Upgrading server dependancies (#4566)

12 files changed, 3618 insertions, 132 deletions

diff --git a/vendor/golang.org/x/image/tiff/reader.go b/vendor/golang.org/x/image/tiff/reader.go
index df39e8284..cfe6fed68 100644
--- a/vendor/golang.org/x/image/tiff/reader.go
+++ b/vendor/golang.org/x/image/tiff/reader.go
@@ -118,8 +118,9 @@ func (d *decoder) ifdUint(p []byte) (u []uint, err error) {
 }
 
 // parseIFD decides whether the the IFD entry in p is "interesting" and
-// stows away the data in the decoder.
-func (d *decoder) parseIFD(p []byte) error {
+// stows away the data in the decoder. It returns the tag number of the
+// entry and an error, if any.
+func (d *decoder) parseIFD(p []byte) (int, error) {
 	tag := d.byteOrder.Uint16(p[0:2])
 	switch tag {
 	case tBitsPerSample,
@@ -138,17 +139,17 @@ func (d *decoder) parseIFD(p []byte) error {
 		tImageWidth:
 		val, err := d.ifdUint(p)
 		if err != nil {
-			return err
+			return 0, err
 		}
 		d.features[int(tag)] = val
 	case tColorMap:
 		val, err := d.ifdUint(p)
 		if err != nil {
-			return err
+			return 0, err
 		}
 		numcolors := len(val) / 3
 		if len(val)%3 != 0 || numcolors <= 0 || numcolors > 256 {
-			return FormatError("bad ColorMap length")
+			return 0, FormatError("bad ColorMap length")
 		}
 		d.palette = make([]color.Color, numcolors)
 		for i := 0; i < numcolors; i++ {
@@ -166,15 +167,15 @@ func (d *decoder) parseIFD(p []byte) error {
 		// must terminate the import process gracefully.
 		val, err := d.ifdUint(p)
 		if err != nil {
-			return err
+			return 0, err
 		}
 		for _, v := range val {
 			if v != 1 {
-				return UnsupportedError("sample format")
+				return 0, UnsupportedError("sample format")
 			}
 		}
 	}
-	return nil
+	return int(tag), nil
 }
 
 // readBits reads n bits from the internal buffer starting at the current offset.
@@ -428,10 +429,16 @@ func newDecoder(r io.Reader) (*decoder, error) {
 		return nil, err
 	}
 
+	prevTag := -1
 	for i := 0; i < len(p); i += ifdLen {
-		if err := d.parseIFD(p[i : i+ifdLen]); err != nil {
+		tag, err := d.parseIFD(p[i : i+ifdLen])
+		if err != nil {
 			return nil, err
 		}
+		if tag <= prevTag {
+			return nil, FormatError("tags are not sorted in ascending order")
+		}
+		prevTag = tag
 	}
 
 	d.config.Width = int(d.firstVal(tImageWidth))
diff --git a/vendor/golang.org/x/image/tiff/reader_test.go b/vendor/golang.org/x/image/tiff/reader_test.go
index f5c02e697..f1cf93bbe 100644
--- a/vendor/golang.org/x/image/tiff/reader_test.go
+++ b/vendor/golang.org/x/image/tiff/reader_test.go
@@ -192,6 +192,24 @@ func TestDecodeLZW(t *testing.T) {
 	compare(t, img0, img1)
 }
 
