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$assert BATCH_TILE % 16 == 0 7$assert BATCH_TILE >= 16 8$SIMD_TILE = BATCH_TILE // 16 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vcvt.h> 16 17 18void xnn_f32_f16_vcvt_ukernel__avx512skx_x${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 > 16: 31 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 32 const __m512 vf0 = _mm512_loadu_ps(input); 33 $for N in range(1, SIMD_TILE): 34 const __m512 vf${N} = _mm512_loadu_ps(input + ${N * 16}); 35 input += ${BATCH_TILE}; 36 37 _mm256_storeu_si256((__m256i*) o, _mm512_cvtps_ph(vf0, _MM_FROUND_NO_EXC)); 38 $for N in range(1, SIMD_TILE): 39 _mm256_storeu_si256((__m256i*) (o + ${N * 16}), _mm512_cvtps_ph(vf${N}, _MM_FROUND_NO_EXC)); 40 o += ${BATCH_TILE}; 41 } 42 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 43 const __m512 vf = _mm512_loadu_ps(input); 44 input += 16; 45 46 _mm256_storeu_si256((__m256i*) o, _mm512_cvtps_ph(vf, _MM_FROUND_NO_EXC)); 47 o += 16; 48 } 49 if XNN_UNLIKELY(n != 0) { 50 assert(n >= 1 * sizeof(float)); 51 assert(n <= 15 * sizeof(float)); 52 53 // Prepare mask for valid elements (depends on n). 54 n >>= 2 /* log2(sizeof(float)) */; 55 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 56 57 const __m512 vf = _mm512_maskz_loadu_ps(vmask, input); 58 const __m256i vh = _mm512_cvtps_ph(vf, _MM_FROUND_NO_EXC); 59 _mm256_mask_storeu_epi16(o, vmask, vh); 60 } 61} 62