Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-raddstoreexpminusmax/scalar-p5.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2020 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <xnnpack/common.h> |
| 13 | #include <xnnpack/raddstoreexpminusmax.h> |
| 14 | |
| 15 | #include <fp16/bitcasts.h> |
| 16 | |
| 17 | |
| 18 | void xnn_f32_raddstoreexpminusmax_ukernel__scalar_p5_x4_acc2( |
| 19 | size_t elements, |
| 20 | const float* input, |
| 21 | float* output, |
| 22 | float* sum, |
| 23 | float vi_max) |
| 24 | { |
| 25 | assert(elements % sizeof(float) == 0); |
| 26 | |
| 27 | const float vmagic_bias = 0x1.8000FEp23f; |
| 28 | // The smallest x for which expf(x) is normalized. |
| 29 | const float vdenorm_cutoff = -0x1.5D589Ep6f; |
| 30 | const float vlog2e = 0x1.715476p+0f; |
| 31 | // Last 7 bits are zeroes |
| 32 | const float vminus_ln2_hi = -0x1.62E400p-1f; |
| 33 | const float vminus_ln2_lo = -0x1.7F7D1Cp-20f; |
| 34 | |
| 35 | const float vc1 = 0x1.FFFFF6p-1f; |
| 36 | const float vc2 = 0x1.FFFDC6p-2f; |
| 37 | const float vc3 = 0x1.555A80p-3f; |
| 38 | const float vc4 = 0x1.573A1Ap-5f; |
| 39 | const float vc5 = 0x1.0F9F9Cp-7f; |
| 40 | |
| 41 | float vacc0 = 0.0f; |
| 42 | float vacc1 = 0.0f; |
| 43 | for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) { |
| 44 | // Load 4 inputs at a time. |
| 45 | const float vi0 = input[0]; |
| 46 | const float vi1 = input[1]; |
| 47 | const float vi2 = input[2]; |
| 48 | const float vi3 = input[3]; |
| 49 | input += 4; |
| 50 | |
| 51 | // Subtract maximum input x := i - i_max. This implies x <= 0. |
| 52 | const float vx0 = vi0 - vi_max; |
| 53 | const float vx1 = vi1 - vi_max; |
| 54 | const float vx2 = vi2 - vi_max; |
| 55 | const float vx3 = vi3 - vi_max; |
| 56 | |
| 57 | // Compute reduced argument n := round(x / log(2)). |
| 58 | // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the result |
| 59 | // to an integer, then subtracing the large number back. The trick with adding large number is valid only within |
| 60 | // certain bounds (|x| <= 2**22), but thats ok, because inputs outside of [-87.336540, 0.0] underflow expf(x) |
| 61 | // anyway. We fixup the result for such inputs at the very end of the algorithm. |
| 62 | float vn0 = vx0 * vlog2e + vmagic_bias; |
| 63 | float vn1 = vx1 * vlog2e + vmagic_bias; |
| 64 | float vn2 = vx2 * vlog2e + vmagic_bias; |
| 65 | float vn3 = vx3 * vlog2e + vmagic_bias; |
| 66 | |
| 67 | // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e. |
| 68 | // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly. |
| 69 | const float vs0 = fp32_from_bits(fp32_to_bits(vn0) << 23); |
| 70 | const float vs1 = fp32_from_bits(fp32_to_bits(vn1) << 23); |
| 71 | const float vs2 = fp32_from_bits(fp32_to_bits(vn2) << 23); |
| 72 | const float vs3 = fp32_from_bits(fp32_to_bits(vn3) << 23); |
| 73 | |
| 74 | // Subtract the large number back to get final n := round(x / log(2)). |
| 75 | vn0 -= vmagic_bias; |
| 76 | vn1 -= vmagic_bias; |
| 77 | vn2 -= vmagic_bias; |
| 78 | vn3 -= vmagic_bias; |
| 79 | |
| 80 | // Compute reduced argument t := x - n * log(2). |
| 81 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 82 | float vt0 = vn0 * vminus_ln2_hi + vx0; |
| 83 | float vt1 = vn1 * vminus_ln2_hi + vx1; |
| 84 | float vt2 = vn2 * vminus_ln2_hi + vx2; |
| 85 | float vt3 = vn3 * vminus_ln2_hi + vx3; |
| 86 | |
| 87 | vt0 = vn0 * vminus_ln2_lo + vt0; |
| 88 | vt1 = vn1 * vminus_ln2_lo + vt1; |
| 89 | vt2 = vn2 * vminus_ln2_lo + vt2; |
| 90 | vt3 = vn3 * vminus_ln2_lo + vt3; |
| 91 | |
Marat Dukhan | 102a739 | 2020-11-20 01:18:10 -0800 | [diff] [blame^] | 92 | // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 93 | float vp0 = vc5 * vt0 + vc4; |
| 94 | float vp1 = vc5 * vt1 + vc4; |
| 95 | float vp2 = vc5 * vt2 + vc4; |
| 96 | float vp3 = vc5 * vt3 + vc4; |
| 97 | |
| 98 | vp0 = vp0 * vt0 + vc3; |
| 99 | vp1 = vp1 * vt1 + vc3; |
| 100 | vp2 = vp2 * vt2 + vc3; |
| 101 | vp3 = vp3 * vt3 + vc3; |
| 102 | |
| 103 | vp0 = vp0 * vt0 + vc2; |
