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$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"] 10#include <assert.h> 11 12#include <arm_neon.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/math.h> 16#include <xnnpack/vunary.h> 17 18 19$VRNDQ_F32 = { 20$ "RNDNE": "vrndnq_f32", 21$ "RNDZ": "vrndq_f32", 22$ "RNDU": "vrndpq_f32", 23$ "RNDD": "vrndmq_f32", 24$}[OP] 25void xnn_f32_v${OP.lower()}_ukernel__neonv8_x${BATCH_TILE}( 26 size_t n, 27 const float* x, 28 float* y, 29 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN 30{ 31 assert(n != 0); 32 assert(n % sizeof(float) == 0); 33 34 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 35 $for N in range(0, BATCH_TILE, 4): 36 const float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4; 37 38 $for N in range(0, BATCH_TILE, 4): 39 const float32x4_t vy${ABC[N:N+4]} = ${VRNDQ_F32}(vx${ABC[N:N+4]}); 40 41 $for N in range(0, BATCH_TILE, 4): 42 vst1q_f32(y, vy${ABC[N:N+4]}); y += 4; 43 } 44 $if BATCH_TILE > 4: 45 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 46 const float32x4_t vx = vld1q_f32(x); x += 4; 47 const float32x4_t vy = ${VRNDQ_F32}(vx); 48 vst1q_f32(y, vy); y += 4; 49 } 50 if XNN_UNLIKELY(n != 0) { 51 const float32x4_t vx = vld1q_f32(x); 52 const float32x4_t vy = ${VRNDQ_F32}(vx); 53 float32x2_t vy_lo = vget_low_f32(vy); 54 if (n & (2 * sizeof(float))) { 55 vst1_f32(y, vy_lo); y += 2; 56 vy_lo = vget_high_f32(vy); 57 } 58 if (n & (1 * sizeof(float))) { 59 vst1_lane_f32(y, vy_lo, 0); 60 } 61 } 62} 63