1*4882a593Smuzhiyun/* SPDX-License-Identifier: GPL-2.0 */ 2*4882a593Smuzhiyun/* 3*4882a593Smuzhiyun * NH - ε-almost-universal hash function, x86_64 AVX2 accelerated 4*4882a593Smuzhiyun * 5*4882a593Smuzhiyun * Copyright 2018 Google LLC 6*4882a593Smuzhiyun * 7*4882a593Smuzhiyun * Author: Eric Biggers <ebiggers@google.com> 8*4882a593Smuzhiyun */ 9*4882a593Smuzhiyun 10*4882a593Smuzhiyun#include <linux/linkage.h> 11*4882a593Smuzhiyun 12*4882a593Smuzhiyun#define PASS0_SUMS %ymm0 13*4882a593Smuzhiyun#define PASS1_SUMS %ymm1 14*4882a593Smuzhiyun#define PASS2_SUMS %ymm2 15*4882a593Smuzhiyun#define PASS3_SUMS %ymm3 16*4882a593Smuzhiyun#define K0 %ymm4 17*4882a593Smuzhiyun#define K0_XMM %xmm4 18*4882a593Smuzhiyun#define K1 %ymm5 19*4882a593Smuzhiyun#define K1_XMM %xmm5 20*4882a593Smuzhiyun#define K2 %ymm6 21*4882a593Smuzhiyun#define K2_XMM %xmm6 22*4882a593Smuzhiyun#define K3 %ymm7 23*4882a593Smuzhiyun#define K3_XMM %xmm7 24*4882a593Smuzhiyun#define T0 %ymm8 25*4882a593Smuzhiyun#define T1 %ymm9 26*4882a593Smuzhiyun#define T2 %ymm10 27*4882a593Smuzhiyun#define T2_XMM %xmm10 28*4882a593Smuzhiyun#define T3 %ymm11 29*4882a593Smuzhiyun#define T3_XMM %xmm11 30*4882a593Smuzhiyun#define T4 %ymm12 31*4882a593Smuzhiyun#define T5 %ymm13 32*4882a593Smuzhiyun#define T6 %ymm14 33*4882a593Smuzhiyun#define T7 %ymm15 34*4882a593Smuzhiyun#define KEY %rdi 35*4882a593Smuzhiyun#define MESSAGE %rsi 36*4882a593Smuzhiyun#define MESSAGE_LEN %rdx 37*4882a593Smuzhiyun#define HASH %rcx 38*4882a593Smuzhiyun 39*4882a593Smuzhiyun.macro _nh_2xstride k0, k1, k2, k3 40*4882a593Smuzhiyun 41*4882a593Smuzhiyun // Add message words to key words 42*4882a593Smuzhiyun vpaddd \k0, T3, T0 43*4882a593Smuzhiyun vpaddd \k1, T3, T1 44*4882a593Smuzhiyun vpaddd \k2, T3, T2 45*4882a593Smuzhiyun vpaddd \k3, T3, T3 46*4882a593Smuzhiyun 47*4882a593Smuzhiyun // Multiply 32x32 => 64 and accumulate 48*4882a593Smuzhiyun vpshufd $0x10, T0, T4 49*4882a593Smuzhiyun vpshufd $0x32, T0, T0 50*4882a593Smuzhiyun vpshufd $0x10, T1, T5 51*4882a593Smuzhiyun vpshufd $0x32, T1, T1 52*4882a593Smuzhiyun vpshufd $0x10, T2, T6 53*4882a593Smuzhiyun vpshufd $0x32, T2, T2 54*4882a593Smuzhiyun vpshufd $0x10, T3, T7 55*4882a593Smuzhiyun vpshufd $0x32, T3, T3 56*4882a593Smuzhiyun vpmuludq T4, T0, T0 57*4882a593Smuzhiyun