1// Copyright 2021 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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <arm_neon.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vcvt.h> 16 17 18$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 19$XINT8X8_T = {"QS8": "int8x8_t", "QU8": "uint8x8_t"}[DATATYPE] 20$XINT8X16_T = {"QS8": "int8x16_t", "QU8": "uint8x16_t"}[DATATYPE] 21$VLD1Q_DUP_X8 = {"QS8": "vld1q_dup_s8", "QU8": "vld1q_dup_u8"}[DATATYPE] 22$VLD1_DUP_X8 = {"QS8": "vld1_dup_s8", "QU8": "vld1_dup_u8"}[DATATYPE] 23$VST1Q_X8 = {"QS8": "vst1q_s8", "QU8": "vst1q_u8"}[DATATYPE] 24$VST1_X8 = {"QS8": "vst1_s8", "QU8": "vst1_u8"}[DATATYPE] 25$VST1_LANE_X8 = {"QS8": "vst1_lane_s8", "QU8": "vst1_lane_u8"}[DATATYPE] 26$VQMOVXN_S16 = {"QS8": "vqmovn_s16", "QU8": "vqmovun_s16"}[DATATYPE] 27$VEXT_X8 = {"QS8": "vext_s8", "QU8": "vext_u8"}[DATATYPE] 28$VCOMBINE_X8 = {"QS8": "vcombine_s8", "QU8": "vcombine_u8"}[DATATYPE] 29$VGET_LOW_X8 = {"QS8": "vget_low_s8", "QU8": "vget_low_u8"}[DATATYPE] 30$VREINTERPRET_U16_X8 = {"QS8": "vreinterpret_u16_s8", "QU8": "vreinterpret_u16_u8"}[DATATYPE] 31$VREINTERPRET_U32_X8 = {"QS8": "vreinterpret_u32_s8", "QU8": "vreinterpret_u32_u8"}[DATATYPE] 32$VMINQ_X8 = {"QS8": "vminq_s8", "QU8": "vminq_u8"}[DATATYPE] 33$VMIN_X8 = {"QS8": "vmin_s8", "QU8": "vmin_u8"}[DATATYPE] 34$VMAXQ_X8 = {"QS8": "vmaxq_s8", "QU8": "vmaxq_u8"}[DATATYPE] 35$VMAX_X8 = {"QS8": "vmax_s8", "QU8": "vmax_u8"}[DATATYPE] 36void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__neon_x${BATCH_TILE}( 37 size_t n, 38 const float* x, 39 ${XINT8_T}* y, 40 const union xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 41{ 42 assert(n != 0); 43 assert(n % sizeof(float) == 0); 44 assert(x != NULL); 45 assert(y != NULL); 46 47 const float32x4_t vscale = vld1q_dup_f32(¶ms->neon.scale); 48 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon.magic_bias); 49 const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(¶ms->neon.magic_bias_less_zero_point); 50 $if BATCH_TILE > 8: 51 const ${XINT8X16_T} voutput_min = ${VLD1Q_DUP_X8}(¶ms->neon.output_min); 52 const ${XINT8X16_T} voutput_max = ${VLD1Q_DUP_X8}(¶ms->neon.output_max); 53 $else: 54 const ${XINT8X8_T} voutput_min = ${VLD1_DUP_X8}(¶ms->neon.output_min); 55 const ${XINT8X8_T} voutput_max = ${VLD1_DUP_X8}(¶ms->neon.output_max); 56 $if BATCH_TILE > 8: 57 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 58 $for N in range(0, BATCH_TILE, 4): 59 float32x4_t vx${ABC[N:N+4]} = vld1q_f32(x); x += 4; 60 61 $for N in range(0, BATCH_TILE, 4): 62 vx${ABC[N:N+4]} = vmulq_f32(vx${ABC[N:N+4]}, vscale); 63 64 $for N in range(0, BATCH_TILE, 4): 65 vx${ABC[N:N+4]} = vaddq_f32(vx${ABC[N:N+4]}, vmagic_bias); 66 67 $for N in range(0, BATCH_TILE, 4): 68 const int32x4_t vacc${ABC[N:N+4]} = vqsubq_s32(vreinterpretq_s32_f32(vx${ABC[N:N+4]}), vmagic_bias_less_zero_point); 69 70 $for N in range(0, BATCH_TILE, 8): 71 const int16x8_t vacc${ABC[N:N+8]} = vcombine_s16(vqmovn_s32(vacc${ABC[N:N+4]}), vqmovn_s32(vacc${ABC[N+4:N+8]})); 72 73 $for N in range(0, BATCH_TILE, 16): 74 $if