1 /* Make the test not meaningless when asserts are disabled. */
2 #undef NDEBUG
3
4 #include <assert.h>
5 #include <inttypes.h>
6 #include <stdio.h>
7 #include <stdlib.h>
8
9 #include <amdgpu.h>
10 #include "drm-uapi/amdgpu_drm.h"
11 #include "drm-uapi/drm_fourcc.h"
12
13 #include "ac_surface.h"
14 #include "util/macros.h"
15 #include "util/u_math.h"
16 #include "util/u_vector.h"
17 #include "util/mesa-sha1.h"
18 #include "addrlib/inc/addrinterface.h"
19
20 #include "ac_surface_test_common.h"
21
22 /*
23 * The main goal of this test is making sure that we do
24 * not change the meaning of existing modifiers.
25 */
26
27 struct test_entry {
28 /* key part */
29 uint64_t modifier;
30 unsigned w;
31 unsigned h;
32 enum pipe_format format;
33
34 /* debug info */
35 const char *name;
36 uint8_t pipes;
37 uint8_t rb;
38 uint8_t banks_or_pkrs;
39 uint8_t se;
40
41 /* value to determine uniqueness */
42 unsigned char hash[20];
43
44 /* u_vector requires power of two sizing */
45 char padding[sizeof(void*) == 8 ? 8 : 16];
46 };
47
48 static uint64_t
block_count(unsigned w,unsigned h,unsigned elem_bits,unsigned block_bits,unsigned * aligned_pitch,unsigned * aligned_height)49 block_count(unsigned w, unsigned h, unsigned elem_bits, unsigned block_bits,
50 unsigned *aligned_pitch, unsigned *aligned_height)
51 {
52 unsigned align_bits = block_bits - elem_bits;
53 unsigned w_align = 1 << (align_bits / 2 + align_bits % 2);
54 unsigned h_align = 1 << (align_bits / 2);
55
56 w = align(w, w_align);
57 h = align(h, h_align);
58
59 if (aligned_pitch)
60 *aligned_pitch = w;
61
62 if (aligned_height)
63 *aligned_height = h;
64 return ((uint64_t)w * h) >> align_bits;
65 }
66
67
68 static ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT
gfx9_get_addr_from_coord_base(ADDR_HANDLE addrlib,const struct radeon_surf * surf,unsigned w,unsigned h,enum pipe_format format,bool rb_aligned,bool pipe_aligned)69 gfx9_get_addr_from_coord_base(ADDR_HANDLE addrlib, const struct radeon_surf *surf,
70 unsigned w, unsigned h, enum pipe_format format,
71 bool rb_aligned, bool pipe_aligned)
72 {
73 ADDR2_COMPUTE_DCCINFO_INPUT din = {0};
74 ADDR2_COMPUTE_DCCINFO_OUTPUT dout = {0};
75 din.size = sizeof(ADDR2_COMPUTE_DCCINFO_INPUT);
76 dout.size = sizeof(ADDR2_COMPUTE_DCCINFO_OUTPUT);
77
78 din.swizzleMode = surf->u.gfx9.swizzle_mode;
79 din.resourceType = ADDR_RSRC_TEX_2D;
80 din.bpp = util_format_get_blocksizebits(format);
81 din.unalignedWidth = w;
82 din.unalignedHeight = h;
83 din.numSlices = 1;
84 din.numMipLevels = 1;
85 din.numFrags = 1;
86 din.dccKeyFlags.pipeAligned = surf->u.gfx9.color.dcc.pipe_aligned;
87 din.dccKeyFlags.rbAligned = surf->u.gfx9.color.dcc.rb_aligned;
88 din.dataSurfaceSize = surf->surf_size;
89
90 ADDR_E_RETURNCODE ret = Addr2ComputeDccInfo(addrlib, &din, &dout);
91 assert(ret == ADDR_OK);
92
93 ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT dcc_input = {0};
94 dcc_input.size = sizeof(dcc_input);
95 dcc_input.swizzleMode = surf->u.gfx9.swizzle_mode;
96 dcc_input.resourceType = ADDR_RSRC_TEX_2D;
97 dcc_input.bpp = din.bpp;
98 dcc_input.numSlices = 1;
99 dcc_input.numMipLevels = 1;
100 dcc_input.numFrags = 1;
101 dcc_input.dccKeyFlags.pipeAligned = pipe_aligned;
102 dcc_input.dccKeyFlags.rbAligned = rb_aligned;
103 dcc_input.pitch = dout.pitch;
104 dcc_input.height = dout.height;
105 dcc_input.compressBlkWidth = dout.compressBlkWidth;
