|
7 | 7 | #include <stdlib.h> |
8 | 8 |
|
9 | 9 | int issue21() { |
10 | | - size_t sz = 110; |
11 | | - size_t i; |
12 | | - uint32_t *in = malloc(sz * sizeof(uint32_t)); |
13 | | - uint32_t *out = malloc(sz * sizeof(uint32_t)); |
14 | | - for (i = 0; i < sz; ++i) |
15 | | - in[i] = 255; |
16 | | - uint32_t b = maxbits_length(in, sz); |
17 | | - uint8_t *buf = malloc(simdpack_compressedbytes(sz, b)); |
18 | | - __m128i *end = simdpack_length(in, sz, (__m128i *)buf, b); |
19 | | - if((uint8_t *)end - buf != simdpack_compressedbytes(sz, b)) { |
20 | | - printf("bad mem usage\n"); |
21 | | - return -1; |
22 | | - } |
23 | | - simdunpack_length((const __m128i *)buf, sz, out, b); |
24 | | - for (i = 0; i < sz; ++i) { |
25 | | - if (in[i] != out[i]) { |
26 | | - printf("bug\n"); |
27 | | - return -1; |
| 10 | + printf("issue21"); |
| 11 | + fflush(stdout); |
| 12 | + for (uint32_t bw = 0; bw < 30; bw++) { |
| 13 | + printf("."); |
| 14 | + fflush(stdout); |
| 15 | + for (size_t sz = 1; sz < 4096; sz++) { |
| 16 | + |
| 17 | + size_t i; |
| 18 | + uint32_t *in = malloc(sz * sizeof(uint32_t)); |
| 19 | + uint32_t *out = malloc(sz * sizeof(uint32_t)); |
| 20 | + for (i = 0; i < sz; ++i) |
| 21 | + in[i] = (1 << bw) - 1; |
| 22 | + uint32_t b = maxbits_length(in, sz); |
| 23 | + uint8_t *buf = malloc(simdpack_compressedbytes(sz, b)); |
| 24 | + __m128i *end = simdpack_length(in, sz, (__m128i *)buf, b); |
| 25 | + if ((uint8_t *)end - buf != simdpack_compressedbytes(sz, b)) { |
| 26 | + printf("bad mem usage\n"); |
| 27 | + return -1; |
| 28 | + } |
| 29 | + simdunpack_length((const __m128i *)buf, sz, out, b); |
| 30 | + for (i = 0; i < sz; ++i) { |
| 31 | + if (in[i] != out[i]) { |
| 32 | + printf("bug\n"); |
| 33 | + return -1; |
| 34 | + } |
| 35 | + } |
| 36 | + free(in); |
| 37 | + free(out); |
| 38 | + free(buf); |
28 | 39 | } |
29 | 40 | } |
30 | | - free(in); |
31 | | - free(out); |
32 | | - free(buf); |
| 41 | + printf("\n"); |
33 | 42 | return 0; |
34 | 43 | } |
35 | 44 |
|
36 | 45 | int issue21FOR() { |
37 | | - size_t i; |
38 | | - size_t sz = 110; |
39 | | - uint32_t *in = malloc(sz * sizeof(uint32_t)); |
40 | | - uint32_t *out = malloc(sz * sizeof(uint32_t)); |
41 | | - in[0] = 0; |
42 | | - for (i = 1; i < sz; ++i) |
43 | | - in[i] = 255; |
44 | | - uint32_t b = maxbits_length(in, sz); |
45 | | - uint8_t *buf = malloc(simdpackFOR_compressedbytes(sz, b)); |
46 | | - __m128i *end = simdpackFOR_length(0, in, sz, (__m128i *)buf, b); |
47 | | - if((uint8_t *)end - buf != simdpackFOR_compressedbytes(sz, b)) { |
48 | | - printf("bad mem usage\n"); |
49 | | - return -1; |
50 | | - } |
51 | | - simdunpackFOR_length(0, (const __m128i *)buf, sz, out, b); |
52 | | - for (i = 0; i < sz; ++i) { |