N + 8 < BATCH_TILE: 75 ${XINT8X16_T} vy${ABC[N:N+16]} = ${VCOMBINE_X8}(${VQMOVXN_S16}(vacc${ABC[N:N+8]}), ${VQMOVXN_S16}(vacc${ABC[N+8:N+16]})); 76 $else: 77 ${XINT8X8_T} vy${ABC[N:N+8]} = ${VQMOVXN_S16}(vacc${ABC[N:N+8]}); 78 79 $for N in range(0, BATCH_TILE, 16): 80 $if N + 8 < BATCH_TILE: 81 vy${ABC[N:N+16]} = ${VMAXQ_X8}(vy${ABC[N:N+16]}, voutput_min); 82 $else: 83 vy${ABC[N:N+8]} = ${VMAX_X8}(vy${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_min)); 84 85 $for N in range(0, BATCH_TILE, 16): 86 $if N + 8 < BATCH_TILE: 87 vy${ABC[N:N+16]} = ${VMINQ_X8}(vy${ABC[N:N+16]}, voutput_max); 88 $else: 89 vy${ABC[N:N+8]} = ${VMIN_X8}(vy${ABC[N:N+8]}, ${VGET_LOW_X8}(voutput_max)); 90 91 $for N in range(0, BATCH_TILE, 16): 92 $if N + 8 < BATCH_TILE: 93 ${VST1Q_X8}(y, vy${ABC[N:N+16]}); y += 16; 94 $else: 95 ${VST1_X8}(y, vy${ABC[N:N+8]}); y += 8; 96 } 97 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 98 float32x4_t vx_lo = vld1q_f32(x); x += 4; 99 float32x4_t vx_hi = vld1q_f32(x); x += 4; 100 101 vx_lo = vmulq_f32(vx_lo, vscale); 102 vx_hi = vmulq_f32(vx_hi, vscale); 103 104 vx_lo = vaddq_f32(vx_lo, vmagic_bias); 105 vx_hi = vaddq_f32(vx_hi, vmagic_bias); 106 107 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point); 108 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point); 109 110 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)); 111 112 ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc); 113 $if BATCH_TILE > 8: 114 vy = ${VMAX_X8}(vy, ${VGET_LOW_X8}(voutput_min)); 115 vy = ${VMIN_X8}(vy, ${VGET_LOW_X8}(voutput_max)); 116 $else: 117 vy = ${VMAX_X8}(vy, voutput_min); 118 vy = ${VMIN_X8}(vy, voutput_max); 119 ${VST1_X8}(y, vy); y += 8; 120 } 121 if XNN_UNLIKELY(n != 0) { 122 assert(n >= 1 * sizeof(float)); 123 assert(n <= 7 * sizeof(float)); 124 float32x4_t vx_lo = vld1q_f32(x); 125 const float* x_hi = (const float*) ((uintptr_t) x + (n & (4 * sizeof(float)))); 126 float32x4_t vx_hi = vld1q_f32(x_hi); 127 128 vx_lo = vmulq_f32(vx_lo, vscale); 129 vx_hi = vmulq_f32(vx_hi, vscale); 130 131 vx_lo = vaddq_f32(vx_lo, vmagic_bias); 132 vx_hi = vaddq_f32(vx_hi, vmagic_bias); 133 134 const int32x4_t vacc_lo = vqsubq_s32(vreinterpretq_s32_f32(vx_lo), vmagic_bias_less_zero_point); 135 const int32x4_t vacc_hi = vqsubq_s32(vreinterpretq_s32_f32(vx_hi), vmagic_bias_less_zero_point); 136 137 const int16x8_t vacc = vcombine_s16(vqmovn_s32(vacc_lo), vqmovn_s32(vacc_hi)); 138 139 ${XINT8X8_T} vy = ${VQMOVXN_S16}(vacc); 140 $if BATCH_TILE > 8: 141 vy = ${VMAX_X8}(vy, ${VGET_LOW_X8}(voutput_min)); 142 vy = ${VMIN_X8}(vy, ${VGET_LOW_X8}(voutput_max)); 143 $else: 144 vy = ${VMAX_X8}(vy, voutput_min); 145 vy = ${VMIN_X8}(vy, voutput_max); 146 147 if (n & (4 * sizeof(float))) { 148 vst1_lane_u32((void*) y, ${VREINTERPRET_U32_X8}(vy), 0); y += 4; 149 vy = ${VEXT_X8}(vy, vy, 4); 150 } 151 if (n & (2 * sizeof(float))) { 152 vst1_lane_u16((void*) y, ${VREINTERPRET_U16_X8}(vy), 0); y += 2; 153 vy = ${VEXT_X8}(vy, vy, 2); 154 } 155 if (n & (1 * sizeof(float))) { 156 ${VST1_LANE_X8}(y, vy, 0); 157 } 158 } 159} 160