Add avx2/avx512 implementation of LENGTH

go.mod

@@ -5,9 +5,12 @@ go 1.15
 require (
 	github.com/aws/aws-sdk-go v1.37.14
 	github.com/klauspost/compress v1.11.7
+	github.com/frankban/quicktest v1.11.3 // indirect
 	github.com/mmcloughlin/avo v0.0.0-20210120082657-d60cc025fc3c // indirect
 	github.com/pierrec/lz4 v2.6.0+incompatible
 	github.com/pilosa/pilosa v1.4.0
 	github.com/traetox/goaio v0.0.0-20171005222435-46641abceb17
-	golang.org/x/text v0.3.3
+	golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9 // indirect
+	golang.org/x/sys v0.0.0-20200930185726-fdedc70b468f
+	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
 )

pkg/vectorize/length/avx2.s

+// Code generated by command: go run avx2.go -out avx2.s -stubs avx2_stubs.go. DO NOT EDIT.
+
+#include "textflag.h"
+
+// func strLengthAvx2Asm(x []uint32, r []int64)
+// Requires: AVX, AVX2, SSE2
+TEXT ·strLengthAvx2Asm(SB), NOSPLIT, $0-48
+	MOVQ x_base+0(FP), AX
+	MOVQ r_base+24(FP), CX
+	MOVQ x_len+8(FP), DX
+
+blockloop:
+	CMPQ      DX, $0x00000030
+	JL        tailloop
+	MOVOU     (AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, (CX)
+	MOVOU     16(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 32(CX)
+	MOVOU     32(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 64(CX)
+	MOVOU     48(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 96(CX)
+	MOVOU     64(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 128(CX)
+	MOVOU     80(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 160(CX)
+	MOVOU     96(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 192(CX)
+	MOVOU     112(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 224(CX)
+	MOVOU     128(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 256(CX)
+	MOVOU     144(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 288(CX)
+	MOVOU     160(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 320(CX)
+	MOVOU     176(AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, 352(CX)
+	ADDQ      $0x000000c0, AX
+	ADDQ      $0x00000180, CX
+	SUBQ      $0x00000030, DX
+	JMP       blockloop
+
+tailloop:
+	CMPQ      DX, $0x00000004
+	JL        done
+	MOVOU     (AX), X0
+	VPMOVZXDQ X0, Y0
+	VMOVDQU   Y0, (CX)
+	ADDQ      $0x00000010, AX
+	ADDQ      $0x00000020, CX
+	SUBQ      $0x00000004, DX
+	JMP       tailloop
+
+done:
+	RET

pkg/vectorize/length/avx2_stubs.go

+// Code generated by command: go run avx2.go -out avx2.s -stubs avx2_stubs.go. DO NOT EDIT.
+
+package length
+
+func strLengthAvx2Asm(x []uint32, r []int64)

pkg/vectorize/length/avx512.s

+// Code generated by command: go run avx512.go -out avx512.s -stubs avx512_stubs.go. DO NOT EDIT.
+
+#include "textflag.h"
+
+// func strLengthAvx512Asm(x []uint32, r []int64)
+// Requires: AVX, AVX512F
+TEXT ·strLengthAvx512Asm(SB), NOSPLIT, $0-48
+	MOVQ x_base+0(FP), AX
+	MOVQ r_base+24(FP), CX
+	MOVQ x_len+8(FP), DX
+
+blockloop:
+	CMPQ      DX, $0x00000060
+	JL        tailloop
+	VMOVDQU   (AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, (CX)
+	VMOVDQU   16(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 32(CX)
+	VMOVDQU   32(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 64(CX)
+	VMOVDQU   48(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 96(CX)
+	VMOVDQU   64(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 128(CX)
+	VMOVDQU   80(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 160(CX)
+	VMOVDQU   96(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 192(CX)
+	VMOVDQU   112(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 224(CX)
+	VMOVDQU   128(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 256(CX)
+	VMOVDQU   144(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 288(CX)
+	VMOVDQU   160(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 320(CX)
+	VMOVDQU   176(AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, 352(CX)
+	ADDQ      $0x00000180, AX
+	ADDQ      $0x00000300, CX
+	SUBQ      $0x00000060, DX
+	JMP       blockloop
+
+tailloop:
+	CMPQ      DX, $0x00000008
+	JL        done
+	VMOVDQU   (AX), Y0
+	VPMOVZXDQ Y0, Z0
+	VMOVDQU64 Z0, (CX)
+	ADDQ      $0x00000020, AX
+	ADDQ      $0x00000040, CX
+	SUBQ      $0x00000008, DX
+	JMP       tailloop
+
+done:
+	RET

pkg/vectorize/length/avx512_stubs.go

+// Code generated by command: go run avx512.go -out avx512.s -stubs avx512_stubs.go. DO NOT EDIT.
+
+package length
+
+func strLengthAvx512Asm(x []uint32, r []int64)

pkg/vectorize/length/length.go

 package length
 
-import "matrixbase/pkg/container/vector"
+import (
+	"matrixbase/pkg/container/types"
+
+	"golang.org/x/sys/cpu"
+)
 
 var (
-	bytesLength func(*vector.Bytes, []int64) []int64
+	strLength func(*types.Bytes, []int64) []int64
 )
 
 func init() {
-	bytesLength = bytesLengthPure
+	if cpu.X86.HasAVX2 {
+		strLength = strLengthAvx512
+	} else if cpu.X86.HasAVX2 {
+		strLength = strLengthAvx2
+	} else {
+		strLength = strLengthPure
+	}
 }
 
-func BytesLength(xs *vector.Bytes, rs []int64) []int64 {
-	return bytesLength(xs, rs)
+func StrLength(xs *types.Bytes, rs []int64) []int64 {
+	return strLength(xs, rs)
+}
+
+func strLengthAvx2(xs *types.Bytes, rs []int64) []int64 {
+	lengths := xs.Lengths
+	n := len(lengths) / 4
+	strLengthAvx2Asm(lengths, rs)
+	for i, j := n*4, len(lengths); i < j; i++ {
+		rs[i] = int64(lengths[i])
+	}
+	return rs
+}
+
+func strLengthAvx512(xs *types.Bytes, rs []int64) []int64 {
+	lengths := xs.Lengths
+	n := len(lengths) / 8
+	strLengthAvx512Asm(lengths, rs)
+	for i, j := n*8, len(lengths); i < j; i++ {
+		rs[i] = int64(lengths[i])
+	}
+	return rs
 }
 
-func bytesLengthPure(xs *vector.Bytes, rs []int64) []int64 {
-	for i, n := range xs.Ns {
+func strLengthPure(xs *types.Bytes, rs []int64) []int64 {
+	for i, n := range xs.Lengths {
 		rs[i] = int64(n)
 	}
 	return rs