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$SIMD_TILE = BATCH_TILE // 8 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vcvt.h> 16 17 18void xnn_f32_f16_vcvt_ukernel__f16c_x${BATCH_TILE}( 19 size_t n, 20 const float* input, 21 void* output, 22 const union xnn_f32_f16_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) 23{ 24 assert(n != 0); 25 assert(n % sizeof(float) == 0); 26 assert(input != NULL); 27 assert(output != NULL); 28 29 uint16_t* o = (uint16_t*) output; 30 $if BATCH_TILE > 8: 31 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 32 const __m256 vf0 = _mm256_loadu_ps(input); 33 $for N in range(1, SIMD_TILE): 34 const __m256 vf${N} = _mm256_loadu_ps(input + ${N * 8}); 35 input += ${BATCH_TILE}; 36 37 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf0, _MM_FROUND_NO_EXC)); 38 $for N in range(1, SIMD_TILE): 39 _mm_storeu_si128((__m128i*) (o + ${N * 8}), _mm256_cvtps_ph(vf${N}, _MM_FROUND_NO_EXC)); 40 o += ${BATCH_TILE}; 41 } 42 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 43 const __m256 vf = _mm256_loadu_ps(input); 44 input += 8; 45 46 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vf, _MM_FROUND_NO_EXC)); 47 o += 8; 48 } 49 if XNN_UNLIKELY(n != 0) { 50 assert(n >= 1 * sizeof(float)); 51 assert(n <= 7 * sizeof(float)); 52 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->f16c.mask_table[7] - n)); 53 54 const __m256 vf = _mm256_maskload_ps(input, vmask); 55 56 __m128 vf_lo = _mm256_castps256_ps128(vf); 57 if (n & (4 * sizeof(float))) { 58 _mm_storel_epi64((__m128i*) o, _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC)); 59 vf_lo = _mm256_extractf128_ps(vf, 1); 60 o += 4; 61 } 62 __m128i vh = _mm_cvtps_ph(vf_lo, _MM_FROUND_NO_EXC); 63 if (n & (2 * sizeof(float))) { 64 _mm_storeu_si32(o, vh); 65 vh = _mm_srli_epi64(vh, 32); 66 o += 2; 67 } 68 if (n & (1 * sizeof(float))) { 69 *((uint16_t*) o) = _mm_extract_epi16(vh, 0); 70 } 71 } 72} 73