+// TestDecodeTagOrder tests that a malformed image with unsorted IFD entries is
+// correctly rejected.
+func TestDecodeTagOrder(t *testing.T) {
+	data, err := ioutil.ReadFile("../testdata/video-001.tiff")
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Swap the first two IFD entries.
+	ifdOffset := int64(binary.LittleEndian.Uint32(data[4:8]))
+	for i := ifdOffset + 2; i < ifdOffset+14; i++ {
+		data[i], data[i+12] = data[i+12], data[i]
+	}
+	if _, _, err := image.Decode(bytes.NewReader(data)); err == nil {
+		t.Fatal("got nil error, want non-nil")
+	}
+}
+
 // TestDecompress tests that decoding some TIFF images that use different
 // compression formats result in the same pixel data.
 func TestDecompress(t *testing.T) {
diff --git a/vendor/golang.org/x/image/vector/acc_amd64.go b/vendor/golang.org/x/image/vector/acc_amd64.go
new file mode 100644
index 000000000..68f6e030c
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/acc_amd64.go
@@ -0,0 +1,34 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build go1.6
+// +build !noasm
+
+package vector
+
+func haveSSE4_1() bool
+
+var haveFixedAccumulateSIMD = haveSSE4_1()
+
+const haveFloatingAccumulateSIMD = true
+
+//go:noescape
+func fixedAccumulateOpOverSIMD(dst []uint8, src []uint32)
+
+//go:noescape
+func fixedAccumulateOpSrcSIMD(dst []uint8, src []uint32)
+
+//go:noescape
+func fixedAccumulateMaskSIMD(buf []uint32)
+
+//go:noescape
+func floatingAccumulateOpOverSIMD(dst []uint8, src []float32)
+
+//go:noescape
+func floatingAccumulateOpSrcSIMD(dst []uint8, src []float32)
+
+//go:noescape
+func floatingAccumulateMaskSIMD(dst []uint32, src []float32)
diff --git a/vendor/golang.org/x/image/vector/acc_amd64.s b/vendor/golang.org/x/image/vector/acc_amd64.s
new file mode 100644
index 000000000..6a424bcdd
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/acc_amd64.s
@@ -0,0 +1,1083 @@
+// generated by go run gen.go; DO NOT EDIT
+
+// +build !appengine
+// +build gc
+// +build go1.6
+// +build !noasm
+
+#include "textflag.h"
+
+// fl is short for floating point math. fx is short for fixed point math.
+
+DATA flAlmost65536<>+0x00(SB)/8, $0x477fffff477fffff
+DATA flAlmost65536<>+0x08(SB)/8, $0x477fffff477fffff
+DATA flOne<>+0x00(SB)/8, $0x3f8000003f800000
+DATA flOne<>+0x08(SB)/8, $0x3f8000003f800000
+DATA flSignMask<>+0x00(SB)/8, $0x7fffffff7fffffff
+DATA flSignMask<>+0x08(SB)/8, $0x7fffffff7fffffff
+
+// scatterAndMulBy0x101 is a PSHUFB mask that brings the low four bytes of an
+// XMM register to the low byte of that register's four uint32 values. It
+// duplicates those bytes, effectively multiplying each uint32 by 0x101.
+//
+// It transforms a little-endian 16-byte XMM value from
+//	ijkl????????????
+// to
+//	ii00jj00kk00ll00
+DATA scatterAndMulBy0x101<>+0x00(SB)/8, $0x8080010180800000
+DATA scatterAndMulBy0x101<>+0x08(SB)/8, $0x8080030380800202
+
+// gather is a PSHUFB mask that brings the second-lowest byte of the XMM
+// register's four uint32 values to the low four bytes of that register.
+//
+// It transforms a little-endian 16-byte XMM value from
+//	?i???j???k???l??
+// to
+//	ijkl000000000000
+DATA gather<>+0x00(SB)/8, $0x808080800d090501
+DATA gather<>+0x08(SB)/8, $0x8080808080808080
+
+DATA fxAlmost65536<>+0x00(SB)/8, $0x0000ffff0000ffff
+DATA fxAlmost65536<>+0x08(SB)/8, $0x0000ffff0000ffff
+DATA inverseFFFF<>+0x00(SB)/8, $0x8000800180008001
+DATA inverseFFFF<>+0x08(SB)/8, $0x8000800180008001
+
+GLOBL flAlmost65536<>(SB), (NOPTR+RODATA), $16
+GLOBL flOne<>(SB), (NOPTR+RODATA), $16
+GLOBL flSignMask<>(SB), (NOPTR+RODATA), $16
+GLOBL scatterAndMulBy0x101<>(SB), (NOPTR+RODATA), $16
+GLOBL gather<>(SB), (NOPTR+RODATA), $16
+GLOBL fxAlmost65536<>(SB), (NOPTR+RODATA), $16
+GLOBL inverseFFFF<>(SB), (NOPTR+RODATA), $16
+
+// func haveSSE4_1() bool
+TEXT ·haveSSE4_1(SB), NOSPLIT, $0
+	MOVQ $1, AX
+	CPUID
+	SHRQ $19, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func fixedAccumulateOpOverSIMD(dst []uint8, src []uint32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	-
+//	xmm4	-
+//	xmm5	fxAlmost65536
+//	xmm6	gather
+//	xmm7	offset
+//	xmm8	scatterAndMulBy0x101
+//	xmm9	fxAlmost65536
+//	xmm10	inverseFFFF
+TEXT ·fixedAccumulateOpOverSIMD(SB), NOSPLIT, $0-48
+
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R10
+
+	// Sanity check that len(dst) >= len(src).
+	CMPQ BX, R10
+	JLT  fxAccOpOverEnd
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
+	MOVOU fxAlmost65536<>(SB), X5
+
+	// gather               := XMM(see above)                      // PSHUFB shuffle mask.
+	// scatterAndMulBy0x101 := XMM(see above)                      // PSHUFB shuffle mask.
+	// fxAlmost65536        := XMM(0x0000ffff repeated four times) // 0xffff.
+	// inverseFFFF          := XMM(0x80008001 repeated four times) // Magic constant for dividing by 0xffff.
+	MOVOU gather<>(SB), X6
+	MOVOU scatterAndMulBy0x101<>(SB), X8
+	MOVOU fxAlmost65536<>(SB), X9
+	MOVOU inverseFFFF<>(SB), X10
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+fxAccOpOverLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  fxAccOpOverLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	PADDD X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	PADDD  X0, X1
+
+	// x += offset
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// Blend over the dst's prior value. SIMD for i in 0..3:
+	//
+	// dstA := uint32(dst[i]) * 0x101
+	// maskA := z@i
+	// outA := dstA*(0xffff-maskA)/0xffff + maskA
+	// dst[i] = uint8(outA >> 8)
+	//
+	// First, set X0 to dstA*(0xfff-maskA).
+	MOVL   (DI), X0
+	PSHUFB X8, X0
+	MOVOU  X9, X11
+	PSUBL  X2, X11
+	PMULLD X11, X0
+
+	// We implement uint32 division by 0xffff as multiplication by a magic
+	// constant (0x800080001) and then a shift by a magic constant (47).
+	// See TestDivideByFFFF for a justification.
+	//
+	// That multiplication widens from uint32 to uint64, so we have to
+	// duplicate and shift our four uint32s from one XMM register (X0) to
+	// two XMM registers (X0 and X11).
+	//
+	// Move the second and fourth uint32s in X0 to be the first and third
+	// uint32s in X11.
+	MOVOU X0, X11
+	PSRLQ $32, X11
+
+	// Multiply by magic, shift by magic.
+	//
+	// pmuludq %xmm10,%xmm0
+	// pmuludq %xmm10,%xmm11
+	BYTE  $0x66; BYTE $0x41; BYTE $0x0f; BYTE $0xf4; BYTE $0xc2
+	BYTE  $0x66; BYTE $0x45; BYTE $0x0f; BYTE $0xf4; BYTE $0xda
+	PSRLQ $47, X0
+	PSRLQ $47, X11
+
+	// Merge the two registers back to one, X11, and add maskA.
+	PSLLQ $32, X11
+	XORPS X0, X11
+	PADDD X11, X2
+
+	// As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
+	PSHUFB X6, X2
+	MOVL   X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $4, DI
+	ADDQ $16, SI
+	JMP  fxAccOpOverLoop4
+
+fxAccOpOverLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  fxAccOpOverEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// Blend over the dst's prior value.
+	//
+	// dstA := uint32(dst[0]) * 0x101
+	// maskA := z
+	// outA := dstA*(0xffff-maskA)/0xffff + maskA
+	// dst[0] = uint8(outA >> 8)
+	MOVBLZX (DI), R12
+	IMULL   $0x101, R12
+	MOVL    X2, R13
+	MOVL    $0xffff, AX
+	SUBL    R13, AX
+	MULL    R12             // MULL's implicit arg is AX, and the result is stored in DX:AX.
+	MOVL    $0x80008001, BX // Divide by 0xffff is to first multiply by a magic constant...
+	MULL    BX              // MULL's implicit arg is AX, and the result is stored in DX:AX.
+	SHRL    $15, DX         // ...and then shift by another magic constant (47 - 32 = 15).
+	ADDL    DX, R13
+	SHRL    $8, R13
+	MOVB    R13, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $1, DI
+	ADDQ $4, SI
+	JMP  fxAccOpOverLoop1
+
+fxAccOpOverEnd:
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func fixedAccumulateOpSrcSIMD(dst []uint8, src []uint32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	-
+//	xmm4	-
+//	xmm5	fxAlmost65536
+//	xmm6	gather
+//	xmm7	offset
+//	xmm8	-
+//	xmm9	-
+//	xmm10	-
+TEXT ·fixedAccumulateOpSrcSIMD(SB), NOSPLIT, $0-48
+
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R10
+
+	// Sanity check that len(dst) >= len(src).
+	CMPQ BX, R10
+	JLT  fxAccOpSrcEnd
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
+	MOVOU fxAlmost65536<>(SB), X5
+
+	// gather := XMM(see above) // PSHUFB shuffle mask.
+	MOVOU gather<>(SB), X6
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+fxAccOpSrcLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  fxAccOpSrcLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	PADDD X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	PADDD  X0, X1
+
+	// x += offset
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
+	// copy(dst[:4], low4BytesOf(z))
+	PSHUFB X6, X2
+	MOVL   X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $4, DI
+	ADDQ $16, SI
+	JMP  fxAccOpSrcLoop4
+
+fxAccOpSrcLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  fxAccOpSrcEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// dst[0] = uint8(z>>8)
+	MOVL X2, BX
+	SHRL $8, BX
+	MOVB BX, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $1, DI
+	ADDQ $4, SI
+	JMP  fxAccOpSrcLoop1
+
+fxAccOpSrcEnd:
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func fixedAccumulateMaskSIMD(buf []uint32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	-
+//	xmm4	-
+//	xmm5	fxAlmost65536
+//	xmm6	-
+//	xmm7	offset
+//	xmm8	-
+//	xmm9	-
+//	xmm10	-
+TEXT ·fixedAccumulateMaskSIMD(SB), NOSPLIT, $0-24
+
+	MOVQ buf_base+0(FP), DI
+	MOVQ buf_len+8(FP), BX
+	MOVQ buf_base+0(FP), SI
+	MOVQ buf_len+8(FP), R10
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
+	MOVOU fxAlmost65536<>(SB), X5
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+fxAccMaskLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  fxAccMaskLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	PADDD X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	PADDD  X0, X1
+
+	// x += offset
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// copy(dst[:4], z)
+	MOVOU X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $16, DI
+	ADDQ $16, SI
+	JMP  fxAccMaskLoop4
+
+fxAccMaskLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  fxAccMaskEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	PADDD X7, X1
+
+	// y = abs(x)
+	// y >>= 2 // Shift by 2*ϕ - 16.
+	// y = min(y, fxAlmost65536)
+	//
+	// pabsd  %xmm1,%xmm2
+	// psrld  $0x2,%xmm2
+	// pminud %xmm5,%xmm2
+	//
+	// Hopefully we'll get these opcode mnemonics into the assembler for Go
+	// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+	// it's similar.
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+	BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+	BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+
+	// z = convertToInt32(y)
+	// No-op.
+
+	// dst[0] = uint32(z)
+	MOVL X2, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $4, DI
+	ADDQ $4, SI
+	JMP  fxAccMaskLoop1
+
+fxAccMaskEnd:
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func floatingAccumulateOpOverSIMD(dst []uint8, src []float32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	flSignMask
+//	xmm4	flOne
+//	xmm5	flAlmost65536
+//	xmm6	gather
+//	xmm7	offset
+//	xmm8	scatterAndMulBy0x101
+//	xmm9	fxAlmost65536
+//	xmm10	inverseFFFF
+TEXT ·floatingAccumulateOpOverSIMD(SB), NOSPLIT, $8-48
+
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R10
+
+	// Sanity check that len(dst) >= len(src).
+	CMPQ BX, R10
+	JLT  flAccOpOverEnd
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
+	// "Round To Zero".
+	STMXCSR mxcsrOrig-8(SP)
+	MOVL    mxcsrOrig-8(SP), AX
+	ORL     $0x6000, AX
+	MOVL    AX, mxcsrNew-4(SP)
+
+	// flSignMask    := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
+	// flOne         := XMM(0x3f800000 repeated four times) // 1 as a float32.
+	// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+	MOVOU flSignMask<>(SB), X3
+	MOVOU flOne<>(SB), X4
+	MOVOU flAlmost65536<>(SB), X5
+
+	// gather               := XMM(see above)                      // PSHUFB shuffle mask.
+	// scatterAndMulBy0x101 := XMM(see above)                      // PSHUFB shuffle mask.
+	// fxAlmost65536        := XMM(0x0000ffff repeated four times) // 0xffff.
+	// inverseFFFF          := XMM(0x80008001 repeated four times) // Magic constant for dividing by 0xffff.
+	MOVOU gather<>(SB), X6
+	MOVOU scatterAndMulBy0x101<>(SB), X8
+	MOVOU fxAlmost65536<>(SB), X9
+	MOVOU inverseFFFF<>(SB), X10
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+flAccOpOverLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  flAccOpOverLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	ADDPS X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	ADDPS  X0, X1
+
+	// x += offset
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// Blend over the dst's prior value. SIMD for i in 0..3:
+	//
+	// dstA := uint32(dst[i]) * 0x101
+	// maskA := z@i
+	// outA := dstA*(0xffff-maskA)/0xffff + maskA
+	// dst[i] = uint8(outA >> 8)
+	//
+	// First, set X0 to dstA*(0xfff-maskA).
+	MOVL   (DI), X0
+	PSHUFB X8, X0
+	MOVOU  X9, X11
+	PSUBL  X2, X11
+	PMULLD X11, X0
+
+	// We implement uint32 division by 0xffff as multiplication by a magic
+	// constant (0x800080001) and then a shift by a magic constant (47).
+	// See TestDivideByFFFF for a justification.
+	//
+	// That multiplication widens from uint32 to uint64, so we have to
+	// duplicate and shift our four uint32s from one XMM register (X0) to
+	// two XMM registers (X0 and X11).
+	//
+	// Move the second and fourth uint32s in X0 to be the first and third
+	// uint32s in X11.
+	MOVOU X0, X11
+	PSRLQ $32, X11
+
+	// Multiply by magic, shift by magic.
+	//
+	// pmuludq %xmm10,%xmm0
+	// pmuludq %xmm10,%xmm11
+	BYTE  $0x66; BYTE $0x41; BYTE $0x0f; BYTE $0xf4; BYTE $0xc2
+	BYTE  $0x66; BYTE $0x45; BYTE $0x0f; BYTE $0xf4; BYTE $0xda
+	PSRLQ $47, X0
+	PSRLQ $47, X11
+
+	// Merge the two registers back to one, X11, and add maskA.
+	PSLLQ $32, X11
+	XORPS X0, X11
+	PADDD X11, X2
+
+	// As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
+	PSHUFB X6, X2
+	MOVL   X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $4, DI
+	ADDQ $16, SI
+	JMP  flAccOpOverLoop4
+
+flAccOpOverLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  flAccOpOverEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// Blend over the dst's prior value.
+	//
+	// dstA := uint32(dst[0]) * 0x101
+	// maskA := z
+	// outA := dstA*(0xffff-maskA)/0xffff + maskA
+	// dst[0] = uint8(outA >> 8)
+	MOVBLZX (DI), R12
+	IMULL   $0x101, R12
+	MOVL    X2, R13
+	MOVL    $0xffff, AX
+	SUBL    R13, AX
+	MULL    R12             // MULL's implicit arg is AX, and the result is stored in DX:AX.
+	MOVL    $0x80008001, BX // Divide by 0xffff is to first multiply by a magic constant...
+	MULL    BX              // MULL's implicit arg is AX, and the result is stored in DX:AX.
+	SHRL    $15, DX         // ...and then shift by another magic constant (47 - 32 = 15).
+	ADDL    DX, R13
+	SHRL    $8, R13
+	MOVB    R13, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $1, DI
+	ADDQ $4, SI
+	JMP  flAccOpOverLoop1
+
+flAccOpOverEnd:
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func floatingAccumulateOpSrcSIMD(dst []uint8, src []float32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	flSignMask
+//	xmm4	flOne
+//	xmm5	flAlmost65536
+//	xmm6	gather
+//	xmm7	offset
+//	xmm8	-
+//	xmm9	-
+//	xmm10	-
+TEXT ·floatingAccumulateOpSrcSIMD(SB), NOSPLIT, $8-48
+
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R10
+
+	// Sanity check that len(dst) >= len(src).
+	CMPQ BX, R10
+	JLT  flAccOpSrcEnd
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
+	// "Round To Zero".
+	STMXCSR mxcsrOrig-8(SP)
+	MOVL    mxcsrOrig-8(SP), AX
+	ORL     $0x6000, AX
+	MOVL    AX, mxcsrNew-4(SP)
+
+	// flSignMask    := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
+	// flOne         := XMM(0x3f800000 repeated four times) // 1 as a float32.
+	// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+	MOVOU flSignMask<>(SB), X3
+	MOVOU flOne<>(SB), X4
+	MOVOU flAlmost65536<>(SB), X5
+
+	// gather := XMM(see above) // PSHUFB shuffle mask.
+	MOVOU gather<>(SB), X6
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+flAccOpSrcLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  flAccOpSrcLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	ADDPS X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	ADDPS  X0, X1
+
+	// x += offset
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
+	// copy(dst[:4], low4BytesOf(z))
+	PSHUFB X6, X2
+	MOVL   X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $4, DI
+	ADDQ $16, SI
+	JMP  flAccOpSrcLoop4
+
+flAccOpSrcLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  flAccOpSrcEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// dst[0] = uint8(z>>8)
+	MOVL X2, BX
+	SHRL $8, BX
+	MOVB BX, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $1, DI
+	ADDQ $4, SI
+	JMP  flAccOpSrcLoop1
+
+flAccOpSrcEnd:
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func floatingAccumulateMaskSIMD(dst []uint32, src []float32)
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	flSignMask
+//	xmm4	flOne
+//	xmm5	flAlmost65536
+//	xmm6	-
+//	xmm7	offset
+//	xmm8	-
+//	xmm9	-
+//	xmm10	-
+TEXT ·floatingAccumulateMaskSIMD(SB), NOSPLIT, $8-48
+
+	MOVQ dst_base+0(FP), DI
+	MOVQ dst_len+8(FP), BX
+	MOVQ src_base+24(FP), SI
+	MOVQ src_len+32(FP), R10
+
+	// Sanity check that len(dst) >= len(src).
+	CMPQ BX, R10
+	JLT  flAccMaskEnd
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
+	// "Round To Zero".
+	STMXCSR mxcsrOrig-8(SP)
+	MOVL    mxcsrOrig-8(SP), AX
+	ORL     $0x6000, AX
+	MOVL    AX, mxcsrNew-4(SP)
+
+	// flSignMask    := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
+	// flOne         := XMM(0x3f800000 repeated four times) // 1 as a float32.
+	// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+	MOVOU flSignMask<>(SB), X3
+	MOVOU flOne<>(SB), X4
+	MOVOU flAlmost65536<>(SB), X5
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+flAccMaskLoop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  flAccMaskLoop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU X1, X0
+	PSLLO $4, X0
+	ADDPS X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS  X0, X0
