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$assert SSE in [2, 4] 7$assert not AVX or SSE == 4 8$assert BATCH_TILE % 8 == 0 9$assert BATCH_TILE >= 8 10$SIMD_TILE = BATCH_TILE // 8 11$SSE_HEADER = {2: "emmintrin.h", 4: "smmintrin.h"}[SSE] 12$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 13#include <assert.h> 14 15#include <${SSE_HEADER}> 16 17#include <xnnpack/common.h> 18#include <xnnpack/unaligned.h> 19#include <xnnpack/vcvt.h> 20 21 22$ISA = "avx" if AVX else {2: "sse2", 4: "sse41"}[SSE] 23void xnn_f32_f16_vcvt_ukernel__${ISA}_x${BATCH_TILE}( 24 size_t n, 25 const float* input, 26 void* output, 27 const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 28{ 29 assert(n != 0); 30 assert(n % sizeof(float) == 0); 31 assert(input != NULL); 32 assert(output != NULL); 33 34 const __m128 vnonsign_mask = _mm_load_ps((const float*) params->sse2.nonsign_mask); 35 const __m128i vexp_bias = _mm_load_si128((const __m128i*) params->sse2.exp_bias); 36 const __m128 vscale_to_inf = _mm_load_ps(params->sse2.scale_to_inf); 37 const __m128i vexpw_max = _mm_load_si128((const __m128i*) params->sse2.expw_max); 38 const __m128 vscale_to_zero = _mm_load_ps(params->sse2.scale_to_zero); 39 const __m128i vbias_min = _mm_load_si128((const __m128i*) params->sse2.bias_min); 40 const __m128i vmanth_mask = _mm_load_si128((const __m128i*) params->sse2.manth_mask); 41 const __m128i vexph_mask = _mm_load_si128((const __m128i*) params->sse2.exph_mask); 42 const __m128i vnanh = _mm_load_si128((const __m128i*) params->sse2.nanh); 43 44 uint16_t* o = (uint16_t*) output; 45 $if BATCH_TILE > 8: 46 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 47 const __m128 vx0 = _mm_loadu_ps(input); 48 $for N in range(1, 2*SIMD_TILE): 49 const __m128 vx${N} = _mm_loadu_ps(input + ${N * 4}); 50 input += ${BATCH_TILE}; 51 52 $for N in range(2*SIMD_TILE): 53 const __m128 vabsx${N} = _mm_and_ps(vx${N}, vnonsign_mask); 54 55 $for N in range(2*SIMD_TILE): 56 const __m128 vsignx${N} = _mm_xor_ps(vx${N}, vabsx${N}); 57 58 $for N in range(2*SIMD_TILE): 59 __m128i vbias${N} = _mm_add_epi32(_mm_castps_si128(vabsx${N}), vexp_bias); 60 61 $for N in range(2*SIMD_TILE): 62 __m128 vf${N} = _mm_mul_ps(vabsx${N}, vscale_to_inf); 63 64 $for N in range(2*SIMD_TILE): 65 const __m128i vnanmaskw${N} = _mm_cmpgt_epi32(_mm_castps_si128(vabsx${N}), vexpw_max); 66 67 $for N in range(2*SIMD_TILE): 68 vbias${N} = _mm_and_si128(vbias${N}, vexpw_max); 69 70 $for N in range(2*SIMD_TILE): 71 vf${N} = _mm_mul_ps(vf${N}, vscale_to_zero); 72 73 $for N in range(SIMD_TILE): 74 const __m128i vnanmaskh${N} = _mm_packs_epi32(vnanmaskw${2*N}, vnanmaskw${2*N+1}); 75 76 $for N in range(SIMD_TILE): 77 const __m128i vsignh${N} = _mm_packs_epi32(_mm_castps_si128(vsignx${2*N}), _mm_castps_si128(vsignx${2*N+1})); 78 79 $for N in range(2*SIMD_TILE): 80 vbias${N} = _mm_max_epi16(vbias${N}, vbias_min); 81 82 $if SSE < 4: 83 $for N in range(SIMD_TILE): 84 __m128i vh${N} = _mm_and_si128(vnanh, vnanmaskh${N}); 85 86 $for N in range(2*SIMD_TILE): 87 vf${N} = _mm_add_ps(vf${N}, _mm_castsi128_ps(vbias${N})); 88 89 $if SSE < 4: 90 $for N in range(SIMD_TILE): 91 vh${N} = _mm_or_si128(vh${N}, vsignh${N}); 92 93 $for N in range(2*SIMD_TILE): 94 __m128i