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$SIMD_TILE = BATCH_TILE // 8 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <arm_neon.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/vcvt.h> 16 17 18void xnn_f16_f32_vcvt_ukernel__neon_int16_x${BATCH_TILE}( 19 size_t n, 20 const void* input, 21 float* output, 22 const union xnn_f16_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 23{ 24 assert(n != 0); 25 assert(n % sizeof(uint16_t) == 0); 26 assert(input != NULL); 27 assert(output != NULL); 28 29 const uint16x8_t vsign_mask = vmovq_n_u16(0x8000); 30 const uint16x8_t vexp_offset = vmovq_n_u16(0x7000); 31 const float32x4_t vexp_scale = vld1q_dup_f32(¶ms->neon.exp_scale); 32 const uint32x4_t vmagic_bias = vmovq_n_u32(0x3F000000); 33 const uint16x8_t vdenorm_cutoff = vmovq_n_u16(0x0400); 34 35 const uint16_t* i = (const uint16_t*) input; 36 $if BATCH_TILE > 8: 37 for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) { 38 $for N in range(SIMD_TILE): 39 const uint16x8_t vh${N} = vld1q_u16(i); i += 8; 40 41 $for N in range(SIMD_TILE): 42 const uint16x8_t vsign${N} = vandq_u16(vh${N}, vsign_mask); 43 44 $for N in range(SIMD_TILE): 45 const uint16x8_t vnonsign${N} = veorq_u16(vh${N}, vsign${N}); 46 47 $for N in range(SIMD_TILE): 48 const uint16x8x2_t vprenorm${N} = vzipq_u16(vshlq_n_u16(vnonsign${N}, 13), vsraq_n_u16(vexp_offset, vnonsign${N}, 3)); 49 50 $for N in range(SIMD_TILE): 51 const float32x4_t vnorm${2*N} = vmulq_f32(vreinterpretq_f32_u16(vprenorm${N}.val[0]), vexp_scale); 52 const float32x4_t vnorm${2*N+1} = vmulq_f32(vreinterpretq_f32_u16(vprenorm${N}.val[1]), vexp_scale); 53 54 $for N in range(SIMD_TILE): 55 const float32x4_t vdenorm${2*N} = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign${N}))), vreinterpretq_f32_u32(vmagic_bias)); 56 const float32x4_t vdenorm${2*N+1} = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign${N}))), vreinterpretq_f32_u32(vmagic_bias)); 57 58 $for N in range(SIMD_TILE): 59 const uint16x8_t vmask${N} = vcgtq_u16(vnonsign${N}, vdenorm_cutoff); 60 61 $for N in range(SIMD_TILE): 62 const uint32x4_t vxmask${2*N} = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask${N})))); 63 const uint32x4_t vf${2*N} = vorrq_u32(vshll_n_u16(vget_low_u16(vsign${N}), 16), 64 vreinterpretq_u32_f32(vbslq_f32(vxmask${2*N}, vnorm${2*N}, vdenorm${2*N}))); 65 66 $for N in range(SIMD_TILE): 67 const uint32x4_t vxmask${2*N+1} = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask${N})))); 68 const uint32x4_t vf${2*N+1} = vorrq_u32(vshll_n_u16(vget_high_u16(vsign${N}), 16), 69 vreinterpretq_u32_f32(vbslq_f32(vxmask${2*N+1}, vnorm${2*N+1}, vdenorm${2*N+1}))); 70 71 $for N in range(2*SIMD_TILE): 72 vst1q_f32(output, vreinterpretq_f32_u32(vf${N})); output += 4; 73 } 74 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { 75 const uint16x8_t vh = vld1q_u16(i); i += 8; 76 77 const uint16x8_t vsign = vandq_u16(vh, vsign_mask); 78 79 const uint16x8_t vnonsign = veorq_u16(vh, vsign); 80 81 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3)); 82 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); 83 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale); 84 85 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias)); 86 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias)); 87 88 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff); 89 90 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask)))); 91 const uint32x4_t vf_lo = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16), 92 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); 93 94 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask)))); 95 const uint32x4_t vf_hi = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16), 96 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi))); 97 98 vst1q_f32(output, vreinterpretq_f32_u32(vf_lo)); output += 4; 99 vst1q_f32(output, vreinterpretq_f32_u32(vf_hi)); output += 4; 100 } 101 if XNN_UNPREDICTABLE(n != 0) { 102 const uint16x8_t vh = vld1q_u16(i); i += 8; 103 104 const uint16x8_t vsign = vandq_u16(vh, vsign_mask); 105 106 const uint16x8_t vnonsign = veorq_u16(vh, vsign); 107 108 const uint16x8x2_t vprenorm = vzipq_u16(vshlq_n_u16(vnonsign, 13), vsraq_n_u16(vexp_offset, vnonsign, 3)); 109 const float32x4_t vnorm_lo = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[0]), vexp_scale); 110 const float32x4_t vnorm_hi = vmulq_f32(vreinterpretq_f32_u16(vprenorm.val[1]), vexp_scale); 111 112 const float32x4_t vdenorm_lo = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_low_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias)); 113 const float32x4_t vdenorm_hi = vsubq_f32(vreinterpretq_f32_u32(vaddw_u16(vmagic_bias, vget_high_u16(vnonsign))), vreinterpretq_f32_u32(vmagic_bias)); 114 115 const uint16x8_t vmask = vcgtq_u16(vnonsign, vdenorm_cutoff); 116 117 const uint32x4_t vxmask_lo = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_low_u16(vmask)))); 118 uint32x4_t vf = vorrq_u32(vshll_n_u16(vget_low_u16(vsign), 16), 119 vreinterpretq_u32_f32(vbslq_f32(vxmask_lo, vnorm_lo, vdenorm_lo))); 120 121 if (n & (4 * sizeof(uint16_t))) { 122 vst1q_f32(output, vreinterpretq_f32_u32(vf)); output += 4; 123 124 const uint32x4_t vxmask_hi = vreinterpretq_u32_s32(vmovl_s16(vreinterpret_s16_u16(vget_high_u16(vmask)))); 125 vf = vorrq_u32(vshll_n_u16(vget_high_u16(vsign), 16), 126 vreinterpretq_u32_f32(vbslq_f32(vxmask_hi, vnorm_hi, vdenorm_hi))); 127 } 128 uint32x2_t vf_lo = vget_low_u32(vf); 129 if (n & (2 * sizeof(uint16_t))) { 130 vst1_f32(output, vreinterpret_f32_u32(vf_lo)); output += 2; 131 vf_lo = vget_high_u32(vf); 132 } 133 if (n & (1 * sizeof(uint16_t))) { 134 vst1_lane_f32(output, vreinterpret_f32_u32(vf_lo), 0); 135 } 136 } 137} 138