+	SHUFPS $0x40, X1, X0
+	ADDPS  X0, X1
+
+	// x += offset
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// copy(dst[:4], z)
+	MOVOU X2, (DI)
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ $16, DI
+	ADDQ $16, SI
+	JMP  flAccMaskLoop4
+
+flAccMaskLoop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  flAccMaskEnd
+
+	// x = src[i] + offset
+	MOVL  (SI), X1
+	ADDPS X7, X1
+
+	// y = x & flSignMask
+	// y = min(y, flOne)
+	// y = mul(y, flAlmost65536)
+	MOVOU X3, X2
+	ANDPS X1, X2
+	MINPS X4, X2
+	MULPS X5, X2
+
+	// z = convertToInt32(y)
+	LDMXCSR  mxcsrNew-4(SP)
+	CVTPS2PL X2, X2
+	LDMXCSR  mxcsrOrig-8(SP)
+
+	// dst[0] = uint32(z)
+	MOVL X2, (DI)
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ $4, DI
+	ADDQ $4, SI
+	JMP  flAccMaskLoop1
+
+flAccMaskEnd:
+	RET
diff --git a/vendor/golang.org/x/image/vector/acc_other.go b/vendor/golang.org/x/image/vector/acc_other.go
new file mode 100644
index 000000000..30425beed
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/acc_other.go
@@ -0,0 +1,17 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !amd64 appengine !gc !go1.6 noasm
+
+package vector
+
+const haveFixedAccumulateSIMD = false
+const haveFloatingAccumulateSIMD = false
+
+func fixedAccumulateOpOverSIMD(dst []uint8, src []uint32)     {}
+func fixedAccumulateOpSrcSIMD(dst []uint8, src []uint32)      {}
+func fixedAccumulateMaskSIMD(buf []uint32)                    {}
+func floatingAccumulateOpOverSIMD(dst []uint8, src []float32) {}
+func floatingAccumulateOpSrcSIMD(dst []uint8, src []float32)  {}
+func floatingAccumulateMaskSIMD(dst []uint32, src []float32)  {}
diff --git a/vendor/golang.org/x/image/vector/acc_test.go b/vendor/golang.org/x/image/vector/acc_test.go
new file mode 100644
index 000000000..d80f7651c
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/acc_test.go
@@ -0,0 +1,651 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package vector
+
+import (
+	"bytes"
+	"fmt"
+	"math/rand"
+	"testing"
+)
+
+// TestDivideByFFFF tests that dividing by 0xffff is equivalent to multiplying
+// and then shifting by magic constants. The Go compiler itself issues this
+// multiply-and-shift for a division by the constant value 0xffff. This trick
+// is used in the asm code as the GOARCH=amd64 SIMD instructions have parallel
+// multiply but not parallel divide.
+//
+// There's undoubtedly a justification somewhere in Hacker's Delight chapter 10
+// "Integer Division by Constants", but I don't have a more specific link.
+//
+// http://www.hackersdelight.org/divcMore.pdf and
+// http://www.hackersdelight.org/magic.htm
+func TestDivideByFFFF(t *testing.T) {
+	const mul, shift = 0x80008001, 47
+	rng := rand.New(rand.NewSource(1))
+	for i := 0; i < 20000; i++ {
+		u := rng.Uint32()
+		got := uint32((uint64(u) * mul) >> shift)
+		want := u / 0xffff
+		if got != want {
+			t.Fatalf("i=%d, u=%#08x: got %#08x, want %#08x", i, u, got, want)
+		}
+	}
+}
+
+// TestXxxSIMDUnaligned tests that unaligned SIMD loads/stores don't crash.
+
+func TestFixedAccumulateSIMDUnaligned(t *testing.T) {
+	if !haveFixedAccumulateSIMD {
+		t.Skip("No SIMD implemention")
+	}
+
+	dst := make([]uint8, 64)
+	src := make([]uint32, 64)
+	for d := 0; d < 16; d++ {
+		for s := 0; s < 16; s++ {
+			fixedAccumulateOpSrcSIMD(dst[d:d+32], src[s:s+32])
+		}
+	}
+}
+
+func TestFloatingAccumulateSIMDUnaligned(t *testing.T) {
+	if !haveFloatingAccumulateSIMD {
+		t.Skip("No SIMD implemention")
+	}
+
+	dst := make([]uint8, 64)
+	src := make([]float32, 64)
+	for d := 0; d < 16; d++ {
+		for s := 0; s < 16; s++ {
+			floatingAccumulateOpSrcSIMD(dst[d:d+32], src[s:s+32])
+		}
+	}
+}
+
+// TestXxxSIMDShortDst tests that the SIMD implementations don't write past the
+// end of the dst buffer.
+
+func TestFixedAccumulateSIMDShortDst(t *testing.T) {
+	if !haveFixedAccumulateSIMD {
+		t.Skip("No SIMD implemention")
+	}
+
+	const oneQuarter = uint32(int2ϕ(fxOne*fxOne)) / 4
+	src := []uint32{oneQuarter, oneQuarter, oneQuarter, oneQuarter}
+	for i := 0; i < 4; i++ {
+		dst := make([]uint8, 4)
+		fixedAccumulateOpSrcSIMD(dst[:i], src[:i])
+		for j := range dst {
+			if j < i {
+				if got := dst[j]; got == 0 {
+					t.Errorf("i=%d, j=%d: got %#02x, want non-zero", i, j, got)
+				}
+			} else {
+				if got := dst[j]; got != 0 {
+					t.Errorf("i=%d, j=%d: got %#02x, want zero", i, j, got)
+				}
+			}
+		}
+	}
+}
+
+func TestFloatingAccumulateSIMDShortDst(t *testing.T) {
+	if !haveFloatingAccumulateSIMD {
+		t.Skip("No SIMD implemention")
+	}
+
+	const oneQuarter = 0.25
+	src := []float32{oneQuarter, oneQuarter, oneQuarter, oneQuarter}
+	for i := 0; i < 4; i++ {
+		dst := make([]uint8, 4)
+		floatingAccumulateOpSrcSIMD(dst[:i], src[:i])
+		for j := range dst {
+			if j < i {
+				if got := dst[j]; got == 0 {
+					t.Errorf("i=%d, j=%d: got %#02x, want non-zero", i, j, got)
+				}
+			} else {
+				if got := dst[j]; got != 0 {
+					t.Errorf("i=%d, j=%d: got %#02x, want zero", i, j, got)
+				}
+			}
+		}
+	}
+}
+
+func TestFixedAccumulateOpOverShort(t *testing.T)    { testAcc(t, fxInShort, fxMaskShort, "over") }
+func TestFixedAccumulateOpSrcShort(t *testing.T)     { testAcc(t, fxInShort, fxMaskShort, "src") }
+func TestFixedAccumulateMaskShort(t *testing.T)      { testAcc(t, fxInShort, fxMaskShort, "mask") }
+func TestFloatingAccumulateOpOverShort(t *testing.T) { testAcc(t, flInShort, flMaskShort, "over") }
+func TestFloatingAccumulateOpSrcShort(t *testing.T)  { testAcc(t, flInShort, flMaskShort, "src") }
+func TestFloatingAccumulateMaskShort(t *testing.T)   { testAcc(t, flInShort, flMaskShort, "mask") }
+
+func TestFixedAccumulateOpOver16(t *testing.T)    { testAcc(t, fxIn16, fxMask16, "over") }
+func TestFixedAccumulateOpSrc16(t *testing.T)     { testAcc(t, fxIn16, fxMask16, "src") }
+func TestFixedAccumulateMask16(t *testing.T)      { testAcc(t, fxIn16, fxMask16, "mask") }
+func TestFloatingAccumulateOpOver16(t *testing.T) { testAcc(t, flIn16, flMask16, "over") }
+func TestFloatingAccumulateOpSrc16(t *testing.T)  { testAcc(t, flIn16, flMask16, "src") }
+func TestFloatingAccumulateMask16(t *testing.T)   { testAcc(t, flIn16, flMask16, "mask") }
+
+func testAcc(t *testing.T, in interface{}, mask []uint32, op string) {
+	for _, simd := range []bool{false, true} {
+		maxN := 0
+		switch in := in.(type) {
+		case []uint32:
+			if simd && !haveFixedAccumulateSIMD {
+				continue
+			}
+			maxN = len(in)
+		case []float32:
+			if simd && !haveFloatingAccumulateSIMD {
+				continue
+			}
+			maxN = len(in)
+		}
+
+		for _, n := range []int{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
+			33, 55, 79, 96, 120, 165, 256, maxN} {
+
+			if n > maxN {
+				continue
+			}
+
+			var (
+				got8, want8   []uint8
+				got32, want32 []uint32
+			)
+			switch op {
+			case "over":
+				const background = 0x40
+				got8 = make([]uint8, n)
+				for i := range got8 {
+					got8[i] = background
+				}
+				want8 = make([]uint8, n)
+				for i := range want8 {
+					dstA := uint32(background * 0x101)
+					maskA := mask[i]
+					outA := dstA*(0xffff-maskA)/0xffff + maskA
+					want8[i] = uint8(outA >> 8)
+				}
+
+			case "src":
+				got8 = make([]uint8, n)
+				want8 = make([]uint8, n)
+				for i := range want8 {
+					want8[i] = uint8(mask[i] >> 8)
+				}
+
+			case "mask":
+				got32 = make([]uint32, n)
+				want32 = mask[:n]
+			}
+
+			switch in := in.(type) {
+			case []uint32:
+				switch op {
+				case "over":
+					if simd {
+						fixedAccumulateOpOverSIMD(got8, in[:n])
+					} else {
+						fixedAccumulateOpOver(got8, in[:n])
+					}
+				case "src":
+					if simd {
+						fixedAccumulateOpSrcSIMD(got8, in[:n])
+					} else {
+						fixedAccumulateOpSrc(got8, in[:n])
+					}
+				case "mask":
+					copy(got32, in[:n])
+					if simd {
+						fixedAccumulateMaskSIMD(got32)
+					} else {
+						fixedAccumulateMask(got32)
+					}
+				}
+			case []float32:
+				switch op {
+				case "over":
+					if simd {
+						floatingAccumulateOpOverSIMD(got8, in[:n])
+					} else {
+						floatingAccumulateOpOver(got8, in[:n])
+					}
+				case "src":
+					if simd {
+						floatingAccumulateOpSrcSIMD(got8, in[:n])
+					} else {
+						floatingAccumulateOpSrc(got8, in[:n])
+					}
+				case "mask":
+					if simd {
+						floatingAccumulateMaskSIMD(got32, in[:n])
+					} else {
+						floatingAccumulateMask(got32, in[:n])
+					}
+				}
+			}
+
+			if op != "mask" {
+				if !bytes.Equal(got8, want8) {
+					t.Errorf("simd=%t, n=%d:\ngot:  % x\nwant: % x", simd, n, got8, want8)
+				}
+			} else {
+				if !uint32sEqual(got32, want32) {
+					t.Errorf("simd=%t, n=%d:\ngot:  % x\nwant: % x", simd, n, got32, want32)
+				}
+			}
+		}
+	}
+}
+
+func uint32sEqual(xs, ys []uint32) bool {
+	if len(xs) != len(ys) {
+		return false
+	}
+	for i := range xs {
+		if xs[i] != ys[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func float32sEqual(xs, ys []float32) bool {
+	if len(xs) != len(ys) {
+		return false
+	}
+	for i := range xs {
+		if xs[i] != ys[i] {
+			return false
+		}
+	}
+	return true
+}
+
+func BenchmarkFixedAccumulateOpOver16(b *testing.B)        { benchAcc(b, fxIn16, "over", false) }
+func BenchmarkFixedAccumulateOpOverSIMD16(b *testing.B)    { benchAcc(b, fxIn16, "over", true) }
+func BenchmarkFixedAccumulateOpSrc16(b *testing.B)         { benchAcc(b, fxIn16, "src", false) }
+func BenchmarkFixedAccumulateOpSrcSIMD16(b *testing.B)     { benchAcc(b, fxIn16, "src", true) }
+func BenchmarkFixedAccumulateMask16(b *testing.B)          { benchAcc(b, fxIn16, "mask", false) }
+func BenchmarkFixedAccumulateMaskSIMD16(b *testing.B)      { benchAcc(b, fxIn16, "mask", true) }
+func BenchmarkFloatingAccumulateOpOver16(b *testing.B)     { benchAcc(b, flIn16, "over", false) }
+func BenchmarkFloatingAccumulateOpOverSIMD16(b *testing.B) { benchAcc(b, flIn16, "over", true) }
+func BenchmarkFloatingAccumulateOpSrc16(b *testing.B)      { benchAcc(b, flIn16, "src", false) }
+func BenchmarkFloatingAccumulateOpSrcSIMD16(b *testing.B)  { benchAcc(b, flIn16, "src", true) }
+func BenchmarkFloatingAccumulateMask16(b *testing.B)       { benchAcc(b, flIn16, "mask", false) }
+func BenchmarkFloatingAccumulateMaskSIMD16(b *testing.B)   { benchAcc(b, flIn16, "mask", true) }
+
+func BenchmarkFixedAccumulateOpOver64(b *testing.B)        { benchAcc(b, fxIn64, "over", false) }
+func BenchmarkFixedAccumulateOpOverSIMD64(b *testing.B)    { benchAcc(b, fxIn64, "over", true) }
+func BenchmarkFixedAccumulateOpSrc64(b *testing.B)         { benchAcc(b, fxIn64, "src", false) }
+func BenchmarkFixedAccumulateOpSrcSIMD64(b *testing.B)     { benchAcc(b, fxIn64, "src", true) }
+func BenchmarkFixedAccumulateMask64(b *testing.B)          { benchAcc(b, fxIn64, "mask", false) }
+func BenchmarkFixedAccumulateMaskSIMD64(b *testing.B)      { benchAcc(b, fxIn64, "mask", true) }
+func BenchmarkFloatingAccumulateOpOver64(b *testing.B)     { benchAcc(b, flIn64, "over", false) }
+func BenchmarkFloatingAccumulateOpOverSIMD64(b *testing.B) { benchAcc(b, flIn64, "over", true) }
+func BenchmarkFloatingAccumulateOpSrc64(b *testing.B)      { benchAcc(b, flIn64, "src", false) }
+func BenchmarkFloatingAccumulateOpSrcSIMD64(b *testing.B)  { benchAcc(b, flIn64, "src", true) }
+func BenchmarkFloatingAccumulateMask64(b *testing.B)       { benchAcc(b, flIn64, "mask", false) }
+func BenchmarkFloatingAccumulateMaskSIMD64(b *testing.B)   { benchAcc(b, flIn64, "mask", true) }
+
+func benchAcc(b *testing.B, in interface{}, op string, simd bool) {
+	var f func()
+
+	switch in := in.(type) {
+	case []uint32:
+		if simd && !haveFixedAccumulateSIMD {
+			b.Skip("No SIMD implemention")
+		}
+
+		switch op {
+		case "over":
+			dst := make([]uint8, len(in))
+			if simd {
+				f = func() { fixedAccumulateOpOverSIMD(dst, in) }
+			} else {
+				f = func() { fixedAccumulateOpOver(dst, in) }
+			}
+		case "src":
+			dst := make([]uint8, len(in))
+			if simd {
+				f = func() { fixedAccumulateOpSrcSIMD(dst, in) }
+			} else {
+				f = func() { fixedAccumulateOpSrc(dst, in) }
+			}
+		case "mask":
+			buf := make([]uint32, len(in))
+			copy(buf, in)
+			if simd {
+				f = func() { fixedAccumulateMaskSIMD(buf) }
+			} else {
+				f = func() { fixedAccumulateMask(buf) }
+			}
+		}
+
+	case []float32:
+		if simd && !haveFloatingAccumulateSIMD {
+			b.Skip("No SIMD implemention")
+		}
+
+		switch op {
+		case "over":
+			dst := make([]uint8, len(in))
+			if simd {
+				f = func() { floatingAccumulateOpOverSIMD(dst, in) }
+			} else {
+				f = func() { floatingAccumulateOpOver(dst, in) }
+			}
+		case "src":
+			dst := make([]uint8, len(in))
+			if simd {
+				f = func() { floatingAccumulateOpSrcSIMD(dst, in) }
+			} else {
+				f = func() { floatingAccumulateOpSrc(dst, in) }
+			}
+		case "mask":
+			dst := make([]uint32, len(in))
+			if simd {
+				f = func() { floatingAccumulateMaskSIMD(dst, in) }
+			} else {
+				f = func() { floatingAccumulateMask(dst, in) }
+			}
+		}
+	}
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		f()
+	}
+}
+
+// itou exists because "uint32(int2ϕ(-1))" doesn't compile: constant -1
+// overflows uint32.
+func itou(i int2ϕ) uint32 {
+	return uint32(i)
+}
+
+var fxInShort = []uint32{
+	itou(+0x08000), // +0.125, // Running sum: +0.125
+	itou(-0x20000), // -0.500, // Running sum: -0.375
+	itou(+0x10000), // +0.250, // Running sum: -0.125
+	itou(+0x18000), // +0.375, // Running sum: +0.250
+	itou(+0x08000), // +0.125, // Running sum: +0.375
+	itou(+0x00000), // +0.000, // Running sum: +0.375
+	itou(-0x40000), // -1.000, // Running sum: -0.625
+	itou(-0x20000), // -0.500, // Running sum: -1.125
+	itou(+0x10000), // +0.250, // Running sum: -0.875
+	itou(+0x38000), // +0.875, // Running sum: +0.000
+	itou(+0x10000), // +0.250, // Running sum: +0.250
+	itou(+0x30000), // +0.750, // Running sum: +1.000
+}
+
+var flInShort = []float32{
+	+0.125, // Running sum: +0.125
+	-0.500, // Running sum: -0.375
+	+0.250, // Running sum: -0.125
+	+0.375, // Running sum: +0.250
+	+0.125, // Running sum: +0.375
+	+0.000, // Running sum: +0.375
+	-1.000, // Running sum: -0.625
+	-0.500, // Running sum: -1.125
+	+0.250, // Running sum: -0.875
+	+0.875, // Running sum: +0.000
+	+0.250, // Running sum: +0.250
+	+0.750, // Running sum: +1.000
+}
+
+// It's OK for fxMaskShort and flMaskShort to have slightly different values.
+// Both the fixed and floating point implementations already have (different)
+// rounding errors in the xxxLineTo methods before we get to accumulation. It's
+// OK for 50% coverage (in ideal math) to be approximated by either 0x7fff or
+// 0x8000. Both slices do contain checks that 0% and 100% map to 0x0000 and
+// 0xffff, as does checkCornersCenter in vector_test.go.
+//
+// It is important, though, for the SIMD and non-SIMD fixed point
+// implementations to give the exact same output, and likewise for the floating
+// point implementations.
+
+var fxMaskShort = []uint32{
+	0x2000,
+	0x6000,
+	0x2000,
+	0x4000,
+	0x6000,
+	0x6000,
+	0xa000,
+	0xffff,
+	0xe000,
+	0x0000,
+	0x4000,
+	0xffff,
+}
+
+var flMaskShort = []uint32{
+	0x1fff,
+	0x5fff,
+	0x1fff,
+	0x3fff,
+	0x5fff,
+	0x5fff,
+	0x9fff,
+	0xffff,
+	0xdfff,
+	0x0000,
+	0x3fff,
+	0xffff,
+}
+
+func TestMakeFxInXxx(t *testing.T) {
+	dump := func(us []uint32) string {
+		var b bytes.Buffer
+		for i, u := range us {
+			if i%8 == 0 {
+				b.WriteByte('\n')
+			}
+			fmt.Fprintf(&b, "%#08x, ", u)
+		}
+		return b.String()
+	}
+
+	if !uint32sEqual(fxIn16, hardCodedFxIn16) {
+		t.Errorf("height 16: got:%v\nwant:%v", dump(fxIn16), dump(hardCodedFxIn16))
+	}
+}
+
+func TestMakeFlInXxx(t *testing.T) {
+	dump := func(fs []float32) string {
+		var b bytes.Buffer
+		for i, f := range fs {
+			if i%8 == 0 {
+				b.WriteByte('\n')
+			}
+			fmt.Fprintf(&b, "%v, ", f)
+		}
+		return b.String()
+	}
+
+	if !float32sEqual(flIn16, hardCodedFlIn16) {
+		t.Errorf("height 16: got:%v\nwant:%v", dump(flIn16), dump(hardCodedFlIn16))
+	}
+}
+
+func makeInXxx(height int, useFloatingPointMath bool) *Rasterizer {
+	width, data := scaledBenchmarkGlyphData(height)
+	z := NewRasterizer(width, height)
+	z.setUseFloatingPointMath(useFloatingPointMath)
+	for _, d := range data {
+		switch d.n {
+		case 0:
+			z.MoveTo(d.px, d.py)
+		case 1:
+			z.LineTo(d.px, d.py)
+		case 2:
+			z.QuadTo(d.px, d.py, d.qx, d.qy)
+		}
+	}
+	return z
+}
+
+func makeFxInXxx(height int) []uint32 {
+	z := makeInXxx(height, false)
+	return z.bufU32
+}
+
+func makeFlInXxx(height int) []float32 {
+	z := makeInXxx(height, true)
+	return z.bufF32
+}
+
+// fxInXxx and flInXxx are the z.bufU32 and z.bufF32 inputs to the accumulate
+// functions when rasterizing benchmarkGlyphData at a height of Xxx pixels.
+//
+// fxMaskXxx and flMaskXxx are the corresponding golden outputs of those
+// accumulateMask functions.
+//
+// The hardCodedEtc versions are a sanity check for unexpected changes in the
+// rasterization implementations up to but not including accumulation.
+
+var (
+	fxIn16 = makeFxInXxx(16)
+	fxIn64 = makeFxInXxx(64)
+	flIn16 = makeFlInXxx(16)
+	flIn64 = makeFlInXxx(64)
+)
+
+var hardCodedFxIn16 = []uint32{
+	0x00000000, 0x00000000, 0xffffe91d, 0xfffe7c4a, 0xfffeaa9f, 0xffff4e33, 0xffffc1c5, 0x00007782,
+	0x00009619, 0x0001a857, 0x000129e9, 0x00000028, 0x00000000, 0x00000000, 0xffff6e70, 0xfffd3199,
+	0xffff5ff8, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00014b29,
+	0x0002acf3, 0x000007e2, 0xffffca5a, 0xfffcab73, 0xffff8a34, 0x00001b55, 0x0001b334, 0x0001449e,
+	0x0000434d, 0xffff62ec, 0xfffe1443, 0xffff325d, 0x00000000, 0x0002234a, 0x0001dcb6, 0xfffe2948,
+	0xfffdd6b8, 0x00000000, 0x00028cc0, 0x00017340, 0x00000000, 0x00000000, 0x00000000, 0xffffd2d6,
+	0xfffcadd0, 0xffff7f5c, 0x00007400, 0x00038c00, 0xfffe9260, 0xffff2da0, 0x0000023a, 0x0002259b,
+	0x0000182a, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0xfffdc600, 0xfffe3a00, 0x00000059,
+	0x0003a44d, 0x00005b59, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
+	0x00000000, 0x00000000, 0xfffe33f3, 0xfffdcc0d, 0x00000000, 0x00033c02, 0x0000c3fe, 0x00000000,
+	0x00000000, 0xffffa13d, 0xfffeeec8, 0xffff8c02, 0xffff8c48, 0xffffc7b5, 0x00000000, 0xffff5b68,
