Unified · Split
| @@ -33,64 +33,119 @@ do { \ · 32 unmodified lines | |||
| 33 | int i; \ | 33 | int i; \ |
| 34 | double bmg[2], stddev = 10.0, mean = 0.0; \ | 34 | double bmg[2], stddev = 10.0, mean = 0.0; \ |
| 35 | \ | 35 | \ |
| 36 | for (i = 0; i < LEN*2+8; i += 2) { \ | 36 | for (i = 0; i < BUF_SIZE; i += 2) { \ |
| 37 | av_bmg_get(&checkasm_lfg, bmg); \ | 37 | av_bmg_get(&checkasm_lfg, bmg); \ |
| 38 | buf[i] = bmg[0] * stddev + mean; \ | 38 | buf[i] = bmg[0] * stddev + mean; \ |
| 39 | buf[i + 1] = bmg[1] * stddev + mean; \ | 39 | buf[i + 1] = bmg[1] * stddev + mean; \ |
| 40 | } \ | 40 | } \ |
| 41 | } while(0); | 41 | } while(0); |
| 42 | 42 | ||
| 43 | static void test_fcmul_add(const float *src0, const float *src1, const float *src2) | 43 | static void test_fcmul_add(AudioFIRDSPContext *fir) |
| 44 | { | 44 | { |
| 45 | LOCAL_ALIGNED_32(float, cdst, [LEN*2+8]); | 45 | #define BUF_SIZE LEN*2+8 |
| 46 | LOCAL_ALIGNED_32(float, odst, [LEN*2+8]); | 46 | LOCAL_ALIGNED_32(float, src0, [BUF_SIZE]); |
| 47 | int i; | 47 | LOCAL_ALIGNED_32(float, src1, [BUF_SIZE]); |
| 48 | 48 | LOCAL_ALIGNED_32(float, src2, [BUF_SIZE]); | |
| 49 | declare_func(void, float *sum, const float *t, const float *c, | 49 | |
| 50 | ptrdiff_t len); | 50 | randomize_buffer(src0); |
| 51 | 51 | randomize_buffer(src1); | |
| 52 | memcpy(cdst, src0, (LEN*2+8) * sizeof(float)); | 52 | randomize_buffer(src2); |
| 53 | memcpy(odst, src0, (LEN*2+8) * sizeof(float)); | 53 | |
| 54 | call_ref(cdst, src1, src2, LEN); | 54 | if (check_func(fir->fcmul_add, "fcmul_add")) { |
| 55 | call_new(odst, src1, src2, LEN); | 55 | LOCAL_ALIGNED_32(float, cdst, [BUF_SIZE]); |
| 56 | for (i = 0; i <= LEN*2; i++) { | 56 | LOCAL_ALIGNED_32(float, odst, [BUF_SIZE]); |
| 57 | int idx = i & ~1; | 57 | int i; |
| 58 | float cre = src2[idx]; | 58 | |
| 59 | float cim = src2[idx + 1]; | 59 | declare_func(void, float *sum, const float *t, const float *c, |
| 60 | float tre = src1[idx]; | 60 | ptrdiff_t len); |
| 61 | float tim = src1[idx + 1]; | 61 | |
| 62 | double t = fabs(src0[i]) + | 62 | memcpy(cdst, src0, (BUF_SIZE) * sizeof(float)); |
| 63 | fabs(tre) + fabs(tim) + fabs(cre) + fabs(cim) + | 63 | memcpy(odst, src0, (BUF_SIZE) * sizeof(float)); |
| 64 | fabs(tre * cre) + fabs(tim * cim) + | 64 | call_ref(cdst, src1, src2, LEN); |
| 65 | fabs(tre * cim) + fabs(tim * cre) + | 65 | call_new(odst, src1, src2, LEN); |
| 66 | fabs(tre * cre - tim * cim) + | 66 | for (i = 0; i <= LEN*2; i++) { |
| 67 | fabs(tre * cim + tim * cre) + | 67 | int idx = i & ~1; |
| 68 | fabs(cdst[i]) + 1.0; | 68 | float cre = src2[idx]; |
| 69 | if (!float_near_abs_eps(cdst[i], odst[i], t * 2 * FLT_EPSILON)) { | 69 | float cim = src2[idx + 1]; |
| 70 | fprintf(stderr, "%d: %- .12f - %- .12f = % .12g\n", | 70 | float tre = src1[idx]; |
| 71 | i, cdst[i], odst[i], cdst[i] - odst[i]); | 71 | float tim = src1[idx + 1]; |
| 72 | fail(); | 72 | double t = fabs(src0[i]) + |
| 73 | break; | 73 | fabs(tre) + fabs(tim) + fabs(cre) + fabs(cim) + |
| 74 | fabs(tre * cre) + fabs(tim * cim) + | ||
| 75 | fabs(tre * cim) + fabs(tim * cre) + | ||
| 76 | fabs(tre * cre - tim * cim) + | ||
| 77 | fabs(tre * cim + tim * cre) + | ||
| 78 | fabs(cdst[i]) + 1.0; | ||