106 dcc_input.compressBlkHeight = dout.compressBlkHeight;
107 dcc_input.compressBlkDepth = dout.compressBlkDepth;
108 dcc_input.metaBlkWidth = dout.metaBlkWidth;
109 dcc_input.metaBlkHeight = dout.metaBlkHeight;
110 dcc_input.metaBlkDepth = dout.metaBlkDepth;
111 return dcc_input;
112 }
113
gfx9_generate_hash(struct ac_addrlib * ac_addrlib,struct test_entry * entry,const struct radeon_surf * surf)114 static void gfx9_generate_hash(struct ac_addrlib *ac_addrlib,
115 struct test_entry *entry,
116 const struct radeon_surf *surf)
117 {
118 ADDR_HANDLE addrlib = ac_addrlib_get_handle(ac_addrlib);
119
120 srandom(53);
121 struct mesa_sha1 ctx;
122 _mesa_sha1_init(&ctx);
123
124 _mesa_sha1_update(&ctx, &surf->total_size, sizeof(surf->total_size));
125 _mesa_sha1_update(&ctx, &surf->meta_offset, sizeof(surf->meta_offset));
126 _mesa_sha1_update(&ctx, &surf->display_dcc_offset, sizeof(surf->display_dcc_offset));
127 _mesa_sha1_update(&ctx, &surf->u.gfx9.color.display_dcc_pitch_max,
128 sizeof(surf->u.gfx9.color.display_dcc_pitch_max));
129
130 ADDR2_COMPUTE_SURFACE_ADDRFROMCOORD_INPUT input = {0};
131 input.size = sizeof(input);
132 input.swizzleMode = surf->u.gfx9.swizzle_mode;
133 input.resourceType = ADDR_RSRC_TEX_2D;
134 input.bpp = util_format_get_blocksizebits(entry->format);
135 input.unalignedWidth = entry->w;
136 input.unalignedHeight = entry->h;
137 input.numSlices = 1;
138 input.numMipLevels = 1;
139 input.numSamples = 1;
140 input.numFrags = 1;
141 input.pitchInElement = surf->u.gfx9.surf_pitch;
142
143 ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT dcc_input = {0};
144 if (surf->meta_offset) {
145 dcc_input = gfx9_get_addr_from_coord_base(addrlib, surf, entry->w,
146 entry->h, entry->format,
147 surf->u.gfx9.color.dcc.rb_aligned,
148 surf->u.gfx9.color.dcc.pipe_aligned);
149 }
150
151 ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT display_dcc_input = {0};
152 if (surf->display_dcc_offset) {
153 display_dcc_input = gfx9_get_addr_from_coord_base(addrlib, surf, entry->w,
154 entry->h, entry->format,
155 false, false);
156 }
157
158 for (unsigned i = 0; i < 1000; ++i) {
159 int32_t x, y;
160 x = random();
161 y = random();
162
163 input.x = (x & INT_MAX) % entry->w;
164 input.y = (y & INT_MAX) % entry->h;
165
166 ADDR2_COMPUTE_SURFACE_ADDRFROMCOORD_OUTPUT output = {0};
167 output.size = sizeof(output);
168
169 ADDR_E_RETURNCODE ret = Addr2ComputeSurfaceAddrFromCoord(addrlib, &input, &output);
170 assert(ret == ADDR_OK);
171
172 _mesa_sha1_update(&ctx, &output.addr, sizeof(output.addr));
173
174 if (surf->meta_offset) {
175 dcc_input.x = (x & INT_MAX) % entry->w;
176 dcc_input.y = (y & INT_MAX) % entry->h;
177
178 ADDR2_COMPUTE_DCC_ADDRFROMCOORD_OUTPUT dcc_output = {0};
179 dcc_output.size = sizeof(dcc_output);
180
181 ret = Addr2ComputeDccAddrFromCoord(addrlib, &dcc_input, &dcc_output);
182 assert(ret == ADDR_OK);
183
184 _mesa_sha1_update(&ctx, &dcc_output.addr, sizeof(dcc_output.addr));
185 }
186
187 if (surf->display_dcc_offset) {
188 display_dcc_input.x = (x & INT_MAX) % entry->w;
189 display_dcc_input.y = (y & INT_MAX) % entry->h;
190
191 ADDR2_COMPUTE_DCC_ADDRFROMCOORD_OUTPUT dcc_output = {0};
192 dcc_output.size = sizeof(dcc_output);
193
194 ret = Addr2ComputeDccAddrFromCoord(addrlib, &display_dcc_input, &dcc_output);
195 assert(ret == ADDR_OK);
196