+	0xffff3498, 0x00000000, 0x00033c00, 0x0000c400, 0xffff9bc4, 0xfffdf4a3, 0xfffe8df3, 0xffffe1a8,
+	0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00033c00,
+	0x000092c7, 0xfffcf373, 0xffff3dc7, 0x00000fcc, 0x00011ae7, 0x000130c3, 0x0000680d, 0x00004a59,
+	0x00000a20, 0xfffe9dc4, 0xfffe4a3c, 0x00000000, 0x00033c00, 0xfffe87ef, 0xfffe3c11, 0x0000105e,
+	0x0002b9c4, 0x000135dc, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0xfffe3600, 0xfffdca00,
+	0x00000000, 0x00033c00, 0xfffd9000, 0xffff3400, 0x0000e400, 0x00031c00, 0x00000000, 0x00000000,
+	0x00000000, 0x00000000, 0x00000000, 0xfffe3600, 0xfffdca00, 0x00000000, 0x00033c00, 0xfffcf9a5,
+	0xffffca5b, 0x000120e6, 0x0002df1a, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
+	0xfffdb195, 0xfffe4e6b, 0x00000000, 0x00033c00, 0xfffd9e00, 0xffff2600, 0x00002f0e, 0x00033ea3,
+	0x0000924d, 0x00000000, 0x00000000, 0x00000000, 0xfffe83b3, 0xfffd881d, 0xfffff431, 0x00000000,
+	0x00031f60, 0xffff297a, 0xfffdb726, 0x00000000, 0x000053a7, 0x0001b506, 0x0000a24b, 0xffffa32d,
+	0xfffead9b, 0xffff0479, 0xffffffc9, 0x00000000, 0x00000000, 0x0002d800, 0x0001249d, 0xfffd67bb,
+	0xfffe9baa, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x0000ac03, 0x0001448b,
+	0xfffe0f70, 0x00000000, 0x000229ea, 0x0001d616, 0xffffff8c, 0xfffebf76, 0xfffe54d9, 0xffff5d9e,
+	0xffffd3eb, 0x0000c65e, 0x0000fc15, 0x0001d491, 0xffffb566, 0xfffd9433, 0x00000000, 0x0000e4ec,
+}
+
+var hardCodedFlIn16 = []float32{
+	0, 0, -0.022306755, -0.3782405, -0.33334962, -0.1741521, -0.0607556, 0.11660573,
+	0.14664596, 0.41462868, 0.2907673, 0.0001568835, 0, 0, -0.14239307, -0.7012868,
+	-0.15632017, 0, 0, 0, 0, 0, 0, 0.3230303,
+	0.6690931, 0.007876594, -0.05189419, -0.832786, -0.11531975, 0.026225802, 0.42518616, 0.3154636,
+	0.06598757, -0.15304244, -0.47969276, -0.20012794, 0, 0.5327272, 0.46727282, -0.45950258,
+	-0.5404974, 0, 0.63484025, 0.36515975, 0, 0, 0, -0.04351709,
+	-0.8293345, -0.12714837, 0.11087036, 0.88912964, -0.35792422, -0.2053554, 0.0022513224, 0.5374398,
+	0.023588525, 0, 0, 0, 0, -0.55346966, -0.44653034, 0.0002531938,
+	0.9088273, 0.090919495, 0, 0, 0, 0, 0, 0,
+	0, 0, -0.44745448, -0.5525455, 0, 0.80748945, 0.19251058, 0,
+	0, -0.092476256, -0.2661464, -0.11322958, -0.11298219, -0.055094406, 0, -0.16045958,
+	-0.1996116, 0, 0.80748653, 0.19251347, -0.09804727, -0.51129663, -0.3610403, -0.029615778,
+	0, 0, 0, 0, 0, 0, 0, 0.80748653,
+	0.14411622, -0.76251525, -0.1890875, 0.01527351, 0.27528667, 0.29730347, 0.101477206, 0.07259522,
+	0.009900213, -0.34395567, -0.42788061, 0, 0.80748653, -0.3648737, -0.44261283, 0.015778137,
+	0.6826565, 0.30156538, 0, 0, 0, 0, -0.44563293, -0.55436707,
+	0, 0.80748653, -0.60703933, -0.20044717, 0.22371745, 0.77628255, 0, 0,
+	0, 0, 0, -0.44563293, -0.55436707, 0, 0.80748653, -0.7550391,
+	-0.05244744, 0.2797074, 0.72029257, 0, 0, 0, 0, 0,
+	-0.57440215, -0.42559785, 0, 0.80748653, -0.59273535, -0.21475118, 0.04544862, 0.81148535,
+	0.14306602, 0, 0, 0, -0.369642, -0.61841226, -0.011945802, 0,
+	0.7791623, -0.20691396, -0.57224834, 0, 0.08218567, 0.42637306, 0.1586175, -0.089709565,
+	-0.32935485, -0.24788953, -0.00022224105, 0, 0, 0.7085409, 0.28821066, -0.64765793,
+	-0.34909368, 0, 0, 0, 0, 0, 0.16679136, 0.31914657,
+	-0.48593786, 0, 0.537915, 0.462085, -0.00041967133, -0.3120329, -0.41914812, -0.15886839,
+	-0.042683028, 0.19370951, 0.24624406, 0.45803425, -0.07049577, -0.6091341, 0, 0.22253075,
+}
+
+var fxMask16 = []uint32{
+	0x0000, 0x0000, 0x05b8, 0x66a6, 0xbbfe, 0xe871, 0xf800, 0xda20, 0xb499, 0x4a84, 0x0009, 0x0000, 0x0000,
+	0x0000, 0x2463, 0xd7fd, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xad35, 0x01f8, 0x0000,
+	0x0d69, 0xe28c, 0xffff, 0xf92a, 0x8c5d, 0x3b36, 0x2a62, 0x51a7, 0xcc97, 0xffff, 0xffff, 0x772d, 0x0000,
+	0x75ad, 0xffff, 0xffff, 0x5ccf, 0x0000, 0x0000, 0x0000, 0x0000, 0x0b4a, 0xdfd6, 0xffff, 0xe2ff, 0x0000,
+	0x5b67, 0x8fff, 0x8f70, 0x060a, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x8e7f, 0xffff, 0xffe9, 0x16d6,
+	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x7303, 0xffff, 0xffff, 0x30ff,
+	0x0000, 0x0000, 0x0000, 0x17b0, 0x5bfe, 0x78fe, 0x95ec, 0xa3fe, 0xa3fe, 0xcd24, 0xfffe, 0xfffe, 0x30fe,
+	0x0001, 0x190d, 0x9be5, 0xf868, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0x30fe,
+	0x0c4c, 0xcf6f, 0xfffe, 0xfc0b, 0xb551, 0x6920, 0x4f1d, 0x3c87, 0x39ff, 0x928e, 0xffff, 0xffff, 0x30ff,
+	0x8f03, 0xffff, 0xfbe7, 0x4d76, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x727f, 0xffff, 0xffff, 0x30ff,
+	0xccff, 0xffff, 0xc6ff, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x727f, 0xffff, 0xffff, 0x30ff,
+	0xf296, 0xffff, 0xb7c6, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x939a, 0xffff, 0xffff, 0x30ff,
+	0xc97f, 0xffff, 0xf43c, 0x2493, 0x0000, 0x0000, 0x0000, 0x0000, 0x5f13, 0xfd0c, 0xffff, 0xffff, 0x3827,
+	0x6dc9, 0xffff, 0xffff, 0xeb16, 0x7dd4, 0x5541, 0x6c76, 0xc10f, 0xfff1, 0xffff, 0xffff, 0xffff, 0x49ff,
+	0x00d8, 0xa6e9, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xfffe, 0xd4fe, 0x83db, 0xffff, 0xffff, 0x7584,
+	0x0000, 0x001c, 0x503e, 0xbb08, 0xe3a1, 0xeea6, 0xbd0e, 0x7e09, 0x08e5, 0x1b8b, 0xb67f, 0xb67f, 0x7d44,
+}
+
+var flMask16 = []uint32{
+	0x0000, 0x0000, 0x05b5, 0x668a, 0xbbe0, 0xe875, 0xf803, 0xda29, 0xb49f, 0x4a7a, 0x000a, 0x0000, 0x0000,
+	0x0000, 0x2473, 0xd7fb, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xad4d, 0x0204, 0x0000,
+	0x0d48, 0xe27a, 0xffff, 0xf949, 0x8c70, 0x3bae, 0x2ac9, 0x51f7, 0xccc4, 0xffff, 0xffff, 0x779f, 0x0000,
+	0x75a1, 0xffff, 0xffff, 0x5d7b, 0x0000, 0x0000, 0x0000, 0x0000, 0x0b23, 0xdf73, 0xffff, 0xe39d, 0x0000,
+	0x5ba0, 0x9033, 0x8f9f, 0x0609, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x8db0, 0xffff, 0xffef, 0x1746,
+	0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x728c, 0xffff, 0xffff, 0x3148,
+	0x0000, 0x0000, 0x0000, 0x17ac, 0x5bce, 0x78cb, 0x95b7, 0xa3d2, 0xa3d2, 0xcce6, 0xffff, 0xffff, 0x3148,
+	0x0000, 0x1919, 0x9bfd, 0xf86b, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0x3148,
+	0x0c63, 0xcf97, 0xffff, 0xfc17, 0xb59d, 0x6981, 0x4f87, 0x3cf1, 0x3a68, 0x9276, 0xffff, 0xffff, 0x3148,
+	0x8eb0, 0xffff, 0xfbf5, 0x4d33, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x7214, 0xffff, 0xffff, 0x3148,
+	0xccaf, 0xffff, 0xc6ba, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x7214, 0xffff, 0xffff, 0x3148,
+	0xf292, 0xffff, 0xb865, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x930c, 0xffff, 0xffff, 0x3148,
+	0xc906, 0xffff, 0xf45d, 0x249f, 0x0000, 0x0000, 0x0000, 0x0000, 0x5ea0, 0xfcf1, 0xffff, 0xffff, 0x3888,
+	0x6d81, 0xffff, 0xffff, 0xeaf5, 0x7dcf, 0x5533, 0x6c2b, 0xc07b, 0xfff1, 0xffff, 0xffff, 0xffff, 0x4a9d,
+	0x00d4, 0xa6a1, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xd54d, 0x8399, 0xffff, 0xffff, 0x764b,
+	0x0000, 0x001b, 0x4ffc, 0xbb4a, 0xe3f5, 0xeee3, 0xbd4c, 0x7e42, 0x0900, 0x1b0c, 0xb6fc, 0xb6fc, 0x7e04,
+}
+
+// TestFixedFloatingCloseness compares the closeness of the fixed point and
+// floating point rasterizer.
+func TestFixedFloatingCloseness(t *testing.T) {
+	if len(fxMask16) != len(flMask16) {
+		t.Fatalf("len(fxMask16) != len(flMask16)")
+	}
+
+	total := uint32(0)
+	for i := range fxMask16 {
+		a := fxMask16[i]
+		b := flMask16[i]
+		if a > b {
+			total += a - b
+		} else {
+			total += b - a
+		}
+	}
+	n := len(fxMask16)
+
+	// This log message is useful when changing the fixed point rasterizer
+	// implementation, such as by changing ϕ. Assuming that the floating point
+	// rasterizer is accurate, the average difference is a measure of how
+	// inaccurate the (faster) fixed point rasterizer is.
+	//
+	// Smaller is better.
+	percent := float64(total*100) / float64(n*65535)
+	t.Logf("Comparing closeness of the fixed point and floating point rasterizer.\n"+
+		"Specifically, the elements of fxMask16 and flMask16.\n"+
+		"Total diff = %d, n = %d, avg = %.5f out of 65535, or %.5f%%.\n",
+		total, n, float64(total)/float64(n), percent)
+
+	const thresholdPercent = 1.0
+	if percent > thresholdPercent {
+		t.Errorf("average difference: got %.5f%%, want <= %.5f%%", percent, thresholdPercent)
+	}
+}
diff --git a/vendor/golang.org/x/image/vector/gen.go b/vendor/golang.org/x/image/vector/gen.go
new file mode 100644
index 000000000..28b298b5e
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/gen.go
@@ -0,0 +1,447 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build ignore
+
+package main
+
+import (
+	"bytes"
+	"io/ioutil"
+	"log"
+	"strings"
+	"text/template"
+)
+
+const (
+	copyright = "" +
+		"// Copyright 2016 The Go Authors. All rights reserved.\n" +
+		"// Use of this source code is governed by a BSD-style\n" +
+		"// license that can be found in the LICENSE file.\n"
+
+	doNotEdit = "// generated by go run gen.go; DO NOT EDIT\n"
+
+	dashDashDash = "// --------"
+)
+
+func main() {
+	tmpl, err := ioutil.ReadFile("gen_acc_amd64.s.tmpl")
+	if err != nil {
+		log.Fatalf("ReadFile: %v", err)
+	}
+	if !bytes.HasPrefix(tmpl, []byte(copyright)) {
+		log.Fatal("source template did not start with the copyright header")
+	}
+	tmpl = tmpl[len(copyright):]
+
+	preamble := []byte(nil)
+	if i := bytes.Index(tmpl, []byte(dashDashDash)); i < 0 {
+		log.Fatalf("source template did not contain %q", dashDashDash)
+	} else {
+		preamble, tmpl = tmpl[:i], tmpl[i:]
+	}
+
+	t, err := template.New("").Parse(string(tmpl))
+	if err != nil {
+		log.Fatalf("Parse: %v", err)
+	}
+
+	out := bytes.NewBuffer(nil)
+	out.WriteString(doNotEdit)
+	out.Write(preamble)
+
+	for i, v := range instances {
+		if i != 0 {
+			out.WriteString("\n")
+		}
+		if strings.Contains(v.LoadArgs, "{{.ShortName}}") {
+			v.LoadArgs = strings.Replace(v.LoadArgs, "{{.ShortName}}", v.ShortName, -1)
+		}
+		if err := t.Execute(out, v); err != nil {
+			log.Fatalf("Execute(%q): %v", v.ShortName, err)
+		}
+	}
+
+	if err := ioutil.WriteFile("acc_amd64.s", out.Bytes(), 0666); err != nil {
+		log.Fatalf("WriteFile: %v", err)
+	}
+}
+
+var instances = []struct {
+	LongName       string
+	ShortName      string
+	FrameSize      string
+	ArgsSize       string
+	Args           string
+	DstElemSize1   int
+	DstElemSize4   int
+	XMM3           string
+	XMM4           string
+	XMM5           string
+	XMM6           string
+	XMM8           string
+	XMM9           string
+	XMM10          string
+	LoadArgs       string
+	Setup          string
+	LoadXMMRegs    string
+	Add            string
+	ClampAndScale  string
+	ConvertToInt32 string
+	Store4         string
+	Store1         string
+}{{
+	LongName:       "fixedAccumulateOpOver",
+	ShortName:      "fxAccOpOver",
+	FrameSize:      fxFrameSize,
+	ArgsSize:       twoArgArgsSize,
+	Args:           "dst []uint8, src []uint32",
+	DstElemSize1:   1 * sizeOfUint8,
+	DstElemSize4:   4 * sizeOfUint8,
+	XMM3:           fxXMM3,
+	XMM4:           fxXMM4,
+	XMM5:           fxXMM5,
+	XMM6:           opOverXMM6,
+	XMM8:           opOverXMM8,
+	XMM9:           opOverXMM9,
+	XMM10:          opOverXMM10,
+	LoadArgs:       twoArgLoadArgs,
+	Setup:          fxSetup,
+	LoadXMMRegs:    fxLoadXMMRegs + "\n" + opOverLoadXMMRegs,
+	Add:            fxAdd,
+	ClampAndScale:  fxClampAndScale,
+	ConvertToInt32: fxConvertToInt32,
+	Store4:         opOverStore4,
+	Store1:         opOverStore1,
+}, {
+	LongName:       "fixedAccumulateOpSrc",
+	ShortName:      "fxAccOpSrc",
+	FrameSize:      fxFrameSize,
+	ArgsSize:       twoArgArgsSize,
+	Args:           "dst []uint8, src []uint32",
+	DstElemSize1:   1 * sizeOfUint8,
+	DstElemSize4:   4 * sizeOfUint8,
+	XMM3:           fxXMM3,
+	XMM4:           fxXMM4,
+	XMM5:           fxXMM5,
+	XMM6:           opSrcXMM6,
+	XMM8:           opSrcXMM8,
+	XMM9:           opSrcXMM9,
+	XMM10:          opSrcXMM10,
+	LoadArgs:       twoArgLoadArgs,
+	Setup:          fxSetup,
+	LoadXMMRegs:    fxLoadXMMRegs + "\n" + opSrcLoadXMMRegs,
+	Add:            fxAdd,
+	ClampAndScale:  fxClampAndScale,
+	ConvertToInt32: fxConvertToInt32,
+	Store4:         opSrcStore4,
+	Store1:         opSrcStore1,
+}, {
+	LongName:       "fixedAccumulateMask",
+	ShortName:      "fxAccMask",
+	FrameSize:      fxFrameSize,
+	ArgsSize:       oneArgArgsSize,
+	Args:           "buf []uint32",
+	DstElemSize1:   1 * sizeOfUint32,
+	DstElemSize4:   4 * sizeOfUint32,
+	XMM3:           fxXMM3,
+	XMM4:           fxXMM4,
+	XMM5:           fxXMM5,
+	XMM6:           maskXMM6,
+	XMM8:           maskXMM8,
+	XMM9:           maskXMM9,
+	XMM10:          maskXMM10,
+	LoadArgs:       oneArgLoadArgs,
+	Setup:          fxSetup,
+	LoadXMMRegs:    fxLoadXMMRegs + "\n" + maskLoadXMMRegs,
+	Add:            fxAdd,
+	ClampAndScale:  fxClampAndScale,
+	ConvertToInt32: fxConvertToInt32,
+	Store4:         maskStore4,
+	Store1:         maskStore1,
+}, {
+	LongName:       "floatingAccumulateOpOver",
+	ShortName:      "flAccOpOver",
+	FrameSize:      flFrameSize,
+	ArgsSize:       twoArgArgsSize,
+	Args:           "dst []uint8, src []float32",
+	DstElemSize1:   1 * sizeOfUint8,
+	DstElemSize4:   4 * sizeOfUint8,
+	XMM3:           flXMM3,
+	XMM4:           flXMM4,
+	XMM5:           flXMM5,
+	XMM6:           opOverXMM6,
+	XMM8:           opOverXMM8,
+	XMM9:           opOverXMM9,
+	XMM10:          opOverXMM10,
+	LoadArgs:       twoArgLoadArgs,
+	Setup:          flSetup,
+	LoadXMMRegs:    flLoadXMMRegs + "\n" + opOverLoadXMMRegs,
+	Add:            flAdd,
+	ClampAndScale:  flClampAndScale,
+	ConvertToInt32: flConvertToInt32,
+	Store4:         opOverStore4,
+	Store1:         opOverStore1,
+}, {
+	LongName:       "floatingAccumulateOpSrc",
+	ShortName:      "flAccOpSrc",
+	FrameSize:      flFrameSize,
+	ArgsSize:       twoArgArgsSize,
+	Args:           "dst []uint8, src []float32",
+	DstElemSize1:   1 * sizeOfUint8,
+	DstElemSize4:   4 * sizeOfUint8,
+	XMM3:           flXMM3,
+	XMM4:           flXMM4,
+	XMM5:           flXMM5,
+	XMM6:           opSrcXMM6,
+	XMM8:           opSrcXMM8,
+	XMM9:           opSrcXMM9,
+	XMM10:          opSrcXMM10,
+	LoadArgs:       twoArgLoadArgs,
+	Setup:          flSetup,
+	LoadXMMRegs:    flLoadXMMRegs + "\n" + opSrcLoadXMMRegs,
+	Add:            flAdd,
+	ClampAndScale:  flClampAndScale,
+	ConvertToInt32: flConvertToInt32,
+	Store4:         opSrcStore4,
+	Store1:         opSrcStore1,
+}, {
+	LongName:       "floatingAccumulateMask",
+	ShortName:      "flAccMask",
+	FrameSize:      flFrameSize,
+	ArgsSize:       twoArgArgsSize,
+	Args:           "dst []uint32, src []float32",
+	DstElemSize1:   1 * sizeOfUint32,
+	DstElemSize4:   4 * sizeOfUint32,
+	XMM3:           flXMM3,
+	XMM4:           flXMM4,
+	XMM5:           flXMM5,
+	XMM6:           maskXMM6,
+	XMM8:           maskXMM8,
+	XMM9:           maskXMM9,
+	XMM10:          maskXMM10,
+	LoadArgs:       twoArgLoadArgs,
+	Setup:          flSetup,
+	LoadXMMRegs:    flLoadXMMRegs + "\n" + maskLoadXMMRegs,
+	Add:            flAdd,
+	ClampAndScale:  flClampAndScale,
+	ConvertToInt32: flConvertToInt32,
+	Store4:         maskStore4,
+	Store1:         maskStore1,
+}}
+
+const (
+	fxFrameSize = `0`
+	flFrameSize = `8`
+
+	oneArgArgsSize = `24`
+	twoArgArgsSize = `48`
+
+	sizeOfUint8  = 1
+	sizeOfUint32 = 4
+
+	fxXMM3 = `-`
+	flXMM3 = `flSignMask`
+
+	fxXMM4 = `-`
+	flXMM4 = `flOne`
+
+	fxXMM5 = `fxAlmost65536`
+	flXMM5 = `flAlmost65536`
+
+	oneArgLoadArgs = `
+		MOVQ buf_base+0(FP), DI
+		MOVQ buf_len+8(FP), BX
+		MOVQ buf_base+0(FP), SI
+		MOVQ buf_len+8(FP), R10
+		`
+	twoArgLoadArgs = `
+		MOVQ dst_base+0(FP), DI
+		MOVQ dst_len+8(FP), BX
+		MOVQ src_base+24(FP), SI
+		MOVQ src_len+32(FP), R10
+		// Sanity check that len(dst) >= len(src).
+		CMPQ BX, R10
+		JLT  {{.ShortName}}End
+		`
+
+	fxSetup = ``
+	flSetup = `
+		// Prepare to set MXCSR bits 13 and 14, so that the CVTPS2PL below is
+		// "Round To Zero".
+		STMXCSR mxcsrOrig-8(SP)
+		MOVL    mxcsrOrig-8(SP), AX
+		ORL     $0x6000, AX
+		MOVL    AX, mxcsrNew-4(SP)
+		`
+
+	fxLoadXMMRegs = `
+		// fxAlmost65536 := XMM(0x0000ffff repeated four times) // Maximum of an uint16.
+		MOVOU fxAlmost65536<>(SB), X5
+		`
+	flLoadXMMRegs = `
+		// flSignMask    := XMM(0x7fffffff repeated four times) // All but the sign bit of a float32.
+		// flOne         := XMM(0x3f800000 repeated four times) // 1 as a float32.
+		// flAlmost65536 := XMM(0x477fffff repeated four times) // 255.99998 * 256 as a float32.
+		MOVOU flSignMask<>(SB), X3
+		MOVOU flOne<>(SB), X4
+		MOVOU flAlmost65536<>(SB), X5
+		`
+
+	fxAdd = `PADDD`
+	flAdd = `ADDPS`
+
+	fxClampAndScale = `
+		// y = abs(x)
+		// y >>= 2 // Shift by 2*ϕ - 16.
+		// y = min(y, fxAlmost65536)
+		//
+		// pabsd  %xmm1,%xmm2
+		// psrld  $0x2,%xmm2
+		// pminud %xmm5,%xmm2
+		//
+		// Hopefully we'll get these opcode mnemonics into the assembler for Go
+		// 1.8. https://golang.org/issue/16007 isn't exactly the same thing, but
+		// it's similar.
+		BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x1e; BYTE $0xd1
+		BYTE $0x66; BYTE $0x0f; BYTE $0x72; BYTE $0xd2; BYTE $0x02
+		BYTE $0x66; BYTE $0x0f; BYTE $0x38; BYTE $0x3b; BYTE $0xd5
+		`
+	flClampAndScale = `
+		// y = x & flSignMask
+		// y = min(y, flOne)
+		// y = mul(y, flAlmost65536)
+		MOVOU X3, X2
+		ANDPS X1, X2
+		MINPS X4, X2
+		MULPS X5, X2
+		`
+
+	fxConvertToInt32 = `
+		// z = convertToInt32(y)
+		// No-op.
+		`
+	flConvertToInt32 = `
+		// z = convertToInt32(y)
+		LDMXCSR  mxcsrNew-4(SP)
+		CVTPS2PL X2, X2
+		LDMXCSR  mxcsrOrig-8(SP)
+		`
+
+	opOverStore4 = `
+		// Blend over the dst's prior value. SIMD for i in 0..3:
+		//
+		// dstA := uint32(dst[i]) * 0x101
+		// maskA := z@i
+		// outA := dstA*(0xffff-maskA)/0xffff + maskA
+		// dst[i] = uint8(outA >> 8)
+		//
+		// First, set X0 to dstA*(0xfff-maskA).
+		MOVL   (DI), X0
+		PSHUFB X8, X0
+		MOVOU  X9, X11
+		PSUBL  X2, X11
+		PMULLD X11, X0
+		// We implement uint32 division by 0xffff as multiplication by a magic
+		// constant (0x800080001) and then a shift by a magic constant (47).
+		// See TestDivideByFFFF for a justification.
+		//
+		// That multiplication widens from uint32 to uint64, so we have to
+		// duplicate and shift our four uint32s from one XMM register (X0) to
+		// two XMM registers (X0 and X11).
+		//
