• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1// Copyright 2022 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$assert BATCH_TILE % 8 == 0
7$assert BATCH_TILE >= 8
8$SIMD_TILE = BATCH_TILE // 8
9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
10#include <assert.h>
11
12#include <arm_neon.h>
13
14#include <xnnpack/common.h>
15#include <xnnpack/raddstoreexpminusmax.h>
16
17
18void xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x${BATCH_TILE}${"" if ACCUMULATORS == 1 else "_acc%d" % ACCUMULATORS}(
19    size_t batch,
20    const void* input,
21    const void* max,
22    void* output,
23    void* sum,
24    const union xnn_f16_expminus_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25{
26  assert(batch % sizeof(__fp16) == 0);
27
28  const float16x8_t vi_max = vld1q_dup_f16(max);
29  const float16x8_t vlog2e = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.log2e));
30  const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.magic_bias));
31  const float16x8_t vminus_ln2_hi = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_ln2_hi));
32  const float16x8_t vminus_ln2_lo = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.minus_ln2_lo));
33  const float16x8_t vc2 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c2));
34  const float16x8_t vc1 = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.c1));
35  const float16x8_t vdenorm_cutoff = vreinterpretq_f16_u16(vld1q_dup_u16(&params->neonfp16arith_rr2_p2.denorm_cutoff));
36
37  const __fp16* i = (const __fp16*) input;
38  __fp16* o = (__fp16*) output;
39  $if BATCH_TILE > 8:
40    $for K in range(ACCUMULATORS):
41      float16x8_t vacc${K} = vmovq_n_f16(0.0f);
42    for (; batch >= ${BATCH_TILE} * sizeof(__fp16); batch -= ${BATCH_TILE} * sizeof(__fp16)) {
43      $for N in range(SIMD_TILE):
44        const float16x8_t vi${ABC[N]} = vld1q_f16(i); i += 8;
45
46      $for N in range(SIMD_TILE):
47        const float16x8_t vx${ABC[N]} = vsubq_f16(vi${ABC[N]}, vi_max);
48
49      $for N in range(SIMD_TILE):
50        float16x8_t vn${ABC[N]} = vfmaq_f16(vmagic_bias, vx${ABC[N]}, vlog2e);
51
52      $for N in range(SIMD_TILE):
53        const float16x8_t vs${ABC[N]} = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn${ABC[N]}), 10));
54
55      $for N in range(SIMD_TILE):
56        vn${ABC[N]} = vsubq_f16(vn${ABC[N]}, vmagic_bias);
57
58      $for N in range(SIMD_TILE):
59        float16x8_t vt${ABC[N]} = vfmaq_f16(vx${ABC[N]}, vn${ABC[N]}, vminus_ln2_hi);
60
61      $for N in range(SIMD_TILE):
62        vt${ABC[N]} = vfmaq_f16(vt${ABC[N]}, vn${ABC[N]}, vminus_ln2_lo);
63
64      $for N in range(SIMD_TILE):
65        const float16x8_t vp${ABC[N]} = vfmaq_f16(vc1, vc2, vt${ABC[N]});
66
67      $for N in range(SIMD_TILE):
68        vt${ABC[N]} = vmulq_f16(vt${ABC[N]}, vs${ABC[N]});
69
70      $for N in range(SIMD_TILE):
71        float16x8_t vf${ABC[N]} = vfmaq_f16(vs${ABC[N]}, vp${ABC[N]}, vt${ABC[N]});
72        const uint16x8_t vm${ABC[N]} = vcltq_f16(vx${ABC[N]}, vdenorm_cutoff);
73
74      $for N in range(SIMD_TILE):
75        vf${ABC[N]} = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf${ABC[N]}), vm${ABC[N]}));
76
77      $for N in range(SIMD_TILE):
78        vst1q_f16(o, vf${ABC[N]}); o += 8;
79
80      $for N in range(SIMD_TILE):
81        vacc${N % ACCUMULATORS} = vaddq_f16(vacc${N % ACCUMULATORS}, vf${ABC[N]});
82    }
83    $if ACCUMULATORS > 1:
84      $ACC_SLICE = 1
85      $while ACC_SLICE < ACCUMULATORS:
86        $for A in range(0, ACCUMULATORS, ACC_SLICE * 2):
87          $if A + ACC_SLICE < ACCUMULATORS:
88            vacc${A} = vaddq_f16(vacc${A}, vacc${A + ACC_SLICE});
89        $ACC_SLICE *= 2
90
91    float16x8_t vacc = vacc0;
92  $else:
93    float16x8_t vacc = vmovq_n_f16(0.0f);
94  for (; batch >= 8 * sizeof(__fp16); batch -= 8 * sizeof(__fp16)) {
95    const float16x8_t vi = vld1q_f16(i); i += 8;
96
97    const float16x8_t vx = vsubq_f16(vi, vi_max);
98
99    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
100    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
101    vn = vsubq_f16(vn, vmagic_bias);
102
103    float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
104    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
105
106    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
107    vt = vmulq_f16(vt, vs);
108
109    float16x8_t vf = vfmaq_f16(vs, vp, vt);
110    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
111    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
112
113    vst1q_f16(o, vf); o += 8;
114
115    vacc = vaddq_f16(vacc, vf);
116  }
117  float16x4_t vacc_lo = vadd_f16(vget_low_f16(vacc), vget_high_f16(vacc));
118  if (batch != 0) {
119    assert(batch >= 1 * sizeof(__fp16));
120    assert(batch <= 7 * sizeof(__fp16));
121    const float16x8_t vi = vld1q_f16(i);
122
123    const float16x8_t vx = vsubq_f16(vi, vi_max);
124
125    float16x8_t vn = vfmaq_f16(vmagic_bias, vx, vlog2e);
126    const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
127    vn = vsubq_f16(vn, vmagic_bias);
128
129    float16x8_t vt = vfmaq_f16(vx, vn, vminus_ln2_hi);
130    vt = vfmaq_f16(vt, vn, vminus_ln2_lo);
131
132    const float16x8_t vp = vfmaq_f16(vc1, vc2, vt);
133    vt = vmulq_f16(vt, vs);
134
135    float16x8_t vf = vfmaq_f16(vs, vp, vt);
136    const uint16x8_t vm = vcltq_f16(vx, vdenorm_cutoff);
137    vf = vreinterpretq_f16_u16(vbicq_u16(vreinterpretq_u16_f16(vf), vm));
138
139    float16x4_t vf_lo = vget_low_f16(vf);
140    if (batch & (4 * sizeof(__fp16))) {
141      vst1_f16(o, vf_lo); o += 4;
142      vacc_lo = vadd_f16(vacc_lo, vf_lo);
143      vf_lo = vget_high_f16(vf);
144    }
145    if (batch & (2 * sizeof(__fp16))) {
146      vst1_lane_u32((void*) o, vreinterpret_u32_f16(vf_lo), 0); o += 2;
147      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 32)));
148      vf_lo = vext_f16(vf_lo, vf_lo, 2);
149    }
150    if (batch & (1 * sizeof(__fp16))) {
151      vst1_lane_f16(o, vf_lo, 0);
152      vacc_lo = vadd_f16(vacc_lo, vreinterpret_f16_u64(vshl_n_u64(vreinterpret_u64_f16(vf_lo), 48)));
153    }
154  }
155  vacc_lo = vpadd_f16(vacc_lo, vacc_lo);
156  *((__fp16*) sum) = vget_lane_f16(vacc_lo, 0) + vget_lane_f16(vacc_lo, 1);
157}
158