197 _mesa_sha1_update(&ctx, &dcc_output.addr, sizeof(dcc_output.addr));
198 }
199 }
200
201 _mesa_sha1_final(&ctx, entry->hash);
202 }
203
gfx12_generate_hash(struct ac_addrlib * ac_addrlib,struct test_entry * entry,const struct radeon_surf * surf)204 static void gfx12_generate_hash(struct ac_addrlib *ac_addrlib,
205 struct test_entry *entry,
206 const struct radeon_surf *surf)
207 {
208 ADDR_HANDLE addrlib = ac_addrlib_get_handle(ac_addrlib);
209
210 srandom(53);
211 struct mesa_sha1 ctx;
212 _mesa_sha1_init(&ctx);
213
214 _mesa_sha1_update(&ctx, &surf->total_size, sizeof(surf->total_size));
215 /* We need to hash these even though they are not used by gfx12. */
216 _mesa_sha1_update(&ctx, &surf->meta_offset, sizeof(surf->meta_offset));
217 _mesa_sha1_update(&ctx, &surf->display_dcc_offset, sizeof(surf->display_dcc_offset));
218 _mesa_sha1_update(&ctx, &surf->u.gfx9.color.display_dcc_pitch_max,
219 sizeof(surf->u.gfx9.color.display_dcc_pitch_max));
220
221 ADDR3_COMPUTE_SURFACE_ADDRFROMCOORD_INPUT input = {0};
222 input.size = sizeof(input);
223 input.swizzleMode = surf->u.gfx9.swizzle_mode;
224 input.flags.color = 1;
225 input.flags.texture = 1;
226 input.resourceType = ADDR_RSRC_TEX_2D;
227 input.bpp = util_format_get_blocksizebits(entry->format);
228 input.unAlignedDims.width = entry->w;
229 input.unAlignedDims.height = entry->h;
230 input.unAlignedDims.depth = 1;
231 input.numMipLevels = 1;
232 input.numSamples = 1;
233 input.pitchInElement = surf->u.gfx9.surf_pitch;
234
235 for (unsigned i = 0; i < 1000; ++i) {
236 int32_t x, y;
237 x = random();
238 y = random();
239
240 input.x = (x & INT_MAX) % entry->w;
241 input.y = (y & INT_MAX) % entry->h;
242
243 ADDR3_COMPUTE_SURFACE_ADDRFROMCOORD_OUTPUT output = {0};
244 output.size = sizeof(output);
245
246 ADDR_E_RETURNCODE ret = Addr3ComputeSurfaceAddrFromCoord(addrlib, &input, &output);
247 assert(ret == ADDR_OK);
248
249 _mesa_sha1_update(&ctx, &output.addr, sizeof(output.addr));
250 }
251
252 _mesa_sha1_final(&ctx, entry->hash);
253 }
254
test_modifier(const struct radeon_info * info,const char * name,struct ac_addrlib * addrlib,uint64_t modifier,enum pipe_format format,struct u_vector * test_entries)255 static void test_modifier(const struct radeon_info *info,
256 const char *name,
257 struct ac_addrlib *addrlib,
258 uint64_t modifier,
259 enum pipe_format format,
260 struct u_vector *test_entries)
261 {
262 unsigned elem_bits = util_logbase2(util_format_get_blocksize(format));
263 const unsigned dims[][2] = {
264 {1, 1},
265 {1920, 1080},
266 {1366, 768},
267 {3840, 2160},
268 {233, 938},
269 };
270 for (unsigned i = 0; i < ARRAY_SIZE(dims); ++i) {
271 struct ac_surf_config config = (struct ac_surf_config) {
272 .info = (struct ac_surf_info) {
273 .width = dims[i][0],
274 .height = dims[i][1],
275 .depth = 1,
276 .samples = 1,
277 .storage_samples = 1,
278 .levels = 1,
279 .num_channels = 3,
280 .array_size = 1
281 },
282 };
283
284 struct test_entry entry = {
285 .modifier = modifier,
286 .w = config.info.width,
287 .h = config.info.height,
288 .format = format,
289 .name = name,
290 .pipes = G_0098F8_NUM_PIPES(info->gb_addr_config),
291 .rb = G_0098F8_NUM_RB_PER_SE(info->gb_addr_config) +
292 G_0098F8_NUM_SHADER_ENGINES_GFX9(info->gb_addr_config),
293 .se = G_0098F8_NUM_SHADER_ENGINES_GFX9(info->gb_addr_config),
294 .banks_or_pkrs = info->gfx_level >= GFX10 ?