+		// Move the second and fourth uint32s in X0 to be the first and third
+		// uint32s in X11.
+		MOVOU X0, X11
+		PSRLQ $32, X11
+		// Multiply by magic, shift by magic.
+		//
+		// pmuludq %xmm10,%xmm0
+		// pmuludq %xmm10,%xmm11
+		BYTE  $0x66; BYTE $0x41; BYTE $0x0f; BYTE $0xf4; BYTE $0xc2
+		BYTE  $0x66; BYTE $0x45; BYTE $0x0f; BYTE $0xf4; BYTE $0xda
+		PSRLQ $47, X0
+		PSRLQ $47, X11
+		// Merge the two registers back to one, X11, and add maskA.
+		PSLLQ $32, X11
+		XORPS X0, X11
+		PADDD X11, X2
+		// As per opSrcStore4, shuffle and copy the 4 second-lowest bytes.
+		PSHUFB X6, X2
+		MOVL   X2, (DI)
+		`
+	opSrcStore4 = `
+		// z = shuffleTheSecondLowestBytesOfEach4ByteElement(z)
+		// copy(dst[:4], low4BytesOf(z))
+		PSHUFB X6, X2
+		MOVL   X2, (DI)
+		`
+	maskStore4 = `
+		// copy(dst[:4], z)
+		MOVOU X2, (DI)
+		`
+
+	opOverStore1 = `
+		// Blend over the dst's prior value.
+		//
+		// dstA := uint32(dst[0]) * 0x101
+		// maskA := z
+		// outA := dstA*(0xffff-maskA)/0xffff + maskA
+		// dst[0] = uint8(outA >> 8)
+		MOVBLZX (DI), R12
+		IMULL   $0x101, R12
+		MOVL    X2, R13
+		MOVL    $0xffff, AX
+		SUBL    R13, AX
+		MULL    R12             // MULL's implicit arg is AX, and the result is stored in DX:AX.
+		MOVL    $0x80008001, BX // Divide by 0xffff is to first multiply by a magic constant...
+		MULL    BX              // MULL's implicit arg is AX, and the result is stored in DX:AX.
+		SHRL    $15, DX         // ...and then shift by another magic constant (47 - 32 = 15).
+		ADDL    DX, R13
+		SHRL    $8, R13
+		MOVB    R13, (DI)
+		`
+	opSrcStore1 = `
+		// dst[0] = uint8(z>>8)
+		MOVL X2, BX
+		SHRL $8, BX
+		MOVB BX, (DI)
+		`
+	maskStore1 = `
+		// dst[0] = uint32(z)
+		MOVL X2, (DI)
+		`
+
+	opOverXMM6 = `gather`
+	opSrcXMM6  = `gather`
+	maskXMM6   = `-`
+
+	opOverXMM8 = `scatterAndMulBy0x101`
+	opSrcXMM8  = `-`
+	maskXMM8   = `-`
+
+	opOverXMM9 = `fxAlmost65536`
+	opSrcXMM9  = `-`
+	maskXMM9   = `-`
+
+	opOverXMM10 = `inverseFFFF`
+	opSrcXMM10  = `-`
+	maskXMM10   = `-`
+
+	opOverLoadXMMRegs = `
+		// gather               := XMM(see above)                      // PSHUFB shuffle mask.
+		// scatterAndMulBy0x101 := XMM(see above)                      // PSHUFB shuffle mask.
+		// fxAlmost65536        := XMM(0x0000ffff repeated four times) // 0xffff.
+		// inverseFFFF          := XMM(0x80008001 repeated four times) // Magic constant for dividing by 0xffff.
+		MOVOU gather<>(SB), X6
+		MOVOU scatterAndMulBy0x101<>(SB), X8
+		MOVOU fxAlmost65536<>(SB), X9
+		MOVOU inverseFFFF<>(SB), X10
+		`
+	opSrcLoadXMMRegs = `
+		// gather := XMM(see above) // PSHUFB shuffle mask.
+		MOVOU gather<>(SB), X6
+		`
+	maskLoadXMMRegs = ``
+)
diff --git a/vendor/golang.org/x/image/vector/gen_acc_amd64.s.tmpl b/vendor/golang.org/x/image/vector/gen_acc_amd64.s.tmpl
new file mode 100644
index 000000000..66b21a13d
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/gen_acc_amd64.s.tmpl
@@ -0,0 +1,171 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build !appengine
+// +build gc
+// +build go1.6
+// +build !noasm
+
+#include "textflag.h"
+
+// fl is short for floating point math. fx is short for fixed point math.
+
+DATA flAlmost65536<>+0x00(SB)/8, $0x477fffff477fffff
+DATA flAlmost65536<>+0x08(SB)/8, $0x477fffff477fffff
+DATA flOne<>+0x00(SB)/8, $0x3f8000003f800000
+DATA flOne<>+0x08(SB)/8, $0x3f8000003f800000
+DATA flSignMask<>+0x00(SB)/8, $0x7fffffff7fffffff
+DATA flSignMask<>+0x08(SB)/8, $0x7fffffff7fffffff
+
+// scatterAndMulBy0x101 is a PSHUFB mask that brings the low four bytes of an
+// XMM register to the low byte of that register's four uint32 values. It
+// duplicates those bytes, effectively multiplying each uint32 by 0x101.
+//
+// It transforms a little-endian 16-byte XMM value from
+//	ijkl????????????
+// to
+//	ii00jj00kk00ll00
+DATA scatterAndMulBy0x101<>+0x00(SB)/8, $0x8080010180800000
+DATA scatterAndMulBy0x101<>+0x08(SB)/8, $0x8080030380800202
+
+// gather is a PSHUFB mask that brings the second-lowest byte of the XMM
+// register's four uint32 values to the low four bytes of that register.
+//
+// It transforms a little-endian 16-byte XMM value from
+//	?i???j???k???l??
+// to
+//	ijkl000000000000
+DATA gather<>+0x00(SB)/8, $0x808080800d090501
+DATA gather<>+0x08(SB)/8, $0x8080808080808080
+
+DATA fxAlmost65536<>+0x00(SB)/8, $0x0000ffff0000ffff
+DATA fxAlmost65536<>+0x08(SB)/8, $0x0000ffff0000ffff
+DATA inverseFFFF<>+0x00(SB)/8, $0x8000800180008001
+DATA inverseFFFF<>+0x08(SB)/8, $0x8000800180008001
+
+GLOBL flAlmost65536<>(SB), (NOPTR+RODATA), $16
+GLOBL flOne<>(SB), (NOPTR+RODATA), $16
+GLOBL flSignMask<>(SB), (NOPTR+RODATA), $16
+GLOBL scatterAndMulBy0x101<>(SB), (NOPTR+RODATA), $16
+GLOBL gather<>(SB), (NOPTR+RODATA), $16
+GLOBL fxAlmost65536<>(SB), (NOPTR+RODATA), $16
+GLOBL inverseFFFF<>(SB), (NOPTR+RODATA), $16
+
+// func haveSSE4_1() bool
+TEXT ·haveSSE4_1(SB), NOSPLIT, $0
+	MOVQ $1, AX
+	CPUID
+	SHRQ $19, CX
+	ANDQ $1, CX
+	MOVB CX, ret+0(FP)
+	RET
+
+// ----------------------------------------------------------------------------
+
+// func {{.LongName}}SIMD({{.Args}})
+//
+// XMM registers. Variable names are per
+// https://github.com/google/font-rs/blob/master/src/accumulate.c
+//
+//	xmm0	scratch
+//	xmm1	x
+//	xmm2	y, z
+//	xmm3	{{.XMM3}}
+//	xmm4	{{.XMM4}}
+//	xmm5	{{.XMM5}}
+//	xmm6	{{.XMM6}}
+//	xmm7	offset
+//	xmm8	{{.XMM8}}
+//	xmm9	{{.XMM9}}
+//	xmm10	{{.XMM10}}
+TEXT ·{{.LongName}}SIMD(SB), NOSPLIT, ${{.FrameSize}}-{{.ArgsSize}}
+	{{.LoadArgs}}
+
+	// R10 = len(src) &^ 3
+	// R11 = len(src)
+	MOVQ R10, R11
+	ANDQ $-4, R10
+
+	{{.Setup}}
+
+	{{.LoadXMMRegs}}
+
+	// offset := XMM(0x00000000 repeated four times) // Cumulative sum.
+	XORPS X7, X7
+
+	// i := 0
+	MOVQ $0, R9
+
+{{.ShortName}}Loop4:
+	// for i < (len(src) &^ 3)
+	CMPQ R9, R10
+	JAE  {{.ShortName}}Loop1
+
+	// x = XMM(s0, s1, s2, s3)
+	//
+	// Where s0 is src[i+0], s1 is src[i+1], etc.
+	MOVOU (SI), X1
+
+	// scratch = XMM(0, s0, s1, s2)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s1+s2, s2+s3)
+	MOVOU    X1, X0
+	PSLLO    $4, X0
+	{{.Add}} X0, X1
+
+	// scratch = XMM(0, 0, 0, 0)
+	// scratch = XMM(scratch@0, scratch@0, x@0, x@1) // yields scratch == XMM(0, 0, s0, s0+s1)
+	// x += scratch                                  // yields x == XMM(s0, s0+s1, s0+s1+s2, s0+s1+s2+s3)
+	XORPS    X0, X0
+	SHUFPS   $0x40, X1, X0
+	{{.Add}} X0, X1
+
+	// x += offset
+	{{.Add}} X7, X1
+
+	{{.ClampAndScale}}
+
+	{{.ConvertToInt32}}
+
+	{{.Store4}}
+
+	// offset = XMM(x@3, x@3, x@3, x@3)
+	MOVOU  X1, X7
+	SHUFPS $0xff, X1, X7
+
+	// i += 4
+	// dst = dst[4:]
+	// src = src[4:]
+	ADDQ $4, R9
+	ADDQ ${{.DstElemSize4}}, DI
+	ADDQ $16, SI
+	JMP  {{.ShortName}}Loop4
+
+{{.ShortName}}Loop1:
+	// for i < len(src)
+	CMPQ R9, R11
+	JAE  {{.ShortName}}End
+
+	// x = src[i] + offset
+	MOVL     (SI), X1
+	{{.Add}} X7, X1
+
+	{{.ClampAndScale}}
+
+	{{.ConvertToInt32}}
+
+	{{.Store1}}
+
+	// offset = x
+	MOVOU X1, X7
+
+	// i += 1
+	// dst = dst[1:]
+	// src = src[1:]
+	ADDQ $1, R9
+	ADDQ ${{.DstElemSize1}}, DI
+	ADDQ $4, SI
+	JMP  {{.ShortName}}Loop1
+
+{{.ShortName}}End:
+	RET
diff --git a/vendor/golang.org/x/image/vector/raster_fixed.go b/vendor/golang.org/x/image/vector/raster_fixed.go
new file mode 100644
index 000000000..5b0fe7a7e
--- /dev/null
+++ b/vendor/golang.org/x/image/vector/raster_fixed.go
@@ -0,0 +1,327 @@
+// Copyright 2016 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package vector
+
+// This file contains a fixed point math implementation of the vector
+// graphics rasterizer.
+
+const (
+	// ϕ is the number of binary digits after the fixed point.
+	//
+	// For example, if ϕ == 10 (and int1ϕ is based on the int32 type) then we
+	// are using 22.10 fixed point math.
+	//
+	// When changing this number, also change the assembly code (search for ϕ
+	// in the .s files).
+	ϕ = 9
+
+	fxOne          int1ϕ = 1 << ϕ
+	fxOneAndAHalf  int1ϕ = 1<<ϕ + 1<<(ϕ-1)
+	fxOneMinusIota int1ϕ = 1<<ϕ - 1 // Used for rounding up.
+)
+
+// int1ϕ is a signed fixed-point number with 1*ϕ binary digits after the fixed
+// point.
+type int1ϕ int32
+
+// int2ϕ is a signed fixed-point number with 2*ϕ binary digits after the fixed
+// point.
+//
+// The Rasterizer's bufU32 field, nominally of type []uint32 (since that slice
+// is also used by other code), can be thought of as a []int2ϕ during the
+// fixedLineTo method. Lines of code that are actually like:
+//	buf[i] += uint32(etc) // buf has type []uint32.
+// can be thought of as
+//	buf[i] += int2ϕ(etc)  // buf has type []int2ϕ.
+type int2ϕ int32
+
+func fixedMax(x, y int1ϕ) int1ϕ {
+	if x > y {
+		return x
+	}
+	return y
+}
+
+func fixedMin(x, y int1ϕ) int1ϕ {
+	if x < y {
+		return x
+	}
+	return y
+}
+
+func fixedFloor(x int1ϕ) int32 { return int32(x >> ϕ) }
+func fixedCeil(x int1ϕ) int32  { return int32((x + fxOneMinusIota) >> ϕ) }
+
+func (z *Rasterizer) fixedLineTo(bx, by float32) {
+	ax, ay := z.penX, z.penY
+	z.penX, z.penY = bx, by
+	dir := int1ϕ(1)
+	if ay > by {
+		dir, ax, ay, bx, by = -1, bx, by, ax, ay
+	}
+	// Horizontal line segments yield no change in coverage. Almost horizontal
+	// segments would yield some change, in ideal math, but the computation
+	// further below, involving 1 / (by - ay), is unstable in fixed point math,
+	// so we treat the segment as if it was perfectly horizontal.
+	if by-ay <= 0.000001 {
+		return
+	}
+	dxdy := (bx - ax) / (by - ay)
+
+	ayϕ := int1ϕ(ay * float32(fxOne))
+	byϕ := int1ϕ(by * float32(fxOne))
+
+	x := int1ϕ(ax * float32(fxOne))
+	y := fixedFloor(ayϕ)
+	yMax := fixedCeil(byϕ)
+	if yMax > int32(z.size.Y) {
+		yMax = int32(z.size.Y)
+	}
+	width := int32(z.size.X)
+
+	for ; y < yMax; y++ {
+		dy := fixedMin(int1ϕ(y+1)<<ϕ, byϕ) - fixedMax(int1ϕ(y)<<ϕ, ayϕ)
+		xNext := x + int1ϕ(float32(dy)*dxdy)
+		if y < 0 {
+			x = xNext
+			continue
+		}
+		buf := z.bufU32[y*width:]
+		d := dy * dir // d ranges up to ±1<<(1*ϕ).
+		x0, x1 := x, xNext
+		if x > xNext {
+			x0, x1 = x1, x0
+		}
+		x0i := fixedFloor(x0)
+		x0Floor := int1ϕ(x0i) << ϕ
+		x1i := fixedCeil(x1)
+		x1Ceil := int1ϕ(x1i) << ϕ
+
+		if x1i <= x0i+1 {
+			xmf := (x+xNext)>>1 - x0Floor
+			if i := clamp(x0i+0, width); i < uint(len(buf)) {
+				buf[i] += uint32(d * (fxOne - xmf))
+			}
+			if i := clamp(x0i+1, width); i < uint(len(buf)) {
+				buf[i] += uint32(d * xmf)
+			}
+		} else {
+			oneOverS := x1 - x0
+			twoOverS := 2 * oneOverS
+			x0f := x0 - x0Floor
+			oneMinusX0f := fxOne - x0f
+			oneMinusX0fSquared := oneMinusX0f * oneMinusX0f
+			x1f := x1 - x1Ceil + fxOne
+			x1fSquared := x1f * x1f
+
+			// These next two variables are unused, as rounding errors are
+			// minimized when we delay the division by oneOverS for as long as
+			// possible. These lines of code (and the "In ideal math" comments
+			// below) are commented out instead of deleted in order to aid the
+			// comparison with the floating point version of the rasterizer.
+			//
+			// a0 := ((oneMinusX0f * oneMinusX0f) >> 1) / oneOverS
+			// am := ((x1f * x1f) >> 1) / oneOverS
+
+			if i := clamp(x0i, width); i < uint(len(buf)) {
+				// In ideal math: buf[i] += uint32(d * a0)
+				D := oneMinusX0fSquared // D ranges up to ±1<<(2*ϕ).
+				D *= d                  // D ranges up to ±1<<(3*ϕ).
+				D /= twoOverS
+				buf[i] += uint32(D)
+			}
+
+			if x1i == x0i+2 {
+				if i := clamp(x0i+1, width); i < uint(len(buf)) {
+					// In ideal math: buf[i] += uint32(d * (fxOne - a0 - am))
+					//
+					// (x1i == x0i+2) and (twoOverS == 2 * (x1 - x0)) implies
+					// that twoOverS ranges up to +1<<(1*ϕ+2).
+					D := twoOverS<<ϕ - oneMinusX0fSquared - x1fSquared // D ranges up to ±1<<(2*ϕ+2).
+					D *= d                                             // D ranges up to ±1<<(3*ϕ+2).
+					D /= twoOverS
+					buf[i] += uint32(D)
+				}
+			} else {
+				// This is commented out for the same reason as a0 and am.
+				//
+				// a1 := ((fxOneAndAHalf - x0f) << ϕ) / oneOverS
+
+				if i := clamp(x0i+1, width); i < uint(len(buf)) {
+					// In ideal math:
+					//	buf[i] += uint32(d * (a1 - a0))
+					// or equivalently (but better in non-ideal, integer math,
+					// with respect to rounding errors),
+					//	buf[i] += uint32(A * d / twoOverS)
+					// where
+					//	A = (a1 - a0) * twoOverS
+					//	  = a1*twoOverS - a0*twoOverS
+					// Noting that twoOverS/oneOverS equals 2, substituting for
+					// a0 and then a1, given above, yields:
+					//	A = a1*twoOverS - oneMinusX0fSquared
+					//	  = (fxOneAndAHalf-x0f)<<(ϕ+1) - oneMinusX0fSquared
+					//	  = fxOneAndAHalf<<(ϕ+1) - x0f<<(ϕ+1) - oneMinusX0fSquared
+					//
+					// This is a positive number minus two non-negative
+					// numbers. For an upper bound on A, the positive number is
+					//	P = fxOneAndAHalf<<(ϕ+1)
+					//	  < (2*fxOne)<<(ϕ+1)
+					//	  = fxOne<<(ϕ+2)
+					//	  = 1<<(2*ϕ+2)
+					//
+					// For a lower bound on A, the two non-negative numbers are
+					//	N = x0f<<(ϕ+1) + oneMinusX0fSquared
+					//	  ≤ x0f<<(ϕ+1) + fxOne*fxOne
+					//	  = x0f<<(ϕ+1) + 1<<(2*ϕ)
+					//	  < x0f<<(ϕ+1) + 1<<(2*ϕ+1)
+					//	  ≤ fxOne<<(ϕ+1) + 1<<(2*ϕ+1)
+					//	  = 1<<(2*ϕ+1) + 1<<(2*ϕ+1)
+					//	  = 1<<(2*ϕ+2)
+					//
+					// Thus, A ranges up to ±1<<(2*ϕ+2). It is possible to
+					// derive a tighter bound, but this bound is sufficient to
+					// reason about overflow.
+					D := (fxOneAndAHalf-x0f)<<(ϕ+1) - oneMinusX0fSquared // D ranges up to ±1<<(2*ϕ+2).
+					D *= d                                               // D ranges up to ±1<<(3*ϕ+2).
+					D /= twoOverS
+					buf[i] += uint32(D)
+				}
+				dTimesS := uint32((d << (2 * ϕ)) / oneOverS)
+				for xi := x0i + 2; xi < x1i-1; xi++ {
+					if i := clamp(xi, width); i < uint(len(buf)) {
+						buf[i] += dTimesS
+					}
+				}
+
+				// This is commented out for the same reason as a0 and am.
+				//
+				// a2 := a1 + (int1ϕ(x1i-x0i-3)<<(2*ϕ))/oneOverS
+
+				if i := clamp(x1i-1, width); i < uint(len(buf)) {
+					// In ideal math:
+					//	buf[i] += uint32(d * (fxOne - a2 - am))
+					// or equivalently (but better in non-ideal, integer math,
+					// with respect to rounding errors),
+					//	buf[i] += uint32(A * d / twoOverS)
+					// where
+					//	A = (fxOne - a2 - am) * twoOverS
+					//	  = twoOverS<<ϕ - a2*twoOverS - am*twoOverS
+					// Noting that twoOverS/oneOverS equals 2, substituting for
+					// am and then a2, given above, yields:
+					//	A = twoOverS<<ϕ - a2*twoOverS - x1f*x1f
+					//	  = twoOverS<<ϕ - a1*twoOverS - (int1ϕ(x1i-x0i-3)<<(2*ϕ))*2 - x1f*x1f
+					//	  = twoOverS<<ϕ - a1*twoOverS - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
+					// Substituting for a1, given above, yields:
+					//	A = twoOverS<<ϕ - ((fxOneAndAHalf-x0f)<<ϕ)*2 - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
+					//	  = twoOverS<<ϕ - (fxOneAndAHalf-x0f)<<(ϕ+1) - int1ϕ(x1i-x0i-3)<<(2*ϕ+1) - x1f*x1f
+					//	  = B<<ϕ - x1f*x1f
+					// where
+					//	B = twoOverS - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
+					//	  = (x1-x0)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
+					//
+					// Re-arranging the defintions given above:
+					//	x0Floor := int1ϕ(x0i) << ϕ
+					//	x0f := x0 - x0Floor
+					//	x1Ceil := int1ϕ(x1i) << ϕ
+					//	x1f := x1 - x1Ceil + fxOne
+					// combined with fxOne = 1<<ϕ yields:
+					//	x0 = x0f + int1ϕ(x0i)<<ϕ
+					//	x1 = x1f + int1ϕ(x1i-1)<<ϕ
+					// so that expanding (x1-x0) yields:
+					//	B = (x1f-x0f + int1ϕ(x1i-x0i-1)<<ϕ)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
+					//	  = (x1f-x0f)<<1 + int1ϕ(x1i-x0i-1)<<(ϕ+1) - (fxOneAndAHalf-x0f)<<1 - int1ϕ(x1i-x0i-3)<<(ϕ+1)
+					// A large part of the second and fourth terms cancel:
+					//	B = (x1f-x0f)<<1 - (fxOneAndAHalf-x0f)<<1 - int1ϕ(-2)<<(ϕ+1)
+					//	  = (x1f-x0f)<<1 - (fxOneAndAHalf-x0f)<<1 + 1<<(ϕ+2)
+					//	  = (x1f - fxOneAndAHalf)<<1 + 1<<(ϕ+2)
+					// The first term, (x1f - fxOneAndAHalf)<<1, is a negative
+					// number, bounded below by -fxOneAndAHalf<<1, which is
+					// greater than -fxOne<<2, or -1<<(ϕ+2). Thus, B ranges up
+					// to ±1<<(ϕ+2). One final simplification:
+					//	B = x1f<<1 + (1<<(ϕ+2) - fxOneAndAHalf<<1)
+					const C = 1<<(ϕ+2) - fxOneAndAHalf<<1
+					D := x1f<<1 + C // D ranges up to ±1<<(1*ϕ+2).
+					D <<= ϕ         // D ranges up to ±1<<(2*ϕ+2).
+					D -= x1fSquared // D ranges up to ±1<<(2*ϕ+3).
+					D *= d          // D ranges up to ±1<<(3*ϕ+3).
+					D /= twoOverS
+					buf[i] += uint32(D)
+				}
+			}
+
+			if i := clamp(x1i, width); i < uint(len(buf)) {
+				// In ideal math: buf[i] += uint32(d * am)
+				D := x1fSquared // D ranges up to ±1<<(2*ϕ).
+				D *= d          // D ranges up to ±1<<(3*ϕ).
+				D /= twoOverS
+				buf[i] += uint32(D)
+			}
+		}
+
+		x = xNext
+	}
+}
+
+func fixedAccumulateOpOver(dst []uint8, src []uint32) {
+	// Sanity check that len(dst) >= len(src).
+	if len(dst) < len(src) {
+		return
+	}
+
+	acc := int2ϕ(0)
+	for i, v := range src {
+		acc += int2ϕ(v)
+		a := acc
+		if a < 0 {
+			a = -a
+		}
+		a >>= 2*ϕ - 16
+		if a > 0xffff {
+			a = 0xffff
+		}
+		// This algorithm comes from the standard library's image/draw package.
+		dstA := uint32(dst[i]) * 0x101
+		maskA := uint32(a)
+		outA := dstA*(0xffff-maskA)/0xffff + maskA
+		dst[i] = uint8(outA >> 8)
+	}
+}
+
+func fixedAccumulateOpSrc(dst []uint8, src []uint32) {
+	// Sanity check that len(dst) >= len(src).
+	if len(dst) < len(src) {
+		return
+	}
+
+	acc := int2ϕ(0)
+	for i, v := range src {
+		acc += int2ϕ(v)
+		a := acc
+		if a < 0 {
+			a = -a
+		}
+		a >>= 2*ϕ - 8
+		if a > 0xff {
+			a = 0xff
+		}
+		dst[i] = uint8(a)
+	}
+}
+
+func fixedAccumulateMask(buf []uint32) {
+	acc := int2ϕ(0)
+	for i, v := range buf {
+		acc += int2ϕ(v)
+		a := acc
+		if a < 0 {
+			a = -a
+		}
+		a >>= 2*ϕ - 16
+		if a > 0xffff {
+			a = 0xffff
+		}
+		buf[i] = uint32(a)
+	}
+}
diff --git a/vendor/golang.org/x/image/vector/raster_floating.go b/vendor/golang.org/x/image/vector/raster_floating.go
index d03936a1e..143b376b5 100644
--- a/vendor/golang.org/x/image/vector/raster_floating.go
+++ b/vendor/golang.org/x/image/vector/raster_floating.go
@@ -9,8 +9,6 @@ package vector
 
