Faster arm64 implementation that does not use PMULL instruction (#140)
* Faster arm64 implementation that does not use PMULL instruction * Add NEON version for sliceXormaster
parent
2df03bd4d1
commit
d5afb5f48e
|
@ -8,10 +8,13 @@
|
||||||
package reedsolomon
|
package reedsolomon
|
||||||
|
|
||||||
//go:noescape
|
//go:noescape
|
||||||
func galMulNEON(c uint64, in, out []byte)
|
func galMulNEON(low, high, in, out []byte)
|
||||||
|
|
||||||
//go:noescape
|
//go:noescape
|
||||||
func galMulXorNEON(c uint64, in, out []byte)
|
func galMulXorNEON(low, high, in, out []byte)
|
||||||
|
|
||||||
|
//go:noescape
|
||||||
|
func galXorNEON(in, out []byte)
|
||||||
|
|
||||||
func galMulSlice(c byte, in, out []byte, o *options) {
|
func galMulSlice(c byte, in, out []byte, o *options) {
|
||||||
if c == 1 {
|
if c == 1 {
|
||||||
|
@ -19,7 +22,7 @@ func galMulSlice(c byte, in, out []byte, o *options) {
|
||||||
return
|
return
|
||||||
}
|
}
|
||||||
var done int
|
var done int
|
||||||
galMulNEON(uint64(c), in, out)
|
galMulNEON(mulTableLow[c][:], mulTableHigh[c][:], in, out)
|
||||||
done = (len(in) >> 5) << 5
|
done = (len(in) >> 5) << 5
|
||||||
|
|
||||||
remain := len(in) - done
|
remain := len(in) - done
|
||||||
|
@ -37,7 +40,7 @@ func galMulSliceXor(c byte, in, out []byte, o *options) {
|
||||||
return
|
return
|
||||||
}
|
}
|
||||||
var done int
|
var done int
|
||||||
galMulXorNEON(uint64(c), in, out)
|
galMulXorNEON(mulTableLow[c][:], mulTableHigh[c][:], in, out)
|
||||||
done = (len(in) >> 5) << 5
|
done = (len(in) >> 5) << 5
|
||||||
|
|
||||||
remain := len(in) - done
|
remain := len(in) - done
|
||||||
|
@ -51,7 +54,14 @@ func galMulSliceXor(c byte, in, out []byte, o *options) {
|
||||||
|
|
||||||
// slice galois add
|
// slice galois add
|
||||||
func sliceXor(in, out []byte, o *options) {
|
func sliceXor(in, out []byte, o *options) {
|
||||||
for n, input := range in {
|
|
||||||
out[n] ^= input
|
galXorNEON(in, out)
|
||||||
|
done := (len(in) >> 5) << 5
|
||||||
|
|
||||||
|
remain := len(in) - done
|
||||||
|
if remain > 0 {
|
||||||
|
for i := done; i < len(in); i++ {
|
||||||
|
out[i] ^= in[i]
|
||||||
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
171
galois_arm64.s
171
galois_arm64.s
|
@ -6,80 +6,44 @@
|
||||||
// Use github.com/minio/asm2plan9s on this file to assemble ARM instructions to
|
// Use github.com/minio/asm2plan9s on this file to assemble ARM instructions to
|
||||||
// the opcodes of their Plan9 equivalents
|
// the opcodes of their Plan9 equivalents
|
||||||
|
|
||||||
// polynomial multiplication
|
// func galMulNEON(low, high, in, out []byte)
|
||||||
#define POLYNOMIAL_MULTIPLICATION \
|
|
||||||
WORD $0x0e3ce340 \ // pmull v0.8h,v26.8b,v28.8b
|
|
||||||
WORD $0x4e3ce346 \ // pmull2 v6.8h,v26.16b,v28.16b