vpmuludq T5, T1, T1 58*4882a593Smuzhiyun vpmuludq T6, T2, T2 59*4882a593Smuzhiyun vpmuludq T7, T3, T3 60*4882a593Smuzhiyun vpaddq T0, PASS0_SUMS, PASS0_SUMS 61*4882a593Smuzhiyun vpaddq T1, PASS1_SUMS, PASS1_SUMS 62*4882a593Smuzhiyun vpaddq T2, PASS2_SUMS, PASS2_SUMS 63*4882a593Smuzhiyun vpaddq T3, PASS3_SUMS, PASS3_SUMS 64*4882a593Smuzhiyun.endm 65*4882a593Smuzhiyun 66*4882a593Smuzhiyun/* 67*4882a593Smuzhiyun * void nh_avx2(const u32 *key, const u8 *message, size_t message_len, 68*4882a593Smuzhiyun * u8 hash[NH_HASH_BYTES]) 69*4882a593Smuzhiyun * 70*4882a593Smuzhiyun * It's guaranteed that message_len % 16 == 0. 71*4882a593Smuzhiyun */ 72*4882a593SmuzhiyunSYM_FUNC_START(nh_avx2) 73*4882a593Smuzhiyun 74*4882a593Smuzhiyun vmovdqu 0x00(KEY), K0 75*4882a593Smuzhiyun vmovdqu 0x10(KEY), K1 76*4882a593Smuzhiyun add $0x20, KEY 77*4882a593Smuzhiyun vpxor PASS0_SUMS, PASS0_SUMS, PASS0_SUMS 78*4882a593Smuzhiyun vpxor PASS1_SUMS, PASS1_SUMS, PASS1_SUMS 79*4882a593Smuzhiyun vpxor PASS2_SUMS, PASS2_SUMS, PASS2_SUMS 80*4882a593Smuzhiyun vpxor PASS3_SUMS, PASS3_SUMS, PASS3_SUMS 81*4882a593Smuzhiyun 82*4882a593Smuzhiyun sub $0x40, MESSAGE_LEN 83*4882a593Smuzhiyun jl .Lloop4_done 84*4882a593Smuzhiyun.Lloop4: 85*4882a593Smuzhiyun vmovdqu (MESSAGE), T3 86*4882a593Smuzhiyun vmovdqu 0x00(KEY), K2 87*4882a593Smuzhiyun vmovdqu 0x10(KEY), K3 88*4882a593Smuzhiyun _nh_2xstride K0, K1, K2, K3 89*4882a593Smuzhiyun 90*4882a593Smuzhiyun vmovdqu 0x20(MESSAGE), T3 91*4882a593Smuzhiyun vmovdqu 0x20(KEY), K0 92*4882a593Smuzhiyun vmovdqu 0x30(KEY), K1 93*4882a593Smuzhiyun _nh_2xstride K2, K3, K0, K1 94*4882a593Smuzhiyun 95*4882a593Smuzhiyun add $0x40, MESSAGE 96*4882a593Smuzhiyun add $0x40, KEY 97*4882a593Smuzhiyun sub $0x40, MESSAGE_LEN 98*4882a593Smuzhiyun jge .Lloop4 99*4882a593Smuzhiyun 100*4882a593Smuzhiyun.Lloop4_done: 101*4882a593Smuzhiyun and $0x3f, MESSAGE_LEN 102*4882a593Smuzhiyun jz .Ldone 103*4882a593Smuzhiyun 104*4882a593Smuzhiyun cmp $0x20, MESSAGE_LEN 105*4882a593Smuzhiyun jl .Llast 106*4882a593Smuzhiyun 107*4882a593Smuzhiyun // 2 or 3 strides remain; do 2 more. 108*4882a593Smuzhiyun vmovdqu (MESSAGE), T3 109*4882a593Smuzhiyun vmovdqu 0x00(KEY), K2 110*4882a593Smuzhiyun vmovdqu 0x10(KEY), K3 111*4882a593Smuzhiyun _nh_2xstride K0, K1, K2, K3 112*4882a593Smuzhiyun