 import (
 	"math"
-
-	"golang.org/x/image/math/f32"
 )
 
 func floatingMax(x, y float32) float32 {
@@ -30,38 +28,38 @@ func floatingMin(x, y float32) float32 {
 func floatingFloor(x float32) int32 { return int32(math.Floor(float64(x))) }
 func floatingCeil(x float32) int32  { return int32(math.Ceil(float64(x))) }
 
-func (z *Rasterizer) floatingLineTo(b f32.Vec2) {
-	a := z.pen
-	z.pen = b
+func (z *Rasterizer) floatingLineTo(bx, by float32) {
+	ax, ay := z.penX, z.penY
+	z.penX, z.penY = bx, by
 	dir := float32(1)
-	if a[1] > b[1] {
-		dir, a, b = -1, b, a
+	if ay > by {
+		dir, ax, ay, bx, by = -1, bx, by, ax, ay
 	}
 	// Horizontal line segments yield no change in coverage. Almost horizontal
 	// segments would yield some change, in ideal math, but the computation
-	// further below, involving 1 / (b[1] - a[1]), is unstable in floating
-	// point math, so we treat the segment as if it was perfectly horizontal.
-	if b[1]-a[1] <= 0.000001 {
+	// further below, involving 1 / (by - ay), is unstable in floating point
+	// math, so we treat the segment as if it was perfectly horizontal.
+	if by-ay <= 0.000001 {
 		return
 	}
-	dxdy := (b[0] - a[0]) / (b[1] - a[1])
+	dxdy := (bx - ax) / (by - ay)
 
