2171 } while ((jc_gpu_va = h->next_job));
2172
2173 pandecode_map_read_write();
2174 }