add $0x20, MESSAGE 113*4882a593Smuzhiyun add $0x20, KEY 114*4882a593Smuzhiyun sub $0x20, MESSAGE_LEN 115*4882a593Smuzhiyun jz .Ldone 116*4882a593Smuzhiyun vmovdqa K2, K0 117*4882a593Smuzhiyun vmovdqa K3, K1 118*4882a593Smuzhiyun.Llast: 119*4882a593Smuzhiyun // Last stride. Zero the high 128 bits of the message and keys so they 120*4882a593Smuzhiyun // don't affect the result when processing them like 2 strides. 121*4882a593Smuzhiyun vmovdqu (MESSAGE), T3_XMM 122*4882a593Smuzhiyun vmovdqa K0_XMM, K0_XMM 123*4882a593Smuzhiyun vmovdqa K1_XMM, K1_XMM 124*4882a593Smuzhiyun vmovdqu 0x00(KEY), K2_XMM 125*4882a593Smuzhiyun vmovdqu 0x10(KEY), K3_XMM 126*4882a593Smuzhiyun _nh_2xstride K0, K1, K2, K3 127*4882a593Smuzhiyun 128*4882a593Smuzhiyun.Ldone: 129*4882a593Smuzhiyun // Sum the accumulators for each pass, then store the sums to 'hash' 130*4882a593Smuzhiyun 131*4882a593Smuzhiyun // PASS0_SUMS is (0A 0B 0C 0D) 132*4882a593Smuzhiyun // PASS1_SUMS is (1A 1B 1C 1D) 133*4882a593Smuzhiyun // PASS2_SUMS is (2A 2B 2C 2D) 134*4882a593Smuzhiyun // PASS3_SUMS is (3A 3B 3C 3D) 135*4882a593Smuzhiyun // We need the horizontal sums: 136*4882a593Smuzhiyun // (0A + 0B + 0C + 0D, 137*4882a593Smuzhiyun // 1A + 1B + 1C + 1D, 138*4882a593Smuzhiyun // 2A + 2B + 2C + 2D, 139*4882a593Smuzhiyun // 3A + 3B + 3C + 3D) 140*4882a593Smuzhiyun // 141*4882a593Smuzhiyun 142*4882a593Smuzhiyun vpunpcklqdq PASS1_SUMS, PASS0_SUMS, T0 // T0 = (0A 1A 0C 1C) 143*4882a593Smuzhiyun vpunpckhqdq PASS1_SUMS, PASS0_SUMS, T1 // T1 = (0B 1B 0D 1D) 144*4882a593Smuzhiyun vpunpcklqdq PASS3_SUMS, PASS2_SUMS, T2 // T2 = (2A 3A 2C 3C) 145*4882a593Smuzhiyun vpunpckhqdq PASS3_SUMS, PASS2_SUMS, T3 // T3 = (2B 3B 2D 3D) 146*4882a593Smuzhiyun 147*4882a593Smuzhiyun vinserti128 $0x1, T2_XMM, T0, T4 // T4 = (0A 1A 2A 3A) 148*4882a593Smuzhiyun vinserti128 $0x1, T3_XMM, T1, T5 // T5 = (0B 1B 2B 3B) 149*4882a593Smuzhiyun vperm2i128 $0x31, T2, T0, T0 // T0 = (0C 1C 2C 3C) 150*4882a593Smuzhiyun vperm2i128 $0x31, T3, T1, T1 // T1 = (0D 1D 2D 3D) 151*4882a593Smuzhiyun 152*4882a593Smuzhiyun vpaddq T5, T4, T4 153*4882a593Smuzhiyun vpaddq T1, T0, T0 154*4882a593Smuzhiyun vpaddq T4, T0, T0 155*4882a593Smuzhiyun vmovdqu T0, (HASH) 156*4882a593Smuzhiyun RET 157*4882a593SmuzhiyunSYM_FUNC_END(nh_avx2) 158