| 79 | if (!float_near_abs_eps(cdst[i], odst[i], t * 2 * FLT_EPSILON)) { | ||
| 80 | fprintf(stderr, "%d: %- .12f - %- .12f = % .12g\n", | ||
| 81 | i, cdst[i], odst[i], cdst[i] - odst[i]); | ||
| 82 | fail(); | ||
| 83 | break; | ||
| 84 | } | ||
| 74 | } | 85 | } |
| 86 | memcpy(odst, src0, (BUF_SIZE) * sizeof(float)); | ||
| 87 | bench_new(odst, src1, src2, LEN); | ||
| 75 | } | 88 | } |
| 76 | memcpy(odst, src0, (LEN*2+8) * sizeof(float)); | 89 | |
| 77 | bench_new(odst, src1, src2, LEN); | 90 | report("fcmul_add"); |
| 78 | } | 91 | } |
| 79 | 92 | ||
| 80 | void checkasm_check_afir(void) | 93 | static void test_dcmul_add(AudioFIRDSPContext *fir) |
| 81 | { | 94 | { |
| 82 | LOCAL_ALIGNED_32(float, src0, [LEN*2+8]); | 95 | #define BUF_SIZE LEN*2+8 |
| 83 | LOCAL_ALIGNED_32(float, src1, [LEN*2+8]); | 96 | LOCAL_ALIGNED_32(double, src0, [BUF_SIZE]); |
| 84 | LOCAL_ALIGNED_32(float, src2, [LEN*2+8]); | 97 | LOCAL_ALIGNED_32(double, src1, [BUF_SIZE]); |
| 85 | AudioFIRDSPContext fir = { 0 }; | 98 | LOCAL_ALIGNED_32(double, src2, [BUF_SIZE]); |
| 86 | |||
| 87 | ff_afir_init(&fir); | ||
| 88 | 99 | ||
| 89 | randomize_buffer(src0); | 100 | randomize_buffer(src0); |
| 90 | randomize_buffer(src1); | 101 | randomize_buffer(src1); |
| 91 | randomize_buffer(src2); | 102 | randomize_buffer(src2); |
| 92 | 103 | ||
| 93 | if (check_func(fir.fcmul_add, "fcmul_add")) | 104 | if (check_func(fir->dcmul_add, "dcmul_add")) { |
| 94 | test_fcmul_add(src0, src1, src2); | 105 | LOCAL_ALIGNED_32(double, cdst, [BUF_SIZE]); |
| 95 | report("fcmul_add"); | 106 | LOCAL_ALIGNED_32(double, odst, [BUF_SIZE]); |
| 107 | int i; | ||
| 108 | |||
| 109 | declare_func(void, double *sum, const double *t, const double *c, | ||
| 110 | ptrdiff_t len); | ||
| 111 | |||
| 112 | memcpy(cdst, src0, (BUF_SIZE) * sizeof(double)); | ||
| 113 | memcpy(odst, src0, (BUF_SIZE) * sizeof(double)); | ||
| 114 | call_ref(cdst, src1, src2, LEN); | ||
| 115 | call_new(odst, src1, src2, LEN); | ||
| 116 | for (i = 0; i <= LEN*2; i++) { | ||
| 117 | int idx = i & ~1; | ||
| 118 | double cre = src2[idx]; | ||
| 119 | double cim = src2[idx + 1]; | ||
| 120 | double tre = src1[idx]; | ||
| 121 | double tim = src1[idx + 1]; | ||
| 122 | double t = fabs(src0[i]) + | ||
| 123 | fabs(tre) + fabs(tim) + fabs(cre) + fabs(cim) + | ||
| 124 | fabs(tre * cre) + fabs(tim * cim) + | ||
| 125 | fabs(tre * cim) + fabs(tim * cre) + | ||
| 126 | fabs(tre * cre - tim * cim) + | ||
| 127 | fabs(tre * cim + tim * cre) + | ||
| 128 | fabs(cdst[i]) + 1.0; | ||
| 129 | if (!double_near_abs_eps(cdst[i], odst[i], t * 2 * FLT_EPSILON)) { | ||
| 130 | fprintf(stderr, "%d: %- .12f - %- .12f = % .12g\n", | ||
| 131 | i, cdst[i], odst[i], cdst[i] - odst[i]); | ||
| 132 | fail(); | ||
| 133 | break; | ||
| 134 | } | ||
| 135 | } | ||
| 136 | memcpy(odst, src0, (BUF_SIZE) * sizeof(double)); | ||
| 137 | bench_new(odst, src1, src2, LEN); | ||
| 138 | } | ||
| 139 | |||
| 140 | report("dcmul_add"); | ||
| 141 | } | ||
| 142 | |||
| 143 | |||
| 144 | void checkasm_check_afir(void) | ||
| 145 | { | ||
| 146 | AudioFIRDSPContext fir = { 0 }; | ||
| 147 | |||
| 148 | ff_afir_init(&fir); | ||
| 149 | test_fcmul_add(&fir); | ||
| 150 | test_dcmul_add(&fir); | ||
| 96 | } | 151 | } |