xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/wasmsimd-rr2-p5.c.in (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1// Copyright 2020 Google LLC
2//
3// This source code is licensed under the BSD-style license found in the
4// LICENSE file in the root directory of this source tree.
5
6$assert ELEMENTS_TILE % 4 == 0
7$assert ELEMENTS_TILE >= 4
8$SIMD_TILE = ELEMENTS_TILE // 4
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <wasm_simd128.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/raddstoreexpminusmax.h>
16
17
18void xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x${ELEMENTS_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
19    size_t elements,
20    const float* input,
21    const float* max,
22    float* output,
23    float* sum,
24    const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25{
26  assert(elements % sizeof(float) == 0);
27
28  const v128_t vi_max = wasm_v128_load32_splat(max);
29  const v128_t vlog2e = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.log2e);
30  const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.magic_bias);
31  const v128_t vminus_ln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_ln2_hi);
32  const v128_t vminus_ln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_ln2_lo);
33  const v128_t vc5 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c5);
34  const v128_t vc4 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c4);
35  const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c3);
36  const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c2);
37  const v128_t vc1 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c1);
38  const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.denorm_cutoff);
39
40  v128_t vacc0 = wasm_f32x4_const_splat(0.0f);
41  $for K in range(1, ACCUMULATORS):
42    v128_t vacc${K} = vacc0;
43  for (; elements >= ${ELEMENTS_TILE} * sizeof(float); elements -= ${ELEMENTS_TILE} * sizeof(float)) {
44    // Load ${ELEMENTS_TILE} (${SIMD_TILE}x4) inputs at a time.
45    const v128_t vi${ABC[0:4]} = wasm_v128_load(input);
46    $for N in range(4, ELEMENTS_TILE, 4):
47      const v128_t vi${ABC[N:N+4]} = wasm_v128_load(input + ${N});
48    input += ${ELEMENTS_TILE};
49
50    // Subtract maximum input x := i - i_max. This implies x <= 0.
51    $for N in range(0, ELEMENTS_TILE, 4):
52      const v128_t vx${ABC[N:N+4]} = wasm_f32x4_sub(vi${ABC[N:N+4]}, vi_max);
53
54    // Compute reduced argument elements := round(x / log(2)).
55    $for N in range(0, ELEMENTS_TILE, 4):
56      v128_t vn${ABC[N:N+4]} = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx${ABC[N:N+4]}, vlog2e));
57
58    // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
59    // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
60    $for N in range(0, ELEMENTS_TILE, 4):
61      const v128_t vs${ABC[N:N+4]} = wasm_i32x4_shl(vn${ABC[N:N+4]}, 23);
62
63    // Subtract the large number back to get final elements := round(x / log(2)).
64    $for N in range(0, ELEMENTS_TILE, 4):
65      vn${ABC[N:N+4]} = wasm_f32x4_sub(vn${ABC[N:N+4]}, vmagic_bias);
66
67    // Compute reduced argument t := x - elements * log(2).
68    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
69    $for N in range(0, ELEMENTS_TILE, 4):
70      v128_t vt${ABC[N:N+4]} = wasm_f32x4_add(vx${ABC[N:N+4]}, wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_hi));
71
72    $for N in range(0, ELEMENTS_TILE, 4):
73      vt${ABC[N:N+4]} = wasm_f32x4_add(vt${ABC[N:N+4]}, wasm_f32x4_mul(vn${ABC[N:N+4]}, vminus_ln2_lo));
74
75    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
76    $for N in range(0, ELEMENTS_TILE, 4):
77      v128_t vp${ABC[N:N+4]} = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt${ABC[N:N+4]}));
78
79    $for N in range(0, ELEMENTS_TILE, 4):
80      vp${ABC[N:N+4]} = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}));
81
82    $for N in range(0, ELEMENTS_TILE, 4):
83      vp${ABC[N:N+4]} = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}));
84
85    $for N in range(0, ELEMENTS_TILE, 4):
86      vp${ABC[N:N+4]} = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}));
87
88    // Reconstruct the final f value:
89    //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
90    //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
91    //     = s + (t * s) * p
92    $for N in range(0, ELEMENTS_TILE, 4):
93      vt${ABC[N:N+4]} = wasm_f32x4_mul(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});
94
95    $for N in range(0, ELEMENTS_TILE, 4):
96      v128_t vf${ABC[N:N+4]} = wasm_f32x4_add(vs${ABC[N:N+4]}, wasm_f32x4_mul(vt${ABC[N:N+4]}, vp${ABC[N:N+4]}));
97
98    // For inputs below zero cutoff, replace output with +0.0f.
99    // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
100    $for N in range(0, ELEMENTS_TILE, 4):
101      vf${ABC[N:N+4]} = wasm_v128_andnot(vf${ABC[N:N+4]}, wasm_f32x4_lt(vx${ABC[N:N+4]}, vdenorm_cutoff));
102
103    // Store ${ELEMENTS_TILE} (${SIMD_TILE}x4) outputs at a time.
