Mercurial > vec

comparison test/test_benchmark.h @ 45:7955bed1d169

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
*: add preliminary floating point support no x86 intrinsics just yet, but I did add altivec since it's (arguably) the simplest :)
author Paper <paper@tflc.us>
date Wed, 30 Apr 2025 18:36:38 -0400
parents c6e0df09b86f
children
comparison
equal deleted inserted replaced
44:b0a3f0248ecc 45:7955bed1d169
1 1
2 /* ------------------------------------------------------------------------ */ 2 /* ------------------------------------------------------------------------ */
3 /* simple benchmark for getting the min/max range of an audio sample. */ 3 /* simple benchmark for getting the min/max range of an audio sample. */
4 4
5 extern void test_benchmark_sample_minmax_simple_impl(int16_t *smpl, uint32_t length, int32_t *pmin, int32_t *pmax); 5 extern void test_benchmark_sample_minmax_int8x2_impl(vec_int8 *smpl, uint32_t length, vec_int8 *pmin, vec_int8 *pmax);
6 extern void test_benchmark_sample_minmax_vec_impl(int16_t *smpl, uint32_t length, int32_t *pmin, int32_t *pmax);
7 6
8 VEC_FUNC_IMPL void test_benchmark_sample_minmax(void) 7 VEC_FUNC_IMPL void test_benchmark_sample_minmax(void)
9 { 8 {
10 int32_t min, max;
11 clock_t start, end;
12 int i; 9 int i;
13 int16_t *q = vec_malloc(16000001u * 2u);
14 10
15 printf("\nsigned 16-bit audio sample min/max - 1 thousand passes - 16000001 samples\n\n"); 11 union {
12 vec_int8 int8[16000001];
13 vec_uint8 uint8[16000001];
14 vec_int16 int16[16000001];
15 vec_uint16 uint16[16000001];
16 vec_int32 int32[16000001];
17 vec_uint32 uint32[16000001];
18 vec_int64 int64[16000001];
19 vec_uint64 uint64[16000001];
20 vec_f32 f32[16000001];
21 vec_f64 f64[16000001];
22 } *q;
16 23
17 start = clock(); 24 q = vec_malloc(sizeof(*q));
18 for (i = 0; i < 100; i++) { 25
19 min = INT32_MAX; 26 for (i = 0; i < 16000001; i++)
20 max = INT32_MIN; 27 q->f64[i] = (double)rand() / RAND_MAX;
21 test_benchmark_sample_minmax_vec_impl(q, 16000001u, &min, &max); 28
29 printf("\naudio sample min/max - 1 thousand passes - 16000001 samples\n\n");
30
31 #define DO_TIMER(TYPE,NAME,MIN,MAX) \
32 { \
33 vec_##TYPE min, max; \
34 clock_t start, end; \
35 \
36 start = clock(); \
37 \
38 for (i = 0; i < 1000; i++) { \
39 extern void test_benchmark_sample_minmax_##NAME##_impl(vec_##TYPE *smpl, uint32_t length, vec_##TYPE *pmin, vec_##TYPE *pmax); \
40 \
41 min = MAX; \
42 max = MIN; \
43 \
44 test_benchmark_sample_minmax_##NAME##_impl(q->TYPE, 16000001u, &min, &max); \
45 } \
46 \
47 end = clock(); \
48 \
49 printf("- %s: took %f secs\n", #NAME, (double)(end - start) / CLOCKS_PER_SEC); \
22 } 50 }
23 end = clock();
24 51
25 printf("- vec: took %f secs\n", (double)(end - start) / CLOCKS_PER_SEC); 52 DO_TIMER(int8, int8x2, INT8_MIN, INT8_MAX)
