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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$ABC = "01234567456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ADD", "DIV", "RDIV", "MAX", "MIN", "MUL", "SUB", "RSUB", "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_f16 = { 20$ "ADD": lambda x: "vaddq_f16(%s, vb)" % x, 21$ "DIV": lambda x: "vdivq_f16(%s, vb)" % x, 22$ "RDIV": lambda x: "vdivq_f16(vb, %s)" % x, 23$ "MAX": lambda x: "vmaxq_f16(%s, vb)" % x, 24$ "MIN": lambda x: "vminq_f16(%s, vb)" % x, 25$ "MUL": lambda x: "vmulq_f16(%s, vb)" % x, 26$ "SUB": lambda x: "vsubq_f16(%s, vb)" % x, 27$ "RSUB": lambda x: "vsubq_f16(vb, %s)" % x, 28$ "SQRDIFF": lambda x: "vsubq_f16(%s, vb)" % x, 29$}[OP] 30$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION] 31$PARAMS = {"LINEAR": "xnn_f16_default_params", "MINMAX": "xnn_f16_minmax_params"}[ACTIVATION] 32void xnn_f16_v${OP.lower()}c${SUFFIX}_ukernel__neonfp16arith_x${BATCH_TILE}( 33 size_t n, 34 const void* restrict a_ptr, 35 const void* restrict b_ptr, 36 void* restrict y_ptr, 37 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 38{ 39 assert(n != 0); 40 assert(n % sizeof(__fp16) == 0); 41 assert(a_ptr != NULL); 42 assert(b_ptr != NULL); 43 assert(y_ptr != NULL); 44 45 const __fp16* a = (const __fp16*) a_ptr; 46 const __fp16* b = (const __fp16*) b_ptr; 47 __fp16* y = (__fp16*) y_ptr; 48 49 $if ACTIVATION == "MINMAX": 50 const float16x8_t vy_min = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.min)); 51 const float16x8_t vy_max = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->neon.max)); 52 53 const float16x8_t vb = vld1q_dup_f16(b); 54 $if BATCH_TILE > 8: 55 for (; n >= ${BATCH_TILE} * sizeof(__fp16); n -= ${BATCH_TILE} * sizeof(__fp16)) { 56 $for N in range(0, BATCH_TILE, 8): 57 const float16x8_t va${ABC[N:N+8]} = vld1q_f16(a); a += 8; 58 59 $for N in range(0, BATCH_TILE, 8): 60 float16x8_t vy${ABC[N:N+8]} = ${VOPQ_f16("va" + ABC[N:N+8])}; 61 62 $if OP == "SQRDIFF": 63 $for N in range(0, BATCH_TILE, 8): 64 vy${ABC[N:N+8]} = vmulq_f16(vy${ABC[N:N+8]}, vy${ABC[N:N+8]}); 65 66 $if ACTIVATION == "MINMAX": 67 $for N in range(0, BATCH_TILE, 8): 68 vy${ABC[N:N+8]} = vmaxq_f16(vy${ABC[N:N+8]}, vy_min); 69 70 $for N in range(0, BATCH_TILE, 8): 71 vy${ABC[N:N+8]} = vminq_f16(vy${ABC[N:N+8]}, vy_max); 72 73 $for N in range(0, BATCH_TILE, 8): 74 vst1q_f16(y, vy${ABC[N:N+8]}); y += 8; 75 } 76 for (; n >= 8 * sizeof(__fp16); n -= 8 * sizeof(__fp16)) { 77 const float16x8_t va01234567 = vld1q_f16(a); a += 8; 78 79 float16x8_t vy01234567 = ${VOPQ_f16("va01234567")}; 80 $if OP == "SQRDIFF": 81 vy01234567 = vmulq_f16(vy01234567, vy01234567); 82 $if ACTIVATION == "MINMAX": 83 vy01234567 = vmaxq_f16(vy01234567, vy_min); 84 vy01234567 = vminq_f16(vy01234567, vy_max); 85 vst1q_f16(y, vy01234567); y += 8; 86 } 87 if XNN_UNLIKELY(n != 0) { 88 const float16x8_t va01234567 = vld1q_f16(a); 89 90 float16x8_t vy01234567 = ${VOPQ_f16("va01234567")}; 91 $if OP == "SQRDIFF": 92 vy01234567 = vmulq_f16(vy01234567, vy01234567); 93 $if ACTIVATION == "MINMAX": 94 vy01234567 = vmaxq_f16(vy01234567, vy_min); 95 vy01234567 = vminq_f16(vy01234567, vy_max); 96 97 float16x4_t vy0123 = vget_low_f16(vy01234567); 98 if (n & (4 * sizeof(__fp16))) { 99 vst1_f16(y, vy0123); y += 4; 100 vy0123 = vget_high_f16(vy01234567); 101 } 102 103 if (n & (2 * sizeof(__fp16))) { 104 vst1_lane_u32((void*) y, vreinterpret_u32_f16(vy0123), 0); y += 2; 105 vy0123 = vext_f16(vy0123, vy0123, 2); 106 } 107 108 if (n & (1 * sizeof(__fp16))) { 109 vst1_lane_f16(y, vy0123, 0); 110 } 111 } 112} 113