Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-raddstoreexpminusmax/wasmsimd-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 <wasm_simd128.h> |
| 13 | |
| 14 | #include <xnnpack/common.h> |
| 15 | #include <xnnpack/raddstoreexpminusmax.h> |
| 16 | |
| 17 | |
| 18 | void xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_p5_x8_acc2( |
| 19 | size_t elements, |
| 20 | const float* input, |
| 21 | float* output, |
| 22 | float* sum, |
| 23 | float max) XNN_DISABLE_TSAN |
| 24 | { |
| 25 | assert(elements % sizeof(float) == 0); |
| 26 | |
| 27 | const v128_t vmagic_bias = wasm_f32x4_splat(0x1.8000FEp23f); |
| 28 | // The smallest x for which expf(x) is normalized. |
| 29 | const v128_t vdenorm_cutoff = wasm_f32x4_splat(-0x1.5D589Ep6f); |
| 30 | const v128_t vlog2e = wasm_f32x4_splat(0x1.715476p+0f); |
| 31 | // Last 7 bits are zeroes |
| 32 | const v128_t vminus_ln2_hi = wasm_f32x4_splat(-0x1.62E400p-1f); |
| 33 | const v128_t vminus_ln2_lo = wasm_f32x4_splat(-0x1.7F7D1Cp-20f); |
| 34 | |
| 35 | const v128_t vc1 = wasm_f32x4_splat(0x1.FFFFF6p-1f); |
| 36 | const v128_t vc2 = wasm_f32x4_splat(0x1.FFFDC6p-2f); |
| 37 | const v128_t vc3 = wasm_f32x4_splat(0x1.555A80p-3f); |
| 38 | const v128_t vc4 = wasm_f32x4_splat(0x1.573A1Ap-5f); |
| 39 | const v128_t vc5 = wasm_f32x4_splat(0x1.0F9F9Cp-7f); |
| 40 | |
| 41 | const v128_t vi_max = wasm_f32x4_splat(max); |
| 42 | |
| 43 | v128_t vacc0 = wasm_f32x4_splat(0.0f); |
| 44 | v128_t vacc1 = vacc0; |
| 45 | for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) { |
| 46 | // Load 8 (2x4) inputs at a time. |
| 47 | const v128_t vi0123 = wasm_v128_load(input); |
| 48 | const v128_t vi4567 = wasm_v128_load(input + 4); |
| 49 | input += 8; |
| 50 | |
| 51 | // Subtract maximum input x := i - i_max. This implies x <= 0. |
| 52 | const v128_t vx0123 = wasm_f32x4_sub(vi0123, vi_max); |
| 53 | const v128_t vx4567 = wasm_f32x4_sub(vi4567, vi_max); |
| 54 | |
| 55 | // Compute reduced argument elements := round(x / log(2)). |
| 56 | v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx0123, vlog2e)); |
| 57 | v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx4567, vlog2e)); |
| 58 | |
| 59 | // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. |
| 60 | // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. |
| 61 | const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23); |
| 62 | const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23); |
| 63 | |
| 64 | // Subtract the large number back to get final elements := round(x / log(2)). |
| 65 | vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias); |
| 66 | vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias); |
| 67 | |
| 68 | // Compute reduced argument t := x - elements * log(2). |
| 69 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 70 | v128_t vt0123 = wasm_f32x4_add(vx0123, wasm_f32x4_mul(vn0123, vminus_ln2_hi)); |
| 71 | v128_t vt4567 = wasm_f32x4_add(vx4567, wasm_f32x4_mul(vn4567, vminus_ln2_hi)); |
| 72 | |
| 73 | vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vminus_ln2_lo)); |
| 74 | vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vminus_ln2_lo)); |
| 75 | |
Marat Dukhan | 102a739 | 2020-11-20 01:18:10 -0800 | [diff] [blame^] | 76 | // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 77 | v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt0123)); |
| 78 | v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt4567)); |
| 79 | |
| 80 | vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp0123, vt0123)); |
| 81 | vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp4567, vt4567)); |
| 82 | |
| 83 | vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp0123, vt0123)); |
| 84 | vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp4567, vt4567)); |
| 85 | |
| 86 | vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp0123, vt0123)); |
| 87 | vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp4567, vt4567)); |
| 88 | |
| 89 | // Reconstruct the final f value: |
| 90 | // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 91 | // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 92 | // = s + (t * s) * p |
| 93 | vt0123 = wasm_f32x4_mul(vt0123, vs0123); |
| 94 | vt4567 = wasm_f32x4_mul(vt4567, vs4567); |
| 95 | |
| 96 | v128_t vf0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123)); |
| 97 | v128_t vf4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567)); |
| 98 | |
| 99 | // For inputs below zero cutoff, replace output with +0.0f. |
| 100 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 101 | vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_lt(vx0123, vdenorm_cutoff)); |
| 102 | vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_lt(vx4567, vdenorm_cutoff)); |
| 103 | |
| 104 | // Store 8 (2x4) outputs at a time. |
| 105 | wasm_v128_store(output, vf0123); |
| 106 | wasm_v128_store(output + 4, vf4567); |
| 107 | output += 8; |
| 108 | |
| 109 | // Accumulate computed exponents. |
| 110 | vacc0 = wasm_f32x4_add(vacc0, vf0123); |
| 111 | vacc0 = wasm_f32x4_add(vacc0, vf4567); |
| 112 | } |
| 113 | // Add up all accumulators to vacc0 |
| 114 | vacc0 = wasm_f32x4_add(vacc0, vacc1); |
| 115 | |
| 116 | v128_t vacc = vacc0; |
| 117 | for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) { |
