1// Copyright 2022 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$assert OP in ["RNDNE", "RNDZ", "RNDU", "RNDD"] 11#include <assert.h> 12 13#include <immintrin.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/intrinsics-polyfill.h> 17#include <xnnpack/math.h> 18#include <xnnpack/vunary.h> 19 20 21$_MM_FROUND_TO_FLAG = { 22$ "RNDNE": "_MM_FROUND_TO_NEAREST_INT", 23$ "RNDZ": "_MM_FROUND_TO_ZERO", 24$ "RNDU": "_MM_FROUND_TO_POS_INF", 25$ "RNDD": "_MM_FROUND_TO_NEG_INF", 26$}[OP] 27void xnn_f16_v${OP.lower()}_ukernel__f16c_x${BATCH_TILE}( 28 size_t n, 29 const void* input, 30 void* output, 31 const union xnn_f16_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) 32{ 33 assert(n != 0); 34 assert(n % sizeof(uint16_t) == 0); 35 36 const uint16_t* i = (const uint16_t*) input; 37 uint16_t* o = (uint16_t*) output; 38 $if BATCH_TILE > 8: 39 for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) { 40 __m256 vacc${ABC[0]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i)); 41 $for N in range(1, SIMD_TILE): 42 __m256 vacc${ABC[N]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i + ${N * 8}))); 43 i += ${BATCH_TILE}; 44 45 $for N in range(SIMD_TILE): 46 vacc${ABC[N]} = _mm256_round_ps(vacc${ABC[N]}, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC); 47 48 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc${ABC[0]}, _MM_FROUND_NO_EXC)); 49 $for N in range(1, SIMD_TILE): 50 _mm_storeu_si128((__m128i*) (o + ${N * 8}), _mm256_cvtps_ph(vacc${ABC[N]}, _MM_FROUND_NO_EXC)); 51 o += ${BATCH_TILE}; 52 } 53 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { 54 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i)); 55 i += 8; 56 57 vacc = _mm256_round_ps(vacc, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC); 58 59 _mm_storeu_si128((__m128i*) o, _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC)); 60 o += 8; 61 } 62 if XNN_UNLIKELY(n != 0) { 63 assert(n >= 1 * sizeof(uint16_t)); 64 assert(n <= 7 * sizeof(uint16_t)); 65 __m256 vacc = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i)); 66 vacc = _mm256_round_ps(vacc, ${_MM_FROUND_TO_FLAG} | _MM_FROUND_NO_EXC); 67 __m128i vh = _mm256_cvtps_ph(vacc, _MM_FROUND_NO_EXC); 68 if (n & (4 * sizeof(uint16_t))) { 69 _mm_storel_epi64((__m128i*) o, vh); 70 vh = _mm_unpackhi_epi64(vh, vh); 71 o += 4; 72 } 73 if (n & (2 * sizeof(uint16_t))) { 74 _mm_storeu_si32(o, vh); 75 vh = _mm_srli_epi64(vh, 32); 76 o += 2; 77 } 78 if (n & (1 * sizeof(uint16_t))) { 79 *o = (uint16_t) _mm_extract_epi16(vh, 0); 80 } 81 } 82} 83