|
|
||||||
WORD $0x0e3ce36c \ // pmull v12.8h,v27.8b,v28.8b
|
|
||||||
WORD $0x4e3ce372 // pmull2 v18.8h,v27.16b,v28.16b
|
|
||||||
|
|
||||||
// first reduction
|
|
||||||
#define FIRST_REDUCTION \
|
|
||||||
WORD $0x0f088402 \ // shrn v2.8b, v0.8h, #8
|
|
||||||
WORD $0x0f0884c8 \ // shrn v8.8b, v6.8h, #8
|
|
||||||
WORD $0x0f08858e \ // shrn v14.8b, v12.8h, #8
|
|
||||||
WORD $0x0f088654 \ // shrn v20.8b, v18.8h, #8
|
|
||||||
WORD $0x0e22e3c3 \ // pmull v3.8h,v30.8b,v2.8b
|
|
||||||
WORD $0x0e28e3c9 \ // pmull v9.8h,v30.8b,v8.8b
|
|
||||||
WORD $0x0e2ee3cf \ // pmull v15.8h,v30.8b,v14.8b
|
|
||||||
WORD $0x0e34e3d5 \ // pmull v21.8h,v30.8b,v20.8b
|
|
||||||
WORD $0x6e201c60 \ // eor v0.16b,v3.16b,v0.16b
|
|
||||||
WORD $0x6e261d26 \ // eor v6.16b,v9.16b,v6.16b
|
|
||||||
WORD $0x6e2c1dec \ // eor v12.16b,v15.16b,v12.16b
|
|
||||||
WORD $0x6e321eb2 // eor v18.16b,v21.16b,v18.16b
|
|
||||||
|
|
||||||
// second reduction
|
|
||||||
#define SECOND_REDUCTION \
|
|
||||||
WORD $0x0f088404 \ // shrn v4.8b, v0.8h, #8
|
|
||||||
WORD $0x0f0884ca \ // shrn v10.8b, v6.8h, #8
|
|
||||||
WORD $0x0f088590 \ // shrn v16.8b, v12.8h, #8
|
|
||||||
WORD $0x0f088656 \ // shrn v22.8b, v18.8h, #8
|
|
||||||
WORD $0x6e241c44 \ // eor v4.16b,v2.16b,v4.16b
|
|
||||||
WORD $0x6e2a1d0a \ // eor v10.16b,v8.16b,v10.16b
|
|
||||||
WORD $0x6e301dd0 \ // eor v16.16b,v14.16b,v16.16b
|
|
||||||
WORD $0x6e361e96 \ // eor v22.16b,v20.16b,v22.16b
|
|
||||||
WORD $0x0e24e3c5 \ // pmull v5.8h,v30.8b,v4.8b
|
|
||||||
WORD $0x0e2ae3cb \ // pmull v11.8h,v30.8b,v10.8b
|
|
||||||
WORD $0x0e30e3d1 \ // pmull v17.8h,v30.8b,v16.8b
|
|
||||||
WORD $0x0e36e3d7 \ // pmull v23.8h,v30.8b,v22.8b
|
|
||||||
WORD $0x6e201ca0 \ // eor v0.16b,v5.16b,v0.16b
|
|
||||||
WORD $0x6e261d61 \ // eor v1.16b,v11.16b,v6.16b
|
|
||||||
WORD $0x6e2c1e22 \ // eor v2.16b,v17.16b,v12.16b
|
|
||||||
WORD $0x6e321ee3 // eor v3.16b,v23.16b,v18.16b
|
|
||||||
|
|
||||||
// func galMulNEON(c uint64, in, out []byte)
|
|
||||||
TEXT ·galMulNEON(SB), 7, $0
|
TEXT ·galMulNEON(SB), 7, $0
|
||||||
MOVD c+0(FP), R0
|
MOVD in_base+48(FP), R1
|
||||||
MOVD in_base+8(FP), R1
|
MOVD in_len+56(FP), R2 // length of message
|
||||||
MOVD in_len+16(FP), R2 // length of message
|
MOVD out_base+72(FP), R5
|
||||||
MOVD out_base+32(FP), R5
|
|
||||||
SUBS $32, R2
|
SUBS $32, R2
|
||||||
BMI complete
|
BMI complete
|
||||||
|
|
||||||
// Load constants table pointer
|
MOVD low+0(FP), R10 // R10: &low
|
||||||
MOVD $·constants(SB), R3
|
MOVD high+24(FP), R11 // R11: &high
|
||||||
|
WORD $0x4c407146 // ld1 {v6.16b}, [x10]
|
||||||
|
WORD $0x4c407167 // ld1 {v7.16b}, [x11]
|
||||||
|
|
||||||
// and load constants into v30 & v31
|
MOVD $0x0f, R3
|
||||||
WORD $0x4c40a07e // ld1 {v30.16b-v31.16b}, [x3]
