| OLD | NEW |
| 1 /* Copyright (c) 2008-2011 Octasic Inc. | 1 /* Copyright (c) 2008-2011 Octasic Inc. |
| 2 Written by Jean-Marc Valin */ | 2 Written by Jean-Marc Valin */ |
| 3 /* | 3 /* |
| 4 Redistribution and use in source and binary forms, with or without | 4 Redistribution and use in source and binary forms, with or without |
| 5 modification, are permitted provided that the following conditions | 5 modification, are permitted provided that the following conditions |
| 6 are met: | 6 are met: |
| 7 | 7 |
| 8 - Redistributions of source code must retain the above copyright | 8 - Redistributions of source code must retain the above copyright |
| 9 notice, this list of conditions and the following disclaimer. | 9 notice, this list of conditions and the following disclaimer. |
| 10 | 10 |
| (...skipping 23 matching lines...) Expand all Loading... |
| 34 | 34 |
| 35 #include <math.h> | 35 #include <math.h> |
| 36 #include "mlp.h" | 36 #include "mlp.h" |
| 37 #include "arch.h" | 37 #include "arch.h" |
| 38 #include "tansig_table.h" | 38 #include "tansig_table.h" |
| 39 #define MAX_NEURONS 100 | 39 #define MAX_NEURONS 100 |
| 40 | 40 |
| 41 #if 0 | 41 #if 0 |
| 42 static OPUS_INLINE opus_val16 tansig_approx(opus_val32 _x) /* Q19 */ | 42 static OPUS_INLINE opus_val16 tansig_approx(opus_val32 _x) /* Q19 */ |
| 43 { | 43 { |
| 44 » int i; | 44 int i; |
| 45 » opus_val16 xx; /* Q11 */ | 45 opus_val16 xx; /* Q11 */ |
| 46 » /*double x, y;*/ | 46 /*double x, y;*/ |
| 47 » opus_val16 dy, yy; /* Q14 */ | 47 opus_val16 dy, yy; /* Q14 */ |
| 48 » /*x = 1.9073e-06*_x;*/ | 48 /*x = 1.9073e-06*_x;*/ |
| 49 » if (_x>=QCONST32(8,19)) | 49 if (_x>=QCONST32(8,19)) |
| 50 » » return QCONST32(1.,14); | 50 return QCONST32(1.,14); |
| 51 » if (_x<=-QCONST32(8,19)) | 51 if (_x<=-QCONST32(8,19)) |
| 52 » » return -QCONST32(1.,14); | 52 return -QCONST32(1.,14); |
| 53 » xx = EXTRACT16(SHR32(_x, 8)); | 53 xx = EXTRACT16(SHR32(_x, 8)); |
| 54 » /*i = lrint(25*x);*/ | 54 /*i = lrint(25*x);*/ |
| 55 » i = SHR32(ADD32(1024,MULT16_16(25, xx)),11); | 55 i = SHR32(ADD32(1024,MULT16_16(25, xx)),11); |
| 56 » /*x -= .04*i;*/ | 56 /*x -= .04*i;*/ |
| 57 » xx -= EXTRACT16(SHR32(MULT16_16(20972,i),8)); | 57 xx -= EXTRACT16(SHR32(MULT16_16(20972,i),8)); |
| 58 » /*x = xx*(1./2048);*/ | 58 /*x = xx*(1./2048);*/ |
| 59 » /*y = tansig_table[250+i];*/ | 59 /*y = tansig_table[250+i];*/ |
| 60 » yy = tansig_table[250+i]; | 60 yy = tansig_table[250+i]; |
| 61 » /*y = yy*(1./16384);*/ | 61 /*y = yy*(1./16384);*/ |
| 62 » dy = 16384-MULT16_16_Q14(yy,yy); | 62 dy = 16384-MULT16_16_Q14(yy,yy); |
| 63 » yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx,dy),(16384 - MULT16_16_Q11(yy,x
x))); | 63 yy = yy + MULT16_16_Q14(MULT16_16_Q11(xx,dy),(16384 - MULT16_16_Q11(yy,xx)))
; |
| 64 » return yy; | 64 return yy; |
| 65 } | 65 } |
| 66 #else | 66 #else |
| 67 /*extern const float tansig_table[501];*/ | 67 /*extern const float tansig_table[501];*/ |
| 68 static OPUS_INLINE float tansig_approx(float x) | 68 static OPUS_INLINE float tansig_approx(float x) |
| 69 { | 69 { |
| 70 » int i; | 70 int i; |
| 71 » float y, dy; | 71 float y, dy; |
| 72 » float sign=1; | 72 float sign=1; |
| 73 » /* Tests are reversed to catch NaNs */ | 73 /* Tests are reversed to catch NaNs */ |
| 74 if (!(x<8)) | 74 if (!(x<8)) |
| 75 return 1; | 75 return 1; |
| 76 if (!(x>-8)) | 76 if (!(x>-8)) |