53 | | - if (in[i] != out[i]) { |
54 | | - printf("bug\n"); |
55 | | - return -1; |
| 46 | + size_t i, j; |
| 47 | + printf("issue21for"); |
| 48 | + fflush(stdout); |
| 49 | + for (uint32_t bw = 0; bw < 30; bw++) { |
| 50 | + printf("."); |
| 51 | + fflush(stdout); |
| 52 | + for (size_t sz = 1; sz < 4096; sz++) { |
| 53 | + |
| 54 | + uint32_t *in = malloc(sz * sizeof(uint32_t)); |
| 55 | + uint32_t *out = malloc(sz * sizeof(uint32_t)); |
| 56 | + in[0] = 0; |
| 57 | + for (i = 1; i < sz; ++i) |
| 58 | + in[i] = (1 << bw) - 1; |
| 59 | + uint32_t b = maxbits_length(in, sz); |
| 60 | + uint8_t *buf = malloc(simdpackFOR_compressedbytes(sz, b)); |
| 61 | + __m128i *end = simdpackFOR_length(0, in, sz, (__m128i *)buf, b); |
| 62 | + if ((uint8_t *)end - buf != simdpackFOR_compressedbytes(sz, b)) { |
| 63 | + printf("bad mem usage\n"); |
| 64 | + return -1; |
| 65 | + } |
| 66 | + simdunpackFOR_length(0, (const __m128i *)buf, sz, out, b); |
| 67 | + for (i = 0; i < sz; ++i) { |
| 68 | + if (in[i] != out[i]) { |
| 69 | + for (j = 0; j < sz; ++j) { |
| 70 | + printf("%zu : %u %u \n", j, in[j], out[j]); |
| 71 | + } |
| 72 | + printf("bug\n"); |
| 73 | + return -1; |
| 74 | + } |
| 75 | + } |
| 76 | + free(in); |
| 77 | + free(out); |
| 78 | + free(buf); |
56 | 79 | } |
57 | 80 | } |
58 | | - free(in); |
59 | | - free(out); |
60 | | - free(buf); |
| 81 | + printf("\n"); |
61 | 82 | return 0; |
62 | 83 | } |
63 | 84 |
|
@@ -404,7 +425,7 @@ int testavx2() { |
404 | 425 | int k; |
405 | 426 | printf(" gap = %u \n", gap); |
406 | 427 | for (k = 0; k < N; ++k) |
407 | | - datain[k] = (uint32_t)(((uint64_t)k * gap)&0xFFFFFFFF); |
| 428 | + datain[k] = (uint32_t)(((uint64_t)k * gap) & 0xFFFFFFFF); |
408 | 429 | for (k = 0; k * AVXBlockSize < N; ++k) { |
409 | 430 | /* |
410 | 431 | First part works for general arrays (sorted or unsorted) |
@@ -567,7 +588,7 @@ int test() { |
567 | 588 | int k; |
568 | 589 | printf(" gap = %u \n", gap); |
569 | 590 | for (k = 0; k < N; ++k) |
570 | | - datain[k] = (uint32_t)(((uint64_t)k * gap)&0xFFFFFFFF); |
| 591 | + datain[k] = (uint32_t)(((uint64_t)k * gap) & 0xFFFFFFFF); |
571 | 592 | for (k = 0; k * SIMDBlockSize < N; ++k) { |
572 | 593 | /* |
573 | 594 | First part works for general arrays (sorted or unsorted) |
@@ -630,7 +651,7 @@ int testFOR() { |
630 | 651 | int k; |
631 | 652 | printf(" gap = %u \n", gap); |
632 | 653 | for (k = 0; k < N; ++k) |
633 | | - datain[k] = (uint32_t)(((uint64_t)k * gap)&0xFFFFFFFF); |
| 654 | + datain[k] = (uint32_t)(((uint64_t)k * gap) & 0xFFFFFFFF); |
634 | 655 | for (k = 0; k * SIMDBlockSize < N; ++k) { |
635 | 656 | int j; |
636 | 657 | simdmaxmin_length(datain + k * SIMDBlockSize, SIMDBlockSize, &tmin, |
|
0 commit comments