53 DO_TIMER(int8, int8x4, INT8_MIN, INT8_MAX)
54 DO_TIMER(int8, int8x8, INT8_MIN, INT8_MAX)
55 DO_TIMER(int8, int8x16, INT8_MIN, INT8_MAX)
56 DO_TIMER(int8, int8x32, INT8_MIN, INT8_MAX)
57 DO_TIMER(int8, int8x64, INT8_MIN, INT8_MAX)
58 DO_TIMER(int8, int8, INT8_MIN, INT8_MAX)
26 59
27 start = clock(); 60 DO_TIMER(int16, int16x2, INT16_MIN, INT16_MAX)
28 for (i = 0; i < 100; i++) { 61 DO_TIMER(int16, int16x4, INT16_MIN, INT16_MAX)
29 min = INT32_MAX; 62 DO_TIMER(int16, int16x8, INT16_MIN, INT16_MAX)
30 max = INT32_MIN; 63 DO_TIMER(int16, int16x16, INT16_MIN, INT16_MAX)
31 test_benchmark_sample_minmax_simple_impl(q, 16000001u, &min, &max); 64 DO_TIMER(int16, int16x32, INT16_MIN, INT16_MAX)
32 } 65 DO_TIMER(int16, int16, INT16_MIN, INT16_MAX)
33 end = clock();
34 66
35 printf("- simple: took %f secs\n", (double)(end - start) / CLOCKS_PER_SEC); 67 DO_TIMER(int32, int32x2, INT32_MIN, INT32_MAX)
68 DO_TIMER(int32, int32x4, INT32_MIN, INT32_MAX)
69 DO_TIMER(int32, int32x8, INT32_MIN, INT32_MAX)
70 DO_TIMER(int32, int32x16, INT32_MIN, INT32_MAX)
71 DO_TIMER(int32, int32, INT32_MIN, INT32_MAX)
72
73 DO_TIMER(int64, int64x2, INT64_MIN, INT64_MAX)
74 DO_TIMER(int64, int64x4, INT64_MIN, INT64_MAX)
75 DO_TIMER(int64, int64x8, INT64_MIN, INT64_MAX)
76 DO_TIMER(int64, int64, INT64_MIN, INT64_MAX)
77
78 DO_TIMER(uint8, uint8x2, 0, UINT8_MAX)
79 DO_TIMER(uint8, uint8x4, 0, UINT8_MAX)
80 DO_TIMER(uint8, uint8x8, 0, UINT8_MAX)
81 DO_TIMER(uint8, uint8x16, 0, UINT8_MAX)
82 DO_TIMER(uint8, uint8x32, 0, UINT8_MAX)
83 DO_TIMER(uint8, uint8x64, 0, UINT8_MAX)
84 DO_TIMER(uint8, uint8, 0, UINT8_MAX)
85
86 DO_TIMER(uint16, uint16x2, 0, UINT16_MAX)
87 DO_TIMER(uint16, uint16x4, 0, UINT16_MAX)
88 DO_TIMER(uint16, uint16x8, 0, UINT16_MAX)
89 DO_TIMER(uint16, uint16x16, 0, UINT16_MAX)
90 DO_TIMER(uint16, uint16x32, 0, UINT16_MAX)
91 DO_TIMER(uint16, uint16, 0, UINT16_MAX)
92
93 DO_TIMER(uint32, uint32x2, 0, UINT32_MAX)
94 DO_TIMER(uint32, uint32x4, 0, UINT32_MAX)
95 DO_TIMER(uint32, uint32x8, 0, UINT32_MAX)
96 DO_TIMER(uint32, uint32x16, 0, UINT32_MAX)
97 DO_TIMER(uint32, uint32, 0, UINT32_MAX)
98
99 DO_TIMER(uint64, uint64x2, 0, UINT64_MAX)
100 DO_TIMER(uint64, uint64x4, 0, UINT64_MAX)
101 DO_TIMER(uint64, uint64x8, 0, UINT64_MAX)
102 DO_TIMER(uint64, uint64, 0, UINT64_MAX)
103
104 DO_TIMER(f32, f32x2, -1.0f, 1.0f)
105 DO_TIMER(f32, f32x4, -1.0f, 1.0f)
106 DO_TIMER(f32, f32x8, -1.0f, 1.0f)
107 DO_TIMER(f32, f32x16, -1.0f, 1.0f)
108 DO_TIMER(f32, f32, -1.0f, 1.0f)
109
110 DO_TIMER(f64, f64x2, -1.0, 1.0)
111 DO_TIMER(f64, f64x4, -1.0, 1.0)
112 DO_TIMER(f64, f64x8, -1.0, 1.0)
113 DO_TIMER(f64, f64, -1.0, 1.0)
36 114
37 printf("\n"); 115 printf("\n");
38 116
39 vec_free(q); 117 vec_free(q);
40 } 118 }