104    wasm_v128_store(output, vf${ABC[0:4]});
105    $for N in range(4, ELEMENTS_TILE, 4):
106      wasm_v128_store(output + ${N}, vf${ABC[N:N+4]});
107    output += ${ELEMENTS_TILE};
108
109    // Accumulate computed exponents.
110    $for N in range(0, ELEMENTS_TILE, 4):
111      vacc${N % ACCUMULATORS} = wasm_f32x4_add(vacc${N % ACCUMULATORS}, vf${ABC[N:N+4]});
112  }
113  $if ACCUMULATORS > 1:
114    // Add up all accumulators to vacc0
115    $ACC_SLICE = 1
116    $while ACC_SLICE < ACCUMULATORS:
117      $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
118        $if A + ACC_SLICE < ACCUMULATORS:
119          vacc${A} = wasm_f32x4_add(vacc${A}, vacc${A + ACC_SLICE});
120      $ACC_SLICE *= 2
121
122  v128_t vacc = vacc0;
123  for (; elements >= 4 * sizeof(float); elements -= 4 * sizeof(float)) {
124    // Load 4 inputs at a time.
125    const v128_t vi = wasm_v128_load(input);
126    input += 4;
127
128    // Subtract maximum input x := i - i_max. This implies x <= 0.
129    const v128_t vx = wasm_f32x4_sub(vi, vi_max);
130
131    // Compute reduced argument elements := round(x / log(2)).
132    v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
133
134    // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
135    // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
136    const v128_t vs = wasm_i32x4_shl(vn, 23);
137
138    // Subtract the large number back to get final elements := round(x / log(2)).
139    vn = wasm_f32x4_sub(vn, vmagic_bias);
140
141    // Compute reduced argument t := x - elements * log(2).
142    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
143    v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
144    vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
145
146    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
147    v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
148    vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
149    vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
150    vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
151
152    // Reconstruct the final f value:
153    //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
154    //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
155    //     = s + (t * s) * p
156    vt = wasm_f32x4_mul(vt, vs);
157    v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
158
159    // For inputs below zero cutoff, replace output with +0.0f.
160    // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
161    vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
162
163    // Store 4 outputs at a time.
164    wasm_v128_store(output, vf);
165    output += 4;
166
167    // Accumulate computed exponents.
168    vacc = wasm_f32x4_add(vacc, vf);
169  }
170  vacc = wasm_f32x4_add(vacc, wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3));
171  float vsum = wasm_f32x4_extract_lane(vacc, 0) + wasm_f32x4_extract_lane(vacc, 1);
172  if (elements != 0) {
173    assert(elements >= 1 * sizeof(float));
174    assert(elements <= 3 * sizeof(float));
175    // Load 4 inputs at a time.
176    const v128_t vi = wasm_v128_load(input);
177
178    // Subtract maximum input x := i - i_max. This implies x <= 0.
179    const v128_t vx = wasm_f32x4_sub(vi, vi_max);
180
181    // Compute reduced argument elements := round(x / log(2)).
182    v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vx, vlog2e));
183
184    // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
185    // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
186    const v128_t vs = wasm_i32x4_shl(vn, 23);
187
188    // Subtract the large number back to get final elements := round(x / log(2)).
189    vn = wasm_f32x4_sub(vn, vmagic_bias);
190
191    // Compute reduced argument t := x - elements * log(2).
192    // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
193    v128_t vt = wasm_f32x4_add(vx, wasm_f32x4_mul(vn, vminus_ln2_hi));
194    vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vminus_ln2_lo));
195
196    // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
197    v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vc5, vt));
198    vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vp, vt));
199    vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vp, vt));
200    vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vp, vt));
201
202    // Reconstruct the final f value:
203    //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
204    //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
205    //     = s + (t * s) * p
206    vt = wasm_f32x4_mul(vt, vs);
207    v128_t vf = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
208
209    // For inputs below zero cutoff, replace output with +0.0f.
210    // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
211    vf = wasm_v128_andnot(vf, wasm_f32x4_lt(vx, vdenorm_cutoff));
212
213    if (elements & (2 * sizeof(float))) {
214      // Store and accumulate 2 outputs at a time.
215      const float vf0 = wasm_f32x4_extract_lane(vf, 0);
216      output[0] = vf0;
217      vsum += vf0;
218
219      const float vf1 = wasm_f32x4_extract_lane(vf, 1);
220      output[1] = vf1;
221      vsum += vf1;
222
223      vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
224      output += 2;
225    }
226    if (elements & (1 * sizeof(float))) {
227      // Store 1 output at a time.
228      const float vf0 = wasm_f32x4_extract_lane(vf, 0);
229      *output = vf0;
230      vsum += vf0;
231    }
232  }
233  // Reduce 4 elements in the SIMD register
234  *sum = vsum;
235}
236