vexpw${N} = _mm_srli_epi32(_mm_castps_si128(vf${N}), 13); 95 96 $for N in range(2*SIMD_TILE): 97 const __m128i vmantw${N} = _mm_and_si128(_mm_castps_si128(vf${N}), vmanth_mask); 98 99 $for N in range(2*SIMD_TILE): 100 vexpw${N} = _mm_and_si128(vexpw${N}, vexph_mask); 101 102 $for N in range(2*SIMD_TILE): 103 const __m128i vnonsignw${N} = _mm_add_epi32(vmantw${N}, vexpw${N}); 104 105 $for N in range(SIMD_TILE): 106 const __m128i vnonsignh${N} = _mm_packs_epi32(vnonsignw${2*N}, vnonsignw${2*N+1}); 107 108 $if SSE == 4: 109 $for N in range(SIMD_TILE): 110 const __m128i vabsh${N} = _mm_blendv_epi8(vnonsignh${N}, vnanh, vnanmaskh${N}); 111 112 $for N in range(SIMD_TILE): 113 const __m128i vh${N} = _mm_or_si128(vabsh${N}, vsignh${N}); 114 $else: 115 $for N in range(SIMD_TILE): 116 vh${N} = _mm_or_si128(vh${N}, _mm_andnot_si128(vnanmaskh${N}, vnonsignh${N})); 117 118 _mm_storeu_si128((__m128i*) o, vh0); 119 $for N in range(1, SIMD_TILE): 120 _mm_storeu_si128((__m128i*) (o + ${N * 8}), vh${N}); 121 o += ${BATCH_TILE}; 122 } 123 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 124 const __m128 vx_lo = _mm_loadu_ps(input); 125 const __m128 vx_hi = _mm_loadu_ps(input + 4); 126 input += 8; 127 128 const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask); 129 const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask); 130 131 const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo); 132 const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi); 133 __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias); 134 __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias); 135 __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf); 136 __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf); 137 const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max); 138 const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max); 139 140 vbias_lo = _mm_and_si128(vbias_lo, vexpw_max); 141 vbias_hi = _mm_and_si128(vbias_hi, vexpw_max); 142 vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero); 143 vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero); 144 const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi); 145 const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi)); 146 147 vbias_lo = _mm_max_epi16(vbias_lo, vbias_min); 148 vbias_hi = _mm_max_epi16(vbias_hi, vbias_min); 149 $if SSE < 4: 150 __m128i vh = _mm_and_si128(vnanh, vnanmaskh); 151 152 vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo)); 153 vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi)); 154 $if SSE < 4: 155 vh = _mm_or_si128(vh, vsignh); 156 157 __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13); 158 __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13); 159 const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask); 160 const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask); 161 162 vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask); 163 vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask); 164 165 const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo); 166 const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi); 167 168 const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi); 169 170 $if SSE == 4: 171 