| 118 | // Load 4 inputs at a time. |
| 119 | const v128_t vi = wasm_v128_load(input); |
| 120 | input += 4; |
| 121 | |
| 122 | // Subtract maximum input x := i - i_max. This implies x <= 0. |
| 123 | const v128_t vx = wasm_f32x4_sub(vi, vi_max); |
| 124 | |
| 125 | // Compute reduced argument elements := round(x / log(2)). |
| 126 | v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e)); |
| 127 | |
| 128 | // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. |
| 129 | // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. |
| 130 | const v128_t vs = wasm_i32x4_shl(vn, 23); |
| 131 | |
| 132 | // Subtract the large number back to get final elements := round(x / log(2)). |
| 133 | vn = wasm_f32x4_sub(vn, vmagic_bias); |
| 134 | |
| 135 | // Compute reduced argument t := x - elements * log(2). |
| 136 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 137 | v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi)); |
| 138 | vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo)); |
| 139 | |
Marat Dukhan | 102a739 | 2020-11-20 01:18:10 -0800 | [diff] [blame^] | 140 | // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 141 | v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt)); |
| 142 | vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt)); |
| 143 | vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt)); |
| 144 | vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt)); |
| 145 | |
| 146 | // Reconstruct the final f value: |
| 147 | // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 148 | // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 149 | // = s + (t * s) * p |
| 150 | vt = wasm_f32x4_mul(vt, vs); |
| 151 | v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp)); |
| 152 | |
| 153 | // For inputs below zero cutoff, replace output with +0.0f. |
| 154 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 155 | vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff)); |
| 156 | |
| 157 | // Store 4 outputs at a time. |
| 158 | wasm_v128_store(output, vf); |
| 159 | output += 4; |
| 160 | |
| 161 | // Accumulate computed exponents. |
| 162 | vacc = wasm_f32x4_add(vacc, vf); |
| 163 | } |
| 164 | vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3)); |
| 165 | float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1); |
| 166 | if (elements != 0) { |
| 167 | assert(elements >= 1 * sizeof(float)); |
| 168 | assert(elements <= 3 * sizeof(float)); |
| 169 | // Load 4 inputs at a time. |
| 170 | const v128_t vi = wasm_v128_load(input); |
| 171 | |
| 172 | // Subtract maximum input x := i - i_max. This implies x <= 0. |
| 173 | const v128_t vx = wasm_f32x4_sub(vi, vi_max); |
| 174 | |
| 175 | // Compute reduced argument elements := round(x / log(2)). |
| 176 | v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e)); |
| 177 | |
| 178 | // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e. |
| 179 | // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly. |
| 180 | const v128_t vs = wasm_i32x4_shl(vn, 23); |
| 181 | |
| 182 | // Subtract the large number back to get final elements := round(x / log(2)). |
| 183 | vn = wasm_f32x4_sub(vn, vmagic_bias); |
| 184 | |
| 185 | // Compute reduced argument t := x - elements * log(2). |
| 186 | // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 187 | v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi)); |
| 188 | vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo)); |
| 189 | |
Marat Dukhan | 102a739 | 2020-11-20 01:18:10 -0800 | [diff] [blame^] | 190 | // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
Marat Dukhan | 52238f0 | 2020-07-16 15:30:28 -0700 | [diff] [blame] | 191 | v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt)); |
| 192 | vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt)); |
| 193 | vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt)); |
| 194 | vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt)); |
| 195 | |
| 196 | // Reconstruct the final f value: |
| 197 | // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 198 | // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 199 | // = s + (t * s) * p |
| 200 | vt = wasm_f32x4_mul(vt, vs); |
| 201 | v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp)); |
| 202 | |
| 203 | // For inputs below zero cutoff, replace output with +0.0f. |
| 204 | // Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 205 | vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff)); |
| 206 | |
| 207 | if (elements & (2 * sizeof(float))) { |
| 208 | // Store and accumulate 2 outputs at a time. |
| 209 | const float vf0 = wasm_f32x4_extract_lane(vf, 0); |
| 210 | output[0] = vf0; |
| 211 | vsum += vf0; |
| 212 | |
| 213 | const float vf1 = wasm_f32x4_extract_lane(vf, 1); |
| 214 | output[1] = vf1; |
| 215 | vsum += vf1; |
| 216 | |
| 217 | vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3); |
| 218 | output += 2; |
| 219 | } |
| 220 | if (elements & (1 * sizeof(float))) { |
| 221 | // Store 1 output at a time. |
| 222 | const float vf0 = wasm_f32x4_extract_lane(vf, 0); |
| 223 | *output = vf0; |
| 224 | vsum += vf0; |
| 225 | } |
| 226 | } |
| 227 | // Reduce 4 elements in the SIMD register |
| 228 | *sum = vsum; |
| 229 | } |