295 G_0098F8_NUM_PKRS(info->gb_addr_config) : G_0098F8_NUM_BANKS(info->gb_addr_config)
296 };
297
298 struct radeon_surf surf = (struct radeon_surf) {
299 .blk_w = 1,
300 .blk_h = 1,
301 .bpe = util_format_get_blocksize(format),
302 .modifier = modifier,
303 };
304
305 int r = ac_compute_surface(addrlib, info, &config, RADEON_SURF_MODE_2D, &surf);
306 assert(!r);
307
308 assert(surf.cmask_offset == 0);
309 assert(surf.fmask_offset == 0);
310
311 uint64_t surf_size;
312 unsigned aligned_pitch, aligned_height;
313 if (modifier != DRM_FORMAT_MOD_LINEAR) {
314 unsigned block_size_bits;
315
316 if (info->gfx_level >= GFX12) {
317 assert(surf.u.gfx9.swizzle_mode == ADDR3_64KB_2D ||
318 surf.u.gfx9.swizzle_mode == ADDR3_256B_2D);
319 block_size_bits = (surf.u.gfx9.swizzle_mode == ADDR3_256B_2D) ? 8 : 16;
320 } else {
321 block_size_bits = surf.u.gfx9.swizzle_mode >= ADDR_SW_256KB_Z_X ? 18 : 16;
322 }
323
324 surf_size = block_count(dims[i][0], dims[i][1],
325 elem_bits, block_size_bits, &aligned_pitch,
326 &aligned_height) << block_size_bits;
327 } else {
328 unsigned alignment = 256;
329
330 aligned_pitch = align(dims[i][0], alignment / util_format_get_blocksize(format));
331 aligned_height = dims[i][1];
332 surf_size = align(dims[i][0] * util_format_get_blocksize(format), alignment) * dims[i][1];
333 }
334
335 assert(surf.u.gfx9.surf_pitch == aligned_pitch);
336 assert(surf.u.gfx9.surf_height == aligned_height);
337 assert(surf.surf_size == surf_size);
338 uint64_t expected_offset = surf_size;
339
340 if (ac_modifier_has_dcc_retile(modifier)) {
341 unsigned dcc_align = info->gfx_level >= GFX10 ? 4096 : 65536;
342 unsigned dcc_pitch;
343 uint64_t dcc_size = block_count(dims[i][0], dims[i][1],
344 elem_bits, 20, &dcc_pitch,
345 NULL) << 12;
346
347 assert(surf.u.gfx9.color.display_dcc_size == align(dcc_size, dcc_align));
348 assert(surf.u.gfx9.color.display_dcc_pitch_max + 1 == dcc_pitch);
349 assert(surf.display_dcc_offset == expected_offset);
350
351 expected_offset += align(dcc_size, dcc_align);
352 } else
353 assert(!surf.display_dcc_offset);
354
355 if (info->gfx_level < GFX12 && ac_modifier_has_dcc(modifier)) {
356 uint64_t dcc_align = 1;
357 unsigned block_bits;
358 if (info->gfx_level >= GFX10) {
359 unsigned num_pipes = G_0098F8_NUM_PIPES(info->gb_addr_config);
360 if (info->gfx_level >= GFX10_3 &&
361 G_0098F8_NUM_PKRS(info->gb_addr_config) == num_pipes && num_pipes > 1)
362 ++num_pipes;
363 block_bits = 16 +
364 num_pipes +
365 G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config);
366 block_bits = MAX2(block_bits, 20);
367 dcc_align = MAX2(4096, 256 <<
368 (num_pipes +
369 G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(info->gb_addr_config)));
370 } else {
371 block_bits = 18 +
372 G_0098F8_NUM_RB_PER_SE(info->gb_addr_config) +
373 G_0098F8_NUM_SHADER_ENGINES_GFX9(info->gb_addr_config);
374 block_bits = MAX2(block_bits, 20);
375 dcc_align = 65536;
376 }
377
378 expected_offset = align(expected_offset, dcc_align);
379 assert(surf.meta_offset == expected_offset);
380
381 uint64_t dcc_size = block_count(dims[i][0], dims[i][1],
382 elem_bits, block_bits,
383 NULL, NULL) << (block_bits - 8);
384 dcc_size = align64(dcc_size, dcc_align);
385 assert(surf.meta_size == dcc_size);
386
387 expected_offset += dcc_size;
388 } else
389 assert(!surf.meta_offset);
390
391 assert(surf.total_size == expected_offset);
392
393 if (info->gfx_level >= GFX12)