const __m128i vabsh = _mm_blendv_epi8(vnonsignh, vnanh, vnanmaskh); 172 173 const __m128i vh = _mm_or_si128(vabsh, vsignh); 174 $else: 175 vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh)); 176 177 _mm_storeu_si128((__m128i*) o, vh); 178 o += 8; 179 } 180 if XNN_UNPREDICTABLE(n != 0) { 181 const __m128 vx_lo = _mm_loadu_ps(input); 182 const float* input_hi = (const float*) ((uintptr_t) input + (n & (4 * sizeof(float)))); 183 const __m128 vx_hi = _mm_loadu_ps(input_hi); 184 185 const __m128 vabsx_lo = _mm_and_ps(vx_lo, vnonsign_mask); 186 const __m128 vabsx_hi = _mm_and_ps(vx_hi, vnonsign_mask); 187 188 const __m128 vsignx_lo = _mm_xor_ps(vx_lo, vabsx_lo); 189 const __m128 vsignx_hi = _mm_xor_ps(vx_hi, vabsx_hi); 190 __m128i vbias_lo = _mm_add_epi32(_mm_castps_si128(vabsx_lo), vexp_bias); 191 __m128i vbias_hi = _mm_add_epi32(_mm_castps_si128(vabsx_hi), vexp_bias); 192 __m128 vf_lo = _mm_mul_ps(vabsx_lo, vscale_to_inf); 193 __m128 vf_hi = _mm_mul_ps(vabsx_hi, vscale_to_inf); 194 const __m128i vnanmaskw_lo = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_lo), vexpw_max); 195 const __m128i vnanmaskw_hi = _mm_cmpgt_epi32(_mm_castps_si128(vabsx_hi), vexpw_max); 196 197 vbias_lo = _mm_and_si128(vbias_lo, vexpw_max); 198 vbias_hi = _mm_and_si128(vbias_hi, vexpw_max); 199 vf_lo = _mm_mul_ps(vf_lo, vscale_to_zero); 200 vf_hi = _mm_mul_ps(vf_hi, vscale_to_zero); 201 const __m128i vnanmaskh = _mm_packs_epi32(vnanmaskw_lo, vnanmaskw_hi); 202 const __m128i vsignh = _mm_packs_epi32(_mm_castps_si128(vsignx_lo), _mm_castps_si128(vsignx_hi)); 203 204 vbias_lo = _mm_max_epi16(vbias_lo, vbias_min); 205 vbias_hi = _mm_max_epi16(vbias_hi, vbias_min); 206 $if SSE < 4: 207 __m128i vh = _mm_and_si128(vnanh, vnanmaskh); 208 209 vf_lo = _mm_add_ps(vf_lo, _mm_castsi128_ps(vbias_lo)); 210 vf_hi = _mm_add_ps(vf_hi, _mm_castsi128_ps(vbias_hi)); 211 $if SSE < 4: 212 vh = _mm_or_si128(vh, vsignh); 213 214 __m128i vexpw_lo = _mm_srli_epi32(_mm_castps_si128(vf_lo), 13); 215 __m128i vexpw_hi = _mm_srli_epi32(_mm_castps_si128(vf_hi), 13); 216 const __m128i vmantw_lo = _mm_and_si128(_mm_castps_si128(vf_lo), vmanth_mask); 217 const __m128i vmantw_hi = _mm_and_si128(_mm_castps_si128(vf_hi), vmanth_mask); 218 219 vexpw_lo = _mm_and_si128(vexpw_lo, vexph_mask); 220 vexpw_hi = _mm_and_si128(vexpw_hi, vexph_mask); 221 222 const __m128i vnonsignw_lo = _mm_add_epi32(vmantw_lo, vexpw_lo); 223 const __m128i vnonsignw_hi = _mm_add_epi32(vmantw_hi, vexpw_hi); 224 225 const __m128i vnonsignh = _mm_packs_epi32(vnonsignw_lo, vnonsignw_hi); 226 227 $if SSE == 4: 228 const __m128i vabsh = _mm_blendv_epi8(vnonsignh, vnanh, vnanmaskh); 229 230 __m128i vh = _mm_or_si128(vabsh, vsignh); 231 $else: 232 vh = _mm_or_si128(vh, _mm_andnot_si128(vnanmaskh, vnonsignh)); 233 234 if (n & (4 * sizeof(float))) { 235 _mm_storel_epi64((__m128i*) o, vh); 236 vh = _mm_unpackhi_epi64(vh, vh); 237 o += 4; 238 } 239 if (n & (2 * sizeof(float))) { 240 unaligned_store_u32(o, (uint32_t) _mm_cvtsi128_si32(vh)); 241 vh = _mm_srli_epi64(vh, 32); 242 o += 2; 243 } 244 if (n & (1 * sizeof(float))) { 245 $if SSE == 4: 246 *o = (uint16_t) _mm_extract_epi16(vh, 0); 247 $else: 248 *o = (uint16_t) _mm_cvtsi128_si32(vh); 249 } 250 } 251} 252