|
WORD $0x4e010c68 // dup v8.16b, w3
|
||||||
|
|
||||||
WORD $0x4e010c1c // dup v28.16b, w0
|
|
||||||
|
|
||||||
loop:
|
loop:
|
||||||
// Main loop
|
// Main loop
|
||||||
WORD $0x4cdfa83a // ld1 {v26.4s-v27.4s}, [x1], #32
|
WORD $0x4cdfa020 // ld1 {v0.16b-v1.16b}, [x1], #32
|
||||||
|
|
||||||
POLYNOMIAL_MULTIPLICATION
|
// Get low input and high input
|
||||||
|
WORD $0x6f0c040a // ushr v10.16b, v0.16b, #4
|
||||||
|
WORD $0x6f0c042b // ushr v11.16b, v1.16b, #4
|
||||||
|
WORD $0x4e281c00 // and v0.16b, v0.16b, v8.16b
|
||||||
|
WORD $0x4e281c21 // and v1.16b, v1.16b, v8.16b
|
||||||
|
|
||||||
FIRST_REDUCTION
|
// Mul low part and mul high part
|
||||||
|
WORD $0x4e0000c4 // tbl v4.16b, {v6.16b}, v0.16b
|
||||||
|
WORD $0x4e0a00e5 // tbl v5.16b, {v7.16b}, v10.16b
|
||||||
|
WORD $0x4e0100ce // tbl v14.16b, {v6.16b}, v1.16b
|
||||||
|
WORD $0x4e0b00ef // tbl v15.16b, {v7.16b}, v11.16b
|
||||||
|
|
||||||
SECOND_REDUCTION
|
// Combine results
|
||||||
|
WORD $0x6e251c84 // eor v4.16b, v4.16b, v5.16b
|
||||||
// combine results
|
WORD $0x6e2f1dc5 // eor v5.16b, v14.16b, v15.16b
|
||||||
WORD $0x4e1f2000 // tbl v0.16b,{v0.16b,v1.16b},v31.16b
|
|
||||||
WORD $0x4e1f2041 // tbl v1.16b,{v2.16b,v3.16b},v31.16b
|
|
||||||
|
|
||||||
// Store result
|
// Store result
|
||||||
WORD $0x4c9faca0 // st1 {v0.2d-v1.2d}, [x5], #32
|
WORD $0x4c9faca4 // st1 {v4.2d-v5.2d}, [x5], #32
|
||||||
|
|
||||||
SUBS $32, R2
|
SUBS $32, R2
|
||||||
BPL loop
|
BPL loop
|
||||||
|
@ -87,42 +51,48 @@ loop:
|
||||||
complete:
|
complete:
|
||||||
RET
|
RET
|
||||||
|
|
||||||
// func galMulXorNEON(c uint64, in, out []byte)
|
|
||||||
|
// func galMulXorNEON(low, high, in, out []byte)
|
||||||
TEXT ·galMulXorNEON(SB), 7, $0
|
TEXT ·galMulXorNEON(SB), 7, $0
|
||||||
MOVD c+0(FP), R0
|
MOVD in_base+48(FP), R1
|
||||||
MOVD in_base+8(FP), R1
|
MOVD in_len+56(FP), R2 // length of message
|
||||||
MOVD in_len+16(FP), R2 // length of message
|
MOVD out_base+72(FP), R5
|
||||||
MOVD out_base+32(FP), R5
|
|
||||||
SUBS $32, R2
|
SUBS $32, R2
|
||||||
BMI completeXor
|
BMI completeXor
|
||||||
|
|
||||||
// Load constants table pointer
|
MOVD low+0(FP), R10 // R10: &low
|
||||||
MOVD $·constants(SB), R3
|
MOVD high+24(FP), R11 // R11: &high
|
||||||
|
WORD $0x4c407146 // ld1 {v6.16b}, [x10]
|
||||||
|
WORD $0x4c407167 // ld1 {v7.16b}, [x11]
|
||||||
|
|
||||||
// and load constants into v30 & v31
|
MOVD $0x0f, R3
|
||||||
WORD $0x4c40a07e // ld1 {v30.16b-v31.16b}, [x3]
|
WORD $0x4e010c68 // dup v8.16b, w3
|
||||||
|
|
||||||
WORD $0x4e010c1c // dup v28.16b, w0
|
|
||||||
|
|
||||||
loopXor:
|
loopXor:
|
||||||
// Main loop
|
// Main loop
|
||||||
WORD $0x4cdfa83a // ld1 {v26.4s-v27.4s}, [x1], #32
|
WORD $0x4cdfa020 // ld1 {v0.16b-v1.16b}, [x1], #32
|
||||||