| 104 | vp1 = vp1 * vt1 + vc2; |
| 105 | vp2 = vp2 * vt2 + vc2; |
| 106 | vp3 = vp3 * vt3 + vc2; |
| 107 | |
| 108 | vp0 = vp0 * vt0 + vc1; |
| 109 | vp1 = vp1 * vt1 + vc1; |
| 110 | vp2 = vp2 * vt2 + vc1; |
| 111 | vp3 = vp3 * vt3 + vc1; |
| 112 | |
| 113 | // Reconstruct the final f value: |
| 114 | // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 115 | // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 116 | // = s + (t * s) * p |
| 117 | vt0 *= vs0; |
| 118 | vt1 *= vs1; |
| 119 | vt2 *= vs2; |
| 120 | vt3 *= vs3; |
| 121 | |
| 122 | float vf0 = vt0 * vp0 + vs0; |
| 123 | float vf1 = vt1 * vp1 + vs1; |
| 124 | float vf2 = vt2 * vp2 + vs2; |
| 125 | float vf3 = vt3 * vp3 + vs3; |
| 126 | |
| 127 | // For inputs below denormal cutoff, replace output with +0.0f. |
| 128 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 129 | if XNN_UNPREDICTABLE(vx0 < vdenorm_cutoff) { |
| 130 | vf0 = 0.0f; |
| 131 | } |
| 132 | if XNN_UNPREDICTABLE(vx1 < vdenorm_cutoff) { |
| 133 | vf1 = 0.0f; |
| 134 | } |
| 135 | if XNN_UNPREDICTABLE(vx2 < vdenorm_cutoff) { |
| 136 | vf2 = 0.0f; |
| 137 | } |
| 138 | if XNN_UNPREDICTABLE(vx3 < vdenorm_cutoff) { |
| 139 | vf3 = 0.0f; |
| 140 | } |
| 141 | |
| 142 | // Store 4 outputs at a time. |
| 143 | output[0] = vf0; |
| 144 | output[1] = vf1; |
| 145 | output[2] = vf2; |
| 146 | output[3] = vf3; |
| 147 | output += 4; |
| 148 | |
| 149 | // Accumulate computed exponents. |
| 150 | vacc0 += vf0; |
| 151 | vacc1 += vf1; |
| 152 | vacc0 += vf2; |
| 153 | vacc1 += vf3; |
| 154 | } |
| 155 | // Add up all accumulators to vacc0 |
| 156 | vacc0 += vacc1; |
| 157 | |
| 158 | float vacc = vacc0; |
| 159 | for (; elements >= sizeof(float); elements -= sizeof(float)) { |
| 160 | // Load 1 input at a time. |
| 161 | const float vi = *input++; |
| 162 | |
| 163 | // Subtract maximum input x := i - i_max. This implies x <= 0. |
| 164 | const float vx = vi - vi_max; |
| 165 | |
| 166 | // Compute reduced argument n := round(x / log(2)). |
| 167 | // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the result |
| 168 | // to an integer, then subtracing the large number back. The trick with adding large number is valid only within |
| 169 | // certain bounds (|x| <= 2**22), but thats ok, because inputs outside of [-87.336540, 0.0] underflow expf(x) |
| 170 | // anyway. We fixup the result for such inputs at the very end of the algorithm. |
| 171 | float vn = vx * vlog2e + vmagic_bias; |
| 172 | |
| 173 | // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e. |
| 174 | // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly. |
| 175 | const float vs = fp32_from_bits(fp32_to_bits(vn) << 23); |
| 176 | |
| 177 | // Subtract the large number back to get final n := round(x / log(2)). |
| 178 | vn -= vmagic_bias; |
| 179 | |
| 180 | // Compute reduced argument t := x - n * log(2). |
| 181 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 182 | float vt = vn * vminus_ln2_hi + vx; |
| 183 | vt = vn * vminus_ln2_lo + vt; |
| 184 | |
Marat Dukhan | 102a739 | 2020-11-20 01:18:10 -0800 | [diff] [blame^] | 185 | // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
Marat Dukhan | f46f675 | 2020-01-21 11:03:49 -0800 | [diff] [blame] | 186 | float vp = vc5 * vt + vc4; |
| 187 | vp = vp * vt + vc3; |
| 188 | vp = vp * vt + vc2; |
| 189 | vp = vp * vt + vc1; |
| 190 | |
| 191 | // Reconstruct the final f value: |
| 192 | // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 193 | // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 194 | // = s + (t * s) * p |
| 195 | vt *= vs; |
| 196 | float vf = vt * vp + vs; |
| 197 | |
| 198 | // For inputs below denormal cutoff, replace output with +0.0f. |
| 199 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 200 | if XNN_UNPREDICTABLE(vx < vdenorm_cutoff) { |
| 201 | vf = 0.0f; |
| 202 | } |
| 203 | |
| 204 | // Store 1 output at a time. |
| 205 | *output++ = vf; |
| 206 | |
| 207 | // Accumulate computed exponents. |
| 208 | vacc += vf; |
| 209 | } |
| 210 | *sum = vacc; |
| 211 | } |