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$assert BATCH_TILE % 16 == 0 7$assert BATCH_TILE >= 16 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB"] 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/intrinsics-polyfill.h> 16#include <xnnpack/vbinary.h> 17 18 19$_MM512_OP_PS = { 20$ "ADD": lambda x, y: "_mm512_add_ps(%s, %s)" % (x, y), 21$ "DIV": lambda x, y: "_mm512_div_ps(%s, %s)" % (x, y), 22$ "MAX": lambda x, y: "_mm512_max_ps(%s, %s)" % (x, y), 23$ "MIN": lambda x, y: "_mm512_min_ps(%s, %s)" % (x, y), 24$ "MUL": lambda x, y: "_mm512_mul_ps(%s, %s)" % (x, y), 25$ "SUB": lambda x, y: "_mm512_sub_ps(%s, %s)" % (x, y), 26$}[OP] 27void xnn_f32_v${OP.lower()}_ukernel__avx512f_x${BATCH_TILE}( 28 size_t n, 29 const float* a, 30 const float* b, 31 float* y, 32 const union xnn_f32_output_params params[restrict static 1]) 33{ 34 assert(n != 0); 35 assert(n % sizeof(float) == 0); 36 37 const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min)); 38 const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max)); 39 40 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 41 const __m512 va${ABC[0:16]} = _mm512_loadu_ps(a); 42 $for N in range(16, BATCH_TILE, 16): 43 const __m512 va${ABC[N:N+16]} = _mm512_loadu_ps(a + ${N}); 44 a += ${BATCH_TILE}; 45 46 const __m512 vb${ABC[0:16]} = _mm512_loadu_ps(b); 47 $for N in range(16, BATCH_TILE, 16): 48 const __m512 vb${ABC[N:N+16]} = _mm512_loadu_ps(b + ${N}); 49 b += ${BATCH_TILE}; 50 51 $for N in range(0, BATCH_TILE, 16): 52 __m512 vy${ABC[N:N+16]} = ${_MM512_OP_PS("va" + ABC[N:N+16], "vb" + ABC[N:N+16])}; 53 54 $for N in range(0, BATCH_TILE, 16): 55 vy${ABC[N:N+16]} = _mm512_max_ps(vy${ABC[N:N+16]}, vy_min); 56 57 $for N in range(0, BATCH_TILE, 16): 58 vy${ABC[N:N+16]} = _mm512_min_ps(vy${ABC[N:N+16]}, vy_max); 59 60 _mm512_storeu_ps(y, vy${ABC[0:16]}); 61 $for N in range(16, BATCH_TILE, 16): 62 _mm512_storeu_ps(y + ${N}, vy${ABC[N:N+16]}); 63 y += ${BATCH_TILE}; 64 } 65 $if BATCH_TILE >= 16: 66 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) { 67 const __m512 va = _mm512_loadu_ps(a); 68 a += 16; 69 70 const __m512 vb = _mm512_loadu_ps(b); 71 b += 16; 72 73 __m512 vy = ${_MM512_OP_PS("va", "vb")}; 74 vy = _mm512_max_ps(vy, vy_min); 75 vy = _mm512_min_ps(vy, vy_max); 76 _mm512_storeu_ps(y, vy); 77 y += 16; 78 } 79 if XNN_UNLIKELY(n != 0) { 80 assert(n >= 1 * sizeof(float)); 81 assert(n <= 15 * sizeof(float)); 82 // Prepare mask for valid 32-bit elements (depends on n). 83 n >>= 2 /* log2(sizeof(float)) */; 84 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1))); 85 86 const __m512 va = _mm512_maskz_loadu_ps(vmask, a); 87 const __m512 vb = _mm512_maskz_loadu_ps(vmask, b); 88 89 __m512 vy = ${_MM512_OP_PS("va", "vb")}; 90 vy = _mm512_max_ps(vy, vy_min); 91 vy = _mm512_min_ps(vy, vy_max); 92 _mm512_mask_storeu_ps(y, vmask, vy); 93 } 94} 95