1// Copyright 2019 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$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"] 10$assert ACTIVATION in ["LINEAR", "MINMAX"] 11#include <assert.h> 12 13#include <arm_neon.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/vbinary.h> 17 18 19$VOPQ_F32 = { 20$ "ADD": lambda x, y: "vaddq_f32(%s, %s)" % (x, y), 21$ "DIV": lambda x, y: "vdivq_f32(%s, %s)" % (x, y), 22$ "MAX": lambda x, y: "vmaxq_f32(%s, %s)" % (x, y), 23$ "MIN": lambda x, y: "vminq_f32(%s, %s)" % (x, y), 24$ "MUL": lambda x, y: "vmulq_f32(%s, %s)" % (x, y), 25$ "SUB": lambda x, y: "vsubq_f32(%s, %s)" % (x, y), 26$ "SQRDIFF": lambda x, y: "vsubq_f32(%s, %s)" % (x, y), 27$}[OP] 28$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION] 29$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION] 30void xnn_f32_v${OP.lower()}${SUFFIX}_ukernel__neon_x${BATCH_TILE}( 31 size_t n, 32 const float* a, 33 const float* b, 34 float* y, 35 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 36{ 37 assert(n != 0); 38 assert(n % sizeof(float) == 0); 39 assert(a != NULL); 40 assert(b != NULL); 41 assert(y != NULL); 42 43 $if ACTIVATION == "MINMAX": 44 const float32x4_t vy_min = vld1q_dup_f32(¶ms->scalar.min); 45 const float32x4_t vy_max = vld1q_dup_f32(¶ms->scalar.max); 46 47 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 48 $for N in range(0, BATCH_TILE, 4): 49 const float32x4_t va${ABC[N:N+4]} = vld1q_f32(a); a += 4; 50 const float32x4_t vb${ABC[N:N+4]} = vld1q_f32(b); b += 4; 51 52 $for N in range(0, BATCH_TILE, 4): 53 float32x4_t vy${ABC[N:N+4]} = ${VOPQ_F32("va" + ABC[N:N+4], "vb" + ABC[N:N+4])}; 54 55 $if OP == "SQRDIFF": 56 $for N in range(0, BATCH_TILE, 4): 57 vy${ABC[N:N+4]} = vmulq_f32(vy${ABC[N:N+4]}, vy${ABC[N:N+4]}); 58 59 $if ACTIVATION == "MINMAX": 60 $for N in range(0, BATCH_TILE, 4): 61 vy${ABC[N:N+4]} = vmaxq_f32(vy${ABC[N:N+4]}, vy_min); 62 63 $for N in range(0, BATCH_TILE, 4): 64 vy${ABC[N:N+4]} = vminq_f32(vy${ABC[N:N+4]}, vy_max); 65 66 $for N in range(0, BATCH_TILE, 4): 67 vst1q_f32(y, vy${ABC[N:N+4]}); y += 4; 68 } 69 $if BATCH_TILE > 4: 70 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 71 const float32x4_t va0123 = vld1q_f32(a); a += 4; 72 const float32x4_t vb0123 = vld1q_f32(b); b += 4; 73 74 float32x4_t vy0123 = ${VOPQ_F32("va0123", "vb0123")}; 75 $if OP == "SQRDIFF": 76 vy0123 = vmulq_f32(vy0123, vy0123); 77 $if ACTIVATION == "MINMAX": 78 vy0123 = vmaxq_f32(vy0123, vy_min); 79 vy0123 = vminq_f32(vy0123, vy_max); 80 vst1q_f32(y, vy0123); y += 4; 81 } 82 if XNN_UNLIKELY(n != 0) { 83 const float32x4_t va0123 = vld1q_f32(a); 84 const float32x4_t vb0123 = vld1q_f32(b); 85 86 float32x4_t vy0123 = ${VOPQ_F32("va0123", "vb0123")}; 87 $if OP == "SQRDIFF": 88 vy0123 = vmulq_f32(vy0123, vy0123); 89 $if ACTIVATION == "MINMAX": 90 vy0123 = vmaxq_f32(vy0123, vy_min); 91 vy0123 = vminq_f32(vy0123, vy_max); 92 93 float32x2_t vy01 = vget_low_f32(vy0123); 94 if (n & (2 * sizeof(float))) { 95 vst1_f32(y, vy01); y += 2; 96 vy01 = vget_high_f32(vy0123); 97 } 98 if (n & (1 * sizeof(float))) { 99 vst1_lane_f32(y, vy01, 0); 100 } 101 } 102} 103