1// Copyright 2019 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$assert BATCH_TILE % 4 == 0 7$assert BATCH_TILE >= 4 8$ABC = "0123456789ABCDEFGHIJKLMN" 9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB"] 10#include <assert.h> 11 12#include <xmmintrin.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/vbinary.h> 16 17 18$_MM_OP_PS = { 19$ "ADD": lambda x, y: "_mm_add_ps(%s, %s)" % (x, y), 20$ "DIV": lambda x, y: "_mm_div_ps(%s, %s)" % (x, y), 21$ "MAX": lambda x, y: "_mm_max_ps(%s, %s)" % (x, y), 22$ "MIN": lambda x, y: "_mm_min_ps(%s, %s)" % (x, y), 23$ "MUL": lambda x, y: "_mm_mul_ps(%s, %s)" % (x, y), 24$ "SUB": lambda x, y: "_mm_sub_ps(%s, %s)" % (x, y), 25$}[OP] 26void xnn_f32_v${OP.lower()}_ukernel__sse_x${BATCH_TILE}( 27 size_t n, 28 const float* a, 29 const float* b, 30 float* y, 31 const union xnn_f32_output_params params[restrict static 1]) 32{ 33 assert(n != 0); 34 assert(n % sizeof(float) == 0); 35 36 const __m128 vy_min = _mm_load_ps(params->sse.min); 37 const __m128 vy_max = _mm_load_ps(params->sse.max); 38 39 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 40 const __m128 va${ABC[0:4]} = _mm_loadu_ps(a); 41 $for N in range(4, BATCH_TILE, 4): 42 const __m128 va${ABC[N:N+4]} = _mm_loadu_ps(a + ${N}); 43 a += ${BATCH_TILE}; 44 45 const __m128 vb${ABC[0:4]} = _mm_loadu_ps(b); 46 $for N in range(4, BATCH_TILE, 4): 47 const __m128 vb${ABC[N:N+4]} = _mm_loadu_ps(b + ${N}); 48 b += ${BATCH_TILE}; 49 50 $for N in range(0, BATCH_TILE, 4): 51 __m128 vy${ABC[N:N+4]} = ${_MM_OP_PS("va" + ABC[N:N+4], "vb" + ABC[N:N+4])}; 52 53 $for N in range(0, BATCH_TILE, 4): 54 vy${ABC[N:N+4]} = _mm_max_ps(vy${ABC[N:N+4]}, vy_min); 55 56 $for N in range(0, BATCH_TILE, 4): 57 vy${ABC[N:N+4]} = _mm_min_ps(vy${ABC[N:N+4]}, vy_max); 58 59 _mm_storeu_ps(y, vy${ABC[0:4]}); 60 $for N in range(4, BATCH_TILE, 4): 61 _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]}); 62 y += ${BATCH_TILE}; 63 } 64 $if BATCH_TILE >= 4: 65 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { 66 const __m128 va0123 = _mm_loadu_ps(a); 67 a += 4; 68 69 const __m128 vb0123 = _mm_loadu_ps(b); 70 b += 4; 71 72 __m128 vy0123 = ${_MM_OP_PS("va0123", "vb0123")}; 73 vy0123 = _mm_max_ps(vy0123, vy_min); 74 vy0123 = _mm_min_ps(vy0123, vy_max); 75 _mm_storeu_ps(y, vy0123); 76 y += 4; 77 } 78 if XNN_UNLIKELY(n != 0) { 79 const __m128 va0123 = _mm_loadu_ps(a); 80 const __m128 vb0123 = _mm_loadu_ps(b); 81 82 __m128 vy0123 = ${_MM_OP_PS("va0123", "vb0123")}; 83 vy0123 = _mm_max_ps(vy0123, vy_min); 84 vy0123 = _mm_min_ps(vy0123, vy_max); 85 if (n & (2 * sizeof(float))) { 86 _mm_storel_pi((__m64*) y, vy0123); 87 vy0123 = _mm_movehl_ps(vy0123, vy0123); 88 y += 2; 89 } 90 if (n & (1 * sizeof(float))) { 91 _mm_store_ss(y, vy0123); 92 } 93 } 94} 95