-	x := a[0]
-	y := floatingFloor(a[1])
-	yMax := floatingCeil(b[1])
+	x := ax
+	y := floatingFloor(ay)
+	yMax := floatingCeil(by)
 	if yMax > int32(z.size.Y) {
 		yMax = int32(z.size.Y)
 	}
 	width := int32(z.size.X)
 
 	for ; y < yMax; y++ {
-		dy := floatingMin(float32(y+1), b[1]) - floatingMax(float32(y), a[1])
+		dy := floatingMin(float32(y+1), by) - floatingMax(float32(y), ay)
 		xNext := x + dy*dxdy
 		if y < 0 {
 			x = xNext
 			continue
 		}
-		buf := z.area[y*width:]
+		buf := z.bufF32[y*width:]
 		d := dy * dir
 		x0, x1 := x, xNext
 		if x > xNext {
@@ -122,7 +120,7 @@ func (z *Rasterizer) floatingLineTo(b f32.Vec2) {
 	}
 }
 
-func floatingAccumulate(dst []uint8, src []float32) {
+const (
 	// almost256 scales a floating point value in the range [0, 1] to a uint8
 	// value in the range [0x00, 0xff].
 	//
@@ -136,7 +134,44 @@ func floatingAccumulate(dst []uint8, src []float32) {
 	// instead of the maximal value 0xff.
 	//
 	// math.Float32bits(almost256) is 0x437fffff.
-	const almost256 = 255.99998
+	almost256 = 255.99998
+
+	// almost65536 scales a floating point value in the range [0, 1] to a
+	// uint16 value in the range [0x0000, 0xffff].
+	//
+	// math.Float32bits(almost65536) is 0x477fffff.
+	almost65536 = almost256 * 256
+)
+
+func floatingAccumulateOpOver(dst []uint8, src []float32) {
+	// Sanity check that len(dst) >= len(src).
+	if len(dst) < len(src) {
+		return
+	}
+
+	acc := float32(0)
+	for i, v := range src {
+		acc += v
+		a := acc
+		if a < 0 {
+			a = -a
+		}
+		if a > 1 {
+			a = 1
+		}
+		// This algorithm comes from the standard library's image/draw package.
+		dstA := uint32(dst[i]) * 0x101
+		maskA := uint32(almost65536 * a)
+		outA := dstA*(0xffff-maskA)/0xffff + maskA
+		dst[i] = uint8(outA >> 8)
+	}
+}
+
+func floatingAccumulateOpSrc(dst []uint8, src []float32) {
+	// Sanity check that len(dst) >= len(src).
+	if len(dst) < len(src) {
+		return
+	}
 
 	acc := float32(0)
 	for i, v := range src {
@@ -151,3 +186,23 @@ func floatingAccumulate(dst []uint8, src []float32) {
 		dst[i] = uint8(almost256 * a)
 	}
 }
+
+func floatingAccumulateMask(dst []uint32, src []float32) {
+	// Sanity check that len(dst) >= len(src).
+	if len(dst) < len(src) {
+		return
+	}
+
+	acc := float32(0)
+	for i, v := range src {
+		acc += v
+		a := acc
+		if a < 0 {
+			a = -a
+		}
+		if a > 1 {
+			a = 1
+		}
+		dst[i] = uint32(almost65536 * a)
+	}
+}
diff --git a/vendor/golang.org/x/image/vector/vector.go b/vendor/golang.org/x/image/vector/vector.go
index 218e76489..852a4f8b7 100644
--- a/vendor/golang.org/x/image/vector/vector.go
+++ b/vendor/golang.org/x/image/vector/vector.go
@@ -2,6 +2,11 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+//go:generate go run gen.go
+//go:generate asmfmt -w acc_amd64.s
+
+// asmfmt is https://github.com/klauspost/asmfmt
+
 // Package vector provides a rasterizer for 2-D vector graphics.
 package vector // import "golang.org/x/image/vector"
 
@@ -18,24 +23,33 @@ package vector // import "golang.org/x/image/vector"
 
 import (
 	"image"
+	"image/color"
 	"image/draw"
 	"math"
-
-	"golang.org/x/image/math/f32"
 )
 
-func midPoint(p, q f32.Vec2) f32.Vec2 {
-	return f32.Vec2{
-		(p[0] + q[0]) * 0.5,
-		(p[1] + q[1]) * 0.5,
-	}
-}
+// floatingPointMathThreshold is the width or height above which the rasterizer
+// chooses to used floating point math instead of fixed point math.
+//
+// Both implementations of line segmentation rasterization (see raster_fixed.go
+// and raster_floating.go) implement the same algorithm (in ideal, infinite
+// precision math) but they perform differently in practice. The fixed point
+// math version is roughtly 1.25x faster (on GOARCH=amd64) on the benchmarks,
+// but at sufficiently large scales, the computations will overflow and hence
+// show rendering artifacts. The floating point math version has more
+// consistent quality over larger scales, but it is significantly slower.
+//
+// This constant determines when to use the faster implementation and when to
+// use the better quality implementation.
+//
+// The rationale for this particular value is that TestRasterizePolygon in
+// vector_test.go checks the rendering quality of polygon edges at various
+// angles, inscribed in a circle of diameter 512. It may be that a higher value
+// would still produce acceptable quality, but 512 seems to work.
+const floatingPointMathThreshold = 512
 
-func lerp(t float32, p, q f32.Vec2) f32.Vec2 {
-	return f32.Vec2{
-		p[0] + t*(q[0]-p[0]),
-		p[1] + t*(q[1]-p[1]),
-	}
+func lerp(t, px, py, qx, qy float32) (x, y float32) {
+	return px + t*(qx-px), py + t*(qy-py)
 }
 
 func clamp(i, width int32) uint {
@@ -51,18 +65,40 @@ func clamp(i, width int32) uint {
 // NewRasterizer returns a new Rasterizer whose rendered mask image is bounded
 // by the given width and height.
 func NewRasterizer(w, h int) *Rasterizer {
-	return &Rasterizer{
-		area: make([]float32, w*h),
-		size: image.Point{w, h},
-	}
+	z := &Rasterizer{}
+	z.Reset(w, h)
+	return z
 }
 
 // Raster is a 2-D vector graphics rasterizer.
+//
+// The zero value is usable, in that it is a Rasterizer whose rendered mask
+// image has zero width and zero height. Call Reset to change its bounds.
 type Rasterizer struct {
-	area  []float32
-	size  image.Point
-	first f32.Vec2
-	pen   f32.Vec2
+	// bufXxx are buffers of float32 or uint32 values, holding either the
+	// individual or cumulative area values.
+	//
+	// We don't actually need both values at any given time, and to conserve
+	// memory, the integration of the individual to the cumulative could modify
+	// the buffer in place. In other words, we could use a single buffer, say
+	// of type []uint32, and add some math.Float32bits and math.Float32frombits
+	// calls to satisfy the compiler's type checking. As of Go 1.7, though,
+	// there is a performance penalty between:
+	//	bufF32[i] += x
+	// and
+	//	bufU32[i] = math.Float32bits(x + math.Float32frombits(bufU32[i]))
+	//
+	// See golang.org/issue/17220 for some discussion.
+	bufF32 []float32
+	bufU32 []uint32
+
+	useFloatingPointMath bool
+
+	size   image.Point
+	firstX float32
+	firstY float32
+	penX   float32
+	penY   float32
 
 	// DrawOp is the operator used for the Draw method.
 	//
@@ -77,18 +113,39 @@ type Rasterizer struct {
 //
 // This includes setting z.DrawOp to draw.Over.
 func (z *Rasterizer) Reset(w, h int) {
-	if n := w * h; n > cap(z.area) {
-		z.area = make([]float32, n)
+	z.size = image.Point{w, h}
+	z.firstX = 0
+	z.firstY = 0
+	z.penX = 0
+	z.penY = 0
+	z.DrawOp = draw.Over
+
+	z.setUseFloatingPointMath(w > floatingPointMathThreshold || h > floatingPointMathThreshold)
+}
+
+func (z *Rasterizer) setUseFloatingPointMath(b bool) {
+	z.useFloatingPointMath = b
+
+	// Make z.bufF32 or z.bufU32 large enough to hold width * height samples.
+	if z.useFloatingPointMath {
+		if n := z.size.X * z.size.Y; n > cap(z.bufF32) {
+			z.bufF32 = make([]float32, n)
+		} else {
+			z.bufF32 = z.bufF32[:n]
+			for i := range z.bufF32 {
+				z.bufF32[i] = 0
+			}
+		}
 	} else {
-		z.area = z.area[:n]
-		for i := range z.area {
-			z.area[i] = 0
+		if n := z.size.X * z.size.Y; n > cap(z.bufU32) {
+			z.bufU32 = make([]uint32, n)
+		} else {
+			z.bufU32 = z.bufU32[:n]
+			for i := range z.bufU32 {
+				z.bufU32[i] = 0
+			}
 		}
 	}
-	z.size = image.Point{w, h}
-	z.first = f32.Vec2{}
-	z.pen = f32.Vec2{}
-	z.DrawOp = draw.Over
 }
 
 // Size returns the width and height passed to NewRasterizer or Reset.
@@ -104,60 +161,66 @@ func (z *Rasterizer) Bounds() image.Rectangle {
 
 // Pen returns the location of the path-drawing pen: the last argument to the
 // most recent XxxTo call.
-func (z *Rasterizer) Pen() f32.Vec2 {
-	return z.pen
+func (z *Rasterizer) Pen() (x, y float32) {
+	return z.penX, z.penY
 }
 
 // ClosePath closes the current path.
 func (z *Rasterizer) ClosePath() {
-	z.LineTo(z.first)
+	z.LineTo(z.firstX, z.firstY)
 }
 
-// MoveTo starts a new path and moves the pen to a.
+// MoveTo starts a new path and moves the pen to (ax, ay).
 //
 // The coordinates are allowed to be out of the Rasterizer's bounds.
-func (z *Rasterizer) MoveTo(a f32.Vec2) {
-	z.first = a
-	z.pen = a
+func (z *Rasterizer) MoveTo(ax, ay float32) {
+	z.firstX = ax
+	z.firstY = ay
+	z.penX = ax
+	z.penY = ay
 }
 
-// LineTo adds a line segment, from the pen to b, and moves the pen to b.
+// LineTo adds a line segment, from the pen to (bx, by), and moves the pen to
+// (bx, by).
 //
 // The coordinates are allowed to be out of the Rasterizer's bounds.
-func (z *Rasterizer) LineTo(b f32.Vec2) {
-	// TODO: add a fixed point math implementation.
-	z.floatingLineTo(b)
+func (z *Rasterizer) LineTo(bx, by float32) {
+	if z.useFloatingPointMath {
+		z.floatingLineTo(bx, by)
+	} else {
+		z.fixedLineTo(bx, by)
+	}
 }
 
-// QuadTo adds a quadratic Bézier segment, from the pen via b to c, and moves
-// the pen to c.
+// QuadTo adds a quadratic Bézier segment, from the pen via (bx, by) to (cx,
+// cy), and moves the pen to (cx, cy).
 //
 // The coordinates are allowed to be out of the Rasterizer's bounds.
-func (z *Rasterizer) QuadTo(b, c f32.Vec2) {
-	a := z.pen
-	devsq := devSquared(a, b, c)
+func (z *Rasterizer) QuadTo(bx, by, cx, cy float32) {
+	ax, ay := z.penX, z.penY
+	devsq := devSquared(ax, ay, bx, by, cx, cy)
 	if devsq >= 0.333 {
 		const tol = 3
 		n := 1 + int(math.Sqrt(math.Sqrt(tol*float64(devsq))))
 		t, nInv := float32(0), 1/float32(n)
 		for i := 0; i < n-1; i++ {
 			t += nInv
-			ab := lerp(t, a, b)
-			bc := lerp(t, b, c)
-			z.LineTo(lerp(t, ab, bc))
+			abx, aby := lerp(t, ax, ay, bx, by)
+			bcx, bcy := lerp(t, bx, by, cx, cy)
+			z.LineTo(lerp(t, abx, aby, bcx, bcy))
 		}
 	}
-	z.LineTo(c)
+	z.LineTo(cx, cy)
 }
 
-// CubeTo adds a cubic Bézier segment, from the pen via b and c to d, and moves
-// the pen to d.
+// CubeTo adds a cubic Bézier segment, from the pen via (bx, by) and (cx, cy)
+// to (dx, dy), and moves the pen to (dx, dy).
 //
 // The coordinates are allowed to be out of the Rasterizer's bounds.
-func (z *Rasterizer) CubeTo(b, c, d f32.Vec2) {
-	a := z.pen
-	devsq := devSquared(a, b, d)
-	if devsqAlt := devSquared(a, c, d); devsq < devsqAlt {
+func (z *Rasterizer) CubeTo(bx, by, cx, cy, dx, dy float32) {
+	ax, ay := z.penX, z.penY
+	devsq := devSquared(ax, ay, bx, by, dx, dy)
+	if devsqAlt := devSquared(ax, ay, cx, cy, dx, dy); devsq < devsqAlt {
 		devsq = devsqAlt
 	}
 	if devsq >= 0.333 {
@@ -166,19 +229,20 @@ func (z *Rasterizer) CubeTo(b, c, d f32.Vec2) {
 		t, nInv := float32(0), 1/float32(n)
 		for i := 0; i < n-1; i++ {
 			t += nInv
-			ab := lerp(t, a, b)
-			bc := lerp(t, b, c)
-			cd := lerp(t, c, d)
-			abc := lerp(t, ab, bc)
-			bcd := lerp(t, bc, cd)
-			z.LineTo(lerp(t, abc, bcd))
+			abx, aby := lerp(t, ax, ay, bx, by)
+			bcx, bcy := lerp(t, bx, by, cx, cy)
+			cdx, cdy := lerp(t, cx, cy, dx, dy)
+			abcx, abcy := lerp(t, abx, aby, bcx, bcy)
+			bcdx, bcdy := lerp(t, bcx, bcy, cdx, cdy)
+			z.LineTo(lerp(t, abcx, abcy, bcdx, bcdy))
 		}
 	}
-	z.LineTo(d)
+	z.LineTo(dx, dy)
 }
 
-// devSquared returns a measure of how curvy the sequnce a to b to c is. It
-// determines how many line segments will approximate a Bézier curve segment.
+// devSquared returns a measure of how curvy the sequence (ax, ay) to (bx, by)
+// to (cx, cy) is. It determines how many line segments will approximate a
+// Bézier curve segment.
 //
 // http://lists.nongnu.org/archive/html/freetype-devel/2016-08/msg00080.html
 // gives the rationale for this evenly spaced heuristic instead of a recursive
@@ -190,9 +254,9 @@ func (z *Rasterizer) CubeTo(b, c, d f32.Vec2) {
 // Taking a circular arc as a simplifying assumption (ie a spherical cow),
 // where I get n, a recursive approach would get 2^⌈lg n⌉, which, if I haven't
 // made any horrible mistakes, is expected to be 33% more in the limit.
-func devSquared(a, b, c f32.Vec2) float32 {
-	devx := a[0] - 2*b[0] + c[0]
-	devy := a[1] - 2*b[1] + c[1]
+func devSquared(ax, ay, bx, by, cx, cy float32) float32 {
+	devx := ax - 2*bx + cx
+	devy := ay - 2*by + cy
 	return devx*devx + devy*devy
 }
 
@@ -202,27 +266,207 @@ func devSquared(a, b, c f32.Vec2) float32 {
 // The vector paths previously added via the XxxTo calls become the mask for
 // drawing src onto dst.
 func (z *Rasterizer) Draw(dst draw.Image, r image.Rectangle, src image.Image, sp image.Point) {
+	// TODO: adjust r and sp (and mp?) if src.Bounds() doesn't contain
+	// r.Add(sp.Sub(r.Min)).
+
 	if src, ok := src.(*image.Uniform); ok {
-		_, _, _, srcA := src.RGBA()
+		srcR, srcG, srcB, srcA := src.RGBA()
 		switch dst := dst.(type) {
 		case *image.Alpha:
 			// Fast path for glyph rendering.
-			if srcA == 0xffff && z.DrawOp == draw.Src {
-				z.rasterizeDstAlphaSrcOpaqueOpSrc(dst, r)
+			if srcA == 0xffff {
+				if z.DrawOp == draw.Over {
+					z.rasterizeDstAlphaSrcOpaqueOpOver(dst, r)
+				} else {
+					z.rasterizeDstAlphaSrcOpaqueOpSrc(dst, r)
+				}
 				return
 			}
+		case *image.RGBA:
+			if z.DrawOp == draw.Over {
+				z.rasterizeDstRGBASrcUniformOpOver(dst, r, srcR, srcG, srcB, srcA)
+			} else {
+				z.rasterizeDstRGBASrcUniformOpSrc(dst, r, srcR, srcG, srcB, srcA)
+			}
+			return
+		}
+	}
+
+	if z.DrawOp == draw.Over {
+		z.rasterizeOpOver(dst, r, src, sp)
+	} else {
+		z.rasterizeOpSrc(dst, r, src, sp)
+	}
+}
+
+func (z *Rasterizer) accumulateMask() {
+	if z.useFloatingPointMath {
+		if n := z.size.X * z.size.Y; n > cap(z.bufU32) {
+			z.bufU32 = make([]uint32, n)
+		} else {
+			z.bufU32 = z.bufU32[:n]
+		}
+		if haveFloatingAccumulateSIMD {
+			floatingAccumulateMaskSIMD(z.bufU32, z.bufF32)
+		} else {
+			floatingAccumulateMask(z.bufU32, z.bufF32)
+		}
+	} else {
+		if haveFixedAccumulateSIMD {
+			fixedAccumulateMaskSIMD(z.bufU32)
+		} else {
+			fixedAccumulateMask(z.bufU32)
+		}
+	}
+}
+
+func (z *Rasterizer) rasterizeDstAlphaSrcOpaqueOpOver(dst *image.Alpha, r image.Rectangle) {
+	// TODO: non-zero vs even-odd winding?
+	if r == dst.Bounds() && r == z.Bounds() {
+		// We bypass the z.accumulateMask step and convert straight from
+		// z.bufF32 or z.bufU32 to dst.Pix.
+		if z.useFloatingPointMath {
+			if haveFloatingAccumulateSIMD {
+				floatingAccumulateOpOverSIMD(dst.Pix, z.bufF32)
+			} else {
+				floatingAccumulateOpOver(dst.Pix, z.bufF32)
+			}
+		} else {
+			if haveFixedAccumulateSIMD {
+				fixedAccumulateOpOverSIMD(dst.Pix, z.bufU32)
+			} else {
+				fixedAccumulateOpOver(dst.Pix, z.bufU32)
+			}
+		}
+		return
+	}
+
+	z.accumulateMask()
+	pix := dst.Pix[dst.PixOffset(r.Min.X, r.Min.Y):]
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			ma := z.bufU32[y*z.size.X+x]
+			i := y*dst.Stride + x
+
+			// This formula is like rasterizeOpOver's, simplified for the
+			// concrete dst type and opaque src assumption.
+			a := 0xffff - ma
+			pix[i] = uint8((uint32(pix[i])*0x101*a/0xffff + ma) >> 8)
 		}
 	}
-	println("TODO: the general case")
 }
 