| 77 return -1; | 77 return -1; |
| 78 » if (x<0) | 78 #ifndef FIXED_POINT |
| 79 » { | 79 /* Another check in case of -ffast-math */ |
| 80 » x=-x; | 80 if (celt_isnan(x)) |
| 81 » sign=-1; | 81 return 0; |
| 82 » } | 82 #endif |
| 83 » i = (int)floor(.5f+25*x); | 83 if (x<0) |
| 84 » x -= .04f*i; | 84 { |
| 85 » y = tansig_table[i]; | 85 x=-x; |
| 86 » dy = 1-y*y; | 86 sign=-1; |
| 87 » y = y + x*dy*(1 - y*x); | 87 } |
| 88 » return sign*y; | 88 i = (int)floor(.5f+25*x); |
| 89 x -= .04f*i; |
| 90 y = tansig_table[i]; |
| 91 dy = 1-y*y; |
| 92 y = y + x*dy*(1 - y*x); |
| 93 return sign*y; |
| 89 } | 94 } |
| 90 #endif | 95 #endif |
| 91 | 96 |
| 92 #if 0 | 97 #if 0 |
| 93 void mlp_process(const MLP *m, const opus_val16 *in, opus_val16 *out) | 98 void mlp_process(const MLP *m, const opus_val16 *in, opus_val16 *out) |
| 94 { | 99 { |
| 95 » int j; | 100 int j; |
| 96 » opus_val16 hidden[MAX_NEURONS]; | 101 opus_val16 hidden[MAX_NEURONS]; |
| 97 » const opus_val16 *W = m->weights; | 102 const opus_val16 *W = m->weights; |
| 98 » /* Copy to tmp_in */ | 103 /* Copy to tmp_in */ |
| 99 » for (j=0;j<m->topo[1];j++) | 104 for (j=0;j<m->topo[1];j++) |
| 100 » { | 105 { |
| 101 » » int k; | 106 int k; |
| 102 » » opus_val32 sum = SHL32(EXTEND32(*W++),8); | 107 opus_val32 sum = SHL32(EXTEND32(*W++),8); |
| 103 » » for (k=0;k<m->topo[0];k++) | 108 for (k=0;k<m->topo[0];k++) |
| 104 » » » sum = MAC16_16(sum, in[k],*W++); | 109 sum = MAC16_16(sum, in[k],*W++); |
| 105 » » hidden[j] = tansig_approx(sum); | 110 hidden[j] = tansig_approx(sum); |
| 106 » } | 111 } |
| 107 » for (j=0;j<m->topo[2];j++) | 112 for (j=0;j<m->topo[2];j++) |
| 108 » { | 113 { |
| 109 » » int k; | 114 int k; |
| 110 » » opus_val32 sum = SHL32(EXTEND32(*W++),14); | 115 opus_val32 sum = SHL32(EXTEND32(*W++),14); |
| 111 » » for (k=0;k<m->topo[1];k++) | 116 for (k=0;k<m->topo[1];k++) |
| 112 » » » sum = MAC16_16(sum, hidden[k], *W++); | 117 sum = MAC16_16(sum, hidden[k], *W++); |
| 113 » » out[j] = tansig_approx(EXTRACT16(PSHR32(sum,17))); | 118 out[j] = tansig_approx(EXTRACT16(PSHR32(sum,17))); |
| 114 » } | 119 } |
| 115 } | 120 } |
| 116 #else | 121 #else |
| 117 void mlp_process(const MLP *m, const float *in, float *out) | 122 void mlp_process(const MLP *m, const float *in, float *out) |
| 118 { | 123 { |
| 119 int j; | 124 int j; |
| 120 float hidden[MAX_NEURONS]; | 125 float hidden[MAX_NEURONS]; |
| 121 const float *W = m->weights; | 126 const float *W = m->weights; |
| 122 /* Copy to tmp_in */ | 127 /* Copy to tmp_in */ |
| 123 for (j=0;j<m->topo[1];j++) | 128 for (j=0;j<m->topo[1];j++) |
| 124 { | 129 { |
| 125 int k; | 130 int k; |
| 126 float sum = *W++; | 131 float sum = *W++; |
| 127 for (k=0;k<m->topo[0];k++) | 132 for (k=0;k<m->topo[0];k++) |
| 128 sum = sum + in[k]**W++; | 133 sum = sum + in[k]**W++; |
| 129 hidden[j] = tansig_approx(sum); | 134 hidden[j] = tansig_approx(sum); |
| 130 } | 135 } |
| 131 for (j=0;j<m->topo[2];j++) | 136 for (j=0;j<m->topo[2];j++) |
| 132 { | 137 { |
| 133 int k; | 138 int k; |
| 134 float sum = *W++; | 139 float sum = *W++; |
| 135 for (k=0;k<m->topo[1];k++) | 140 for (k=0;k<m->topo[1];k++) |
| 136 sum = sum + hidden[k]**W++; | 141 sum = sum + hidden[k]**W++; |
| 137 out[j] = tansig_approx(sum); | 142 out[j] = tansig_approx(sum); |
| 138 } | 143 } |
| 139 } | 144 } |
| 140 #endif | 145 #endif |
| OLD | NEW |