WORD $0x4c40a8b8 // ld1 {v24.4s-v25.4s}, [x5]
|
WORD $0x4c40a0b4 // ld1 {v20.16b-v21.16b}, [x5]
|
||||||
|
|
||||||
POLYNOMIAL_MULTIPLICATION
|
// Get low input and high input
|
||||||
|
WORD $0x6f0c040a // ushr v10.16b, v0.16b, #4
|
||||||
|
WORD $0x6f0c042b // ushr v11.16b, v1.16b, #4
|
||||||
|
WORD $0x4e281c00 // and v0.16b, v0.16b, v8.16b
|
||||||
|
WORD $0x4e281c21 // and v1.16b, v1.16b, v8.16b
|
||||||
|
|
||||||
FIRST_REDUCTION
|
// Mul low part and mul high part
|
||||||
|
WORD $0x4e0000c4 // tbl v4.16b, {v6.16b}, v0.16b
|
||||||
|
WORD $0x4e0a00e5 // tbl v5.16b, {v7.16b}, v10.16b
|
||||||
|
WORD $0x4e0100ce // tbl v14.16b, {v6.16b}, v1.16b
|
||||||
|
WORD $0x4e0b00ef // tbl v15.16b, {v7.16b}, v11.16b
|
||||||
|
|
||||||
SECOND_REDUCTION
|
// Combine results
|
||||||
|
WORD $0x6e251c84 // eor v4.16b, v4.16b, v5.16b
|
||||||
|
WORD $0x6e2f1dc5 // eor v5.16b, v14.16b, v15.16b
|
||||||
|
WORD $0x6e341c84 // eor v4.16b, v4.16b, v20.16b
|
||||||
|
WORD $0x6e351ca5 // eor v5.16b, v5.16b, v21.16b
|
||||||
|
|
||||||
// combine results
|
// Store result
|
||||||
WORD $0x4e1f2000 // tbl v0.16b,{v0.16b,v1.16b},v31.16b
|
WORD $0x4c9faca4 // st1 {v4.2d-v5.2d}, [x5], #32
|
||||||
WORD $0x4e1f2041 // tbl v1.16b,{v2.16b,v3.16b},v31.16b
|
|
||||||
|
|
||||||
// Xor result and store
|
|
||||||
WORD $0x6e381c00 // eor v0.16b,v0.16b,v24.16b
|
|
||||||
WORD $0x6e391c21 // eor v1.16b,v1.16b,v25.16b
|
|
||||||
WORD $0x4c9faca0 // st1 {v0.2d-v1.2d}, [x5], #32
|
|
||||||
|
|
||||||
SUBS $32, R2
|
SUBS $32, R2
|
||||||
BPL loopXor
|
BPL loopXor
|
||||||
|
@ -130,12 +100,27 @@ loopXor:
|
||||||
completeXor:
|
completeXor:
|
||||||
RET
|
RET
|
||||||
|
|
||||||
// Constants table
|
// func galXorNEON(in, out []byte)
|
||||||
// generating polynomial is 29 (= 0x1d)
|
TEXT ·galXorNEON(SB), 7, $0
|
||||||
DATA ·constants+0x0(SB)/8, $0x1d1d1d1d1d1d1d1d
|
MOVD in_base+0(FP), R1
|
||||||
DATA ·constants+0x8(SB)/8, $0x1d1d1d1d1d1d1d1d
|
MOVD in_len+8(FP), R2 // length of message
|
||||||
// constant for TBL instruction
|
MOVD out_base+24(FP), R5
|
||||||
DATA ·constants+0x10(SB)/8, $0x0e0c0a0806040200
|
SUBS $32, R2
|
||||||
DATA ·constants+0x18(SB)/8, $0x1e1c1a1816141210
|
BMI completeXor
|
||||||
|
|
||||||
GLOBL ·constants(SB), 8, $32
|
loopXor:
|
||||||
|
// Main loop
|
||||||
|
WORD $0x4cdfa020 // ld1 {v0.16b-v1.16b}, [x1], #32
|
||||||
|
WORD $0x4c40a0b4 // ld1 {v20.16b-v21.16b}, [x5]
|
||||||
|
|
||||||
|
WORD $0x6e341c04 // eor v4.16b, v0.16b, v20.16b
|
||||||
|
WORD $0x6e351c25 // eor v5.16b, v1.16b, v21.16b
|
||||||
|
|
||||||
|
// Store result
|
||||||
|
WORD $0x4c9faca4 // st1 {v4.2d-v5.2d}, [x5], #32
|
||||||
|
|
||||||
|
SUBS $32, R2
|
||||||
|
BPL loopXor
|
||||||
|
|
||||||
|
completeXor:
|
||||||
|
RET
|
||||||
|
|
Loading…
Reference in New Issue