394 gfx12_generate_hash(addrlib, &entry, &surf);
395 else
396 gfx9_generate_hash(addrlib, &entry, &surf);
397
398 *(struct test_entry*)u_vector_add(test_entries) = entry;
399 }
400 }
401
run_modifier_test(struct u_vector * test_entries,const char * name,const struct radeon_info * info)402 static void run_modifier_test(struct u_vector *test_entries, const char *name,
403 const struct radeon_info *info)
404 {
405 struct ac_addrlib *addrlib = ac_addrlib_create(info, NULL);
406 assert(addrlib);
407
408 const struct ac_modifier_options options = {
409 .dcc = true,
410 .dcc_retile = true,
411 };
412
413 enum pipe_format formats[] = {
414 PIPE_FORMAT_R8_UNORM,
415 PIPE_FORMAT_R16_UNORM,
416 PIPE_FORMAT_R32_FLOAT,
417 PIPE_FORMAT_R32G32_FLOAT,
418 PIPE_FORMAT_R32G32B32A32_FLOAT
419 };
420 for (unsigned j = 0; j < ARRAY_SIZE(formats); ++j) {
421 unsigned mod_count = 0;
422 ac_get_supported_modifiers(info, &options, formats[j], &mod_count, NULL);
423
424 uint64_t *modifiers = malloc(sizeof(uint64_t) * mod_count);
425 ac_get_supported_modifiers(info, &options, formats[j], &mod_count, modifiers);
426
427 for (unsigned i = 0; i < mod_count; ++i) {
428 test_modifier(info, name, addrlib, modifiers[i], formats[j], test_entries);
429 }
430
431 free(modifiers);
432 }
433 ac_addrlib_destroy(addrlib);
434 }
435
compare_test_entry(const void * a,const void * b)436 static int compare_test_entry(const void *a, const void *b)
437 {
438 return memcmp(a, b, sizeof(struct test_entry));
439 }
440
test_entry_key_equal(const struct test_entry * a,const struct test_entry * b)441 static bool test_entry_key_equal(const struct test_entry *a, const struct test_entry *b)
442 {
443 return a->modifier == b->modifier && a->w == b->w && a->h == b->h && a->format == b->format;
444 }
445
test_entry_value_equal(const struct test_entry * a,const struct test_entry * b)446 static bool test_entry_value_equal(const struct test_entry *a, const struct test_entry *b)
447 {
448 if (memcmp(a->hash, b->hash, sizeof(a->hash)))
449 return false;
450 return true;
451 }
452
print_test_entry(const struct test_entry * e)453 static void print_test_entry(const struct test_entry *e)
454 {
455 fprintf(stderr, "%.16" PRIx64 " %.4d %.4d %.2d %s %d %d %d %d\n", e->modifier, e->w, e->h,
456 util_format_get_blocksize(e->format), e->name, e->pipes, e->rb, e->se, e->banks_or_pkrs);
457 }
458
main()459 int main()
460 {
461 STATIC_ASSERT(sizeof(struct test_entry) == 64);
462
463 struct u_vector test_entries;
464 u_vector_init_pow2(&test_entries, 64, sizeof(struct test_entry));
465
466 for (unsigned i = 0; i < ARRAY_SIZE(testcases); ++i) {
467 struct radeon_info info = get_radeon_info(&testcases[i]);
468
469 run_modifier_test(&test_entries, testcases[i].name, &info);
470 }
471
472 qsort(u_vector_tail(&test_entries),
473 u_vector_length(&test_entries),
474 sizeof(struct test_entry),
475 compare_test_entry);
476
477 struct test_entry *cur, *prev = NULL, *prevprev = NULL;
478 bool mismatched_duplicates = false;
479 u_vector_foreach(cur, &test_entries) {
480 if (prev && test_entry_key_equal(cur, prev) &&
481 !test_entry_value_equal(cur, prev)) {
482 if (!prevprev || !test_entry_key_equal(prev, prevprev)) {
483 print_test_entry(prev);
484 }
485 print_test_entry(cur);
486 mismatched_duplicates = true;
487 }
488 prevprev = prev;
489 prev = cur;
490 }
491 assert(!mismatched_duplicates);
492
493 u_vector_finish(&test_entries);
494
495 return 0;
496 }
497