 func (z *Rasterizer) rasterizeDstAlphaSrcOpaqueOpSrc(dst *image.Alpha, r image.Rectangle) {
-	// TODO: add SIMD implementations.
-	// TODO: add a fixed point math implementation.
 	// TODO: non-zero vs even-odd winding?
 	if r == dst.Bounds() && r == z.Bounds() {
-		floatingAccumulate(dst.Pix, z.area)
+		// We bypass the z.accumulateMask step and convert straight from
+		// z.bufF32 or z.bufU32 to dst.Pix.
+		if z.useFloatingPointMath {
+			if haveFloatingAccumulateSIMD {
+				floatingAccumulateOpSrcSIMD(dst.Pix, z.bufF32)
+			} else {
+				floatingAccumulateOpSrc(dst.Pix, z.bufF32)
+			}
+		} else {
+			if haveFixedAccumulateSIMD {
+				fixedAccumulateOpSrcSIMD(dst.Pix, z.bufU32)
+			} else {
+				fixedAccumulateOpSrc(dst.Pix, z.bufU32)
+			}
+		}
 		return
 	}
-	println("TODO: the general case")
+
+	z.accumulateMask()
+	pix := dst.Pix[dst.PixOffset(r.Min.X, r.Min.Y):]
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			ma := z.bufU32[y*z.size.X+x]
+
+			// This formula is like rasterizeOpSrc's, simplified for the
+			// concrete dst type and opaque src assumption.
+			pix[y*dst.Stride+x] = uint8(ma >> 8)
+		}
+	}
+}
+
+func (z *Rasterizer) rasterizeDstRGBASrcUniformOpOver(dst *image.RGBA, r image.Rectangle, sr, sg, sb, sa uint32) {
+	z.accumulateMask()
+	pix := dst.Pix[dst.PixOffset(r.Min.X, r.Min.Y):]
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			ma := z.bufU32[y*z.size.X+x]
+
+			// This formula is like rasterizeOpOver's, simplified for the
+			// concrete dst type and uniform src assumption.
+			a := 0xffff - (sa * ma / 0xffff)
+			i := y*dst.Stride + 4*x
+			pix[i+0] = uint8(((uint32(pix[i+0])*0x101*a + sr*ma) / 0xffff) >> 8)
+			pix[i+1] = uint8(((uint32(pix[i+1])*0x101*a + sg*ma) / 0xffff) >> 8)
+			pix[i+2] = uint8(((uint32(pix[i+2])*0x101*a + sb*ma) / 0xffff) >> 8)
+			pix[i+3] = uint8(((uint32(pix[i+3])*0x101*a + sa*ma) / 0xffff) >> 8)
+		}
+	}
+}
+
+func (z *Rasterizer) rasterizeDstRGBASrcUniformOpSrc(dst *image.RGBA, r image.Rectangle, sr, sg, sb, sa uint32) {
+	z.accumulateMask()
+	pix := dst.Pix[dst.PixOffset(r.Min.X, r.Min.Y):]
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			ma := z.bufU32[y*z.size.X+x]
+
+			// This formula is like rasterizeOpSrc's, simplified for the
+			// concrete dst type and uniform src assumption.
+			i := y*dst.Stride + 4*x
+			pix[i+0] = uint8((sr * ma / 0xffff) >> 8)
+			pix[i+1] = uint8((sg * ma / 0xffff) >> 8)
+			pix[i+2] = uint8((sb * ma / 0xffff) >> 8)
+			pix[i+3] = uint8((sa * ma / 0xffff) >> 8)
+		}
+	}
+}
+
+func (z *Rasterizer) rasterizeOpOver(dst draw.Image, r image.Rectangle, src image.Image, sp image.Point) {
+	z.accumulateMask()
+	out := color.RGBA64{}
+	outc := color.Color(&out)
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			sr, sg, sb, sa := src.At(sp.X+x, sp.Y+y).RGBA()
+			ma := z.bufU32[y*z.size.X+x]
+
+			// This algorithm comes from the standard library's image/draw
+			// package.
+			dr, dg, db, da := dst.At(r.Min.X+x, r.Min.Y+y).RGBA()
+			a := 0xffff - (sa * ma / 0xffff)
+			out.R = uint16((dr*a + sr*ma) / 0xffff)
+			out.G = uint16((dg*a + sg*ma) / 0xffff)
+			out.B = uint16((db*a + sb*ma) / 0xffff)
+			out.A = uint16((da*a + sa*ma) / 0xffff)
+
+			dst.Set(r.Min.X+x, r.Min.Y+y, outc)
+		}
+	}
+}
+
+func (z *Rasterizer) rasterizeOpSrc(dst draw.Image, r image.Rectangle, src image.Image, sp image.Point) {
+	z.accumulateMask()
+	out := color.RGBA64{}
+	outc := color.Color(&out)
+	for y, y1 := 0, r.Max.Y-r.Min.Y; y < y1; y++ {
+		for x, x1 := 0, r.Max.X-r.Min.X; x < x1; x++ {
+			sr, sg, sb, sa := src.At(sp.X+x, sp.Y+y).RGBA()
+			ma := z.bufU32[y*z.size.X+x]
+
+			// This algorithm comes from the standard library's image/draw
+			// package.
+			out.R = uint16(sr * ma / 0xffff)
+			out.G = uint16(sg * ma / 0xffff)
+			out.B = uint16(sb * ma / 0xffff)
+			out.A = uint16(sa * ma / 0xffff)
+
+			dst.Set(r.Min.X+x, r.Min.Y+y, outc)
+		}
+	}
 }
diff --git a/vendor/golang.org/x/image/vector/vector_test.go b/vendor/golang.org/x/image/vector/vector_test.go
index 3aa8c192f..f32b992f8 100644
--- a/vendor/golang.org/x/image/vector/vector_test.go
+++ b/vendor/golang.org/x/image/vector/vector_test.go
@@ -7,13 +7,16 @@ package vector
 // TODO: add tests for NaN and Inf coordinates.
 
 import (
+	"fmt"
 	"image"
+	"image/color"
 	"image/draw"
 	"image/png"
+	"math"
+	"math/rand"
 	"os"
+	"path/filepath"
 	"testing"
-
-	"golang.org/x/image/math/f32"
 )
 
 // encodePNG is useful for manually debugging the tests.
@@ -30,47 +33,476 @@ func encodePNG(dstFilename string, src image.Image) error {
 	return closeErr
 }
 
-func TestBasicPath(t *testing.T) {
-	want := []byte{
-		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xe3, 0xaa, 0x3e, 0x00, 0x00, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfa, 0x5f, 0x00, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfc, 0x24, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xa1, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfc, 0x14, 0x00, 0x00,
-		0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x4a, 0x00, 0x00,
-		0x00, 0x00, 0xcc, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x81, 0x00, 0x00,
-		0x00, 0x00, 0x66, 0xff, 0xff, 0xff, 0xff, 0xff, 0xef, 0xe4, 0xff, 0xff, 0xff, 0xb6, 0x00, 0x00,
-		0x00, 0x00, 0x0c, 0xf2, 0xff, 0xff, 0xfe, 0x9e, 0x15, 0x00, 0x15, 0x96, 0xff, 0xce, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x88, 0xfc, 0xe3, 0x43, 0x00, 0x00, 0x00, 0x00, 0x06, 0xcd, 0xdc, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x00, 0x10, 0x0f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x25, 0xde, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x56, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
-		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+func pointOnCircle(center, radius, index, number int) (x, y float32) {
+	c := float64(center)
+	r := float64(radius)
+	i := float64(index)
+	n := float64(number)
+	return float32(c + r*(math.Cos(2*math.Pi*i/n))),
+		float32(c + r*(math.Sin(2*math.Pi*i/n)))
+}
+
+func TestRasterizeOutOfBounds(t *testing.T) {
+	// Set this to a non-empty string such as "/tmp" to manually inspect the
+	// rasterization.
+	//
+	// If empty, this test simply checks that calling LineTo with points out of
+	// the rasterizer's bounds doesn't panic.
+	const tmpDir = ""
+
+	const center, radius, n = 16, 20, 16
+	var z Rasterizer
+	for i := 0; i < n; i++ {
+		for j := 1; j < n/2; j++ {
+			z.Reset(2*center, 2*center)
+			z.MoveTo(1*center, 1*center)
+			z.LineTo(pointOnCircle(center, radius, i+0, n))
+			z.LineTo(pointOnCircle(center, radius, i+j, n))
+			z.ClosePath()
+
+			z.MoveTo(0*center, 0*center)
+			z.LineTo(0*center, 2*center)
+			z.LineTo(2*center, 2*center)
+			z.LineTo(2*center, 0*center)
+			z.ClosePath()
+
+			dst := image.NewAlpha(z.Bounds())
+			z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
+
+			if tmpDir == "" {
+				continue
+			}
+
+			filename := filepath.Join(tmpDir, fmt.Sprintf("out-%02d-%02d.png", i, j))
+			if err := encodePNG(filename, dst); err != nil {
+				t.Error(err)
+			}
+			t.Logf("wrote %s", filename)
+		}
 	}
+}
 
-	z := NewRasterizer(16, 16)
-	z.MoveTo(f32.Vec2{2, 2})
-	z.LineTo(f32.Vec2{8, 2})
-	z.QuadTo(f32.Vec2{14, 2}, f32.Vec2{14, 14})
-	z.CubeTo(f32.Vec2{8, 2}, f32.Vec2{5, 20}, f32.Vec2{2, 8})
+func TestRasterizePolygon(t *testing.T) {
+	var z Rasterizer
+	for radius := 4; radius <= 256; radius *= 2 {
+		for n := 3; n <= 19; n += 4 {
+			z.Reset(2*radius, 2*radius)
+			z.MoveTo(float32(2*radius), float32(1*radius))
+			for i := 1; i < n; i++ {
+				z.LineTo(pointOnCircle(radius, radius, i, n))
+			}
+			z.ClosePath()
+
+			dst := image.NewAlpha(z.Bounds())
+			z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
+
+			if err := checkCornersCenter(dst); err != nil {
+				t.Errorf("radius=%d, n=%d: %v", radius, n, err)
+			}
+		}
+	}
+}
+
+func TestRasterizeAlmostAxisAligned(t *testing.T) {
+	z := NewRasterizer(8, 8)
+	z.MoveTo(2, 2)
+	z.LineTo(6, math.Nextafter32(2, 0))
+	z.LineTo(6, 6)
+	z.LineTo(math.Nextafter32(2, 0), 6)
 	z.ClosePath()
 
 	dst := image.NewAlpha(z.Bounds())
-	z.DrawOp = draw.Src
 	z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
 
-	got := dst.Pix
+	if err := checkCornersCenter(dst); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestRasterizeWideAlmostHorizontalLines(t *testing.T) {
+	var z Rasterizer
+	for i := uint(3); i < 16; i++ {
+		x := float32(int(1 << i))
+
+		z.Reset(8, 8)
+		z.MoveTo(-x, 3)
+		z.LineTo(+x, 4)
+		z.LineTo(+x, 6)
+		z.LineTo(-x, 6)
+		z.ClosePath()
+
+		dst := image.NewAlpha(z.Bounds())
+		z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
+
+		if err := checkCornersCenter(dst); err != nil {
+			t.Errorf("i=%d: %v", i, err)
+		}
+	}
+}
+
+func TestRasterize30Degrees(t *testing.T) {
+	z := NewRasterizer(8, 8)
+	z.MoveTo(4, 4)
+	z.LineTo(8, 4)
+	z.LineTo(4, 6)
+	z.ClosePath()
+
+	dst := image.NewAlpha(z.Bounds())
+	z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
+
+	if err := checkCornersCenter(dst); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestRasterizeRandomLineTos(t *testing.T) {
+	var z Rasterizer
+	for i := 5; i < 50; i++ {
+		n, rng := 0, rand.New(rand.NewSource(int64(i)))
+
+		z.Reset(i+2, i+2)
+		z.MoveTo(float32(i/2), float32(i/2))
+		for ; rng.Intn(16) != 0; n++ {
+			x := 1 + rng.Intn(i)
+			y := 1 + rng.Intn(i)
+			z.LineTo(float32(x), float32(y))
+		}
+		z.ClosePath()
+
+		dst := image.NewAlpha(z.Bounds())
+		z.Draw(dst, dst.Bounds(), image.Opaque, image.Point{})
+
+		if err := checkCorners(dst); err != nil {
+			t.Errorf("i=%d (%d nodes): %v", i, n, err)
+		}
+	}
+}
+
+// checkCornersCenter checks that the corners of the image are all 0x00 and the
+// center is 0xff.
+func checkCornersCenter(m *image.Alpha) error {
+	if err := checkCorners(m); err != nil {
+		return err
+	}
+	size := m.Bounds().Size()
+	center := m.Pix[(size.Y/2)*m.Stride+(size.X/2)]
+	if center != 0xff {
+		return fmt.Errorf("center: got %#02x, want 0xff", center)
+	}
+	return nil
+}
+
+// checkCorners checks that the corners of the image are all 0x00.
+func checkCorners(m *image.Alpha) error {
+	size := m.Bounds().Size()
+	corners := [4]uint8{
+		m.Pix[(0*size.Y+0)*m.Stride+(0*size.X+0)],
+		m.Pix[(0*size.Y+0)*m.Stride+(1*size.X-1)],
+		m.Pix[(1*size.Y-1)*m.Stride+(0*size.X+0)],
+		m.Pix[(1*size.Y-1)*m.Stride+(1*size.X-1)],
+	}
+	if corners != [4]uint8{} {
+		return fmt.Errorf("corners were not all zero: %v", corners)
+	}
+	return nil
+}
+
+var basicMask = []byte{
+	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xe3, 0xaa, 0x3e, 0x00, 0x00, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfa, 0x5f, 0x00, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfc, 0x24, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xa1, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xfc, 0x14, 0x00, 0x00,
+	0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x4a, 0x00, 0x00,
+	0x00, 0x00, 0xcc, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x81, 0x00, 0x00,
+	0x00, 0x00, 0x66, 0xff, 0xff, 0xff, 0xff, 0xff, 0xef, 0xe4, 0xff, 0xff, 0xff, 0xb6, 0x00, 0x00,
+	0x00, 0x00, 0x0c, 0xf2, 0xff, 0xff, 0xfe, 0x9e, 0x15, 0x00, 0x15, 0x96, 0xff, 0xce, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x88, 0xfc, 0xe3, 0x43, 0x00, 0x00, 0x00, 0x00, 0x06, 0xcd, 0xdc, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x00, 0x10, 0x0f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x25, 0xde, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x56, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+	0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+}
+
+func testBasicPath(t *testing.T, prefix string, dst draw.Image, src image.Image, op draw.Op, want []byte) {
+	z := NewRasterizer(16, 16)
+	z.MoveTo(2, 2)
+	z.LineTo(8, 2)
+	z.QuadTo(14, 2, 14, 14)
+	z.CubeTo(8, 2, 5, 20, 2, 8)
+	z.ClosePath()
+
+	z.DrawOp = op
+	z.Draw(dst, z.Bounds(), src, image.Point{})
+
+	var got []byte
+	switch dst := dst.(type) {
+	case *image.Alpha:
+		got = dst.Pix
+	case *image.RGBA:
+		got = dst.Pix
+	default:
+		t.Errorf("%s: unrecognized dst image type %T", prefix, dst)
+	}
+
 	if len(got) != len(want) {
-		t.Fatalf("len(got)=%d and len(want)=%d differ", len(got), len(want))
+		t.Errorf("%s: len(got)=%d and len(want)=%d differ", prefix, len(got), len(want))
+		return
 	}
 	for i := range got {
 		delta := int(got[i]) - int(want[i])
 		// The +/- 2 allows different implementations to give different
 		// rounding errors.
 		if delta < -2 || +2 < delta {
-			t.Errorf("i=%d: got %#02x, want %#02x", i, got[i], want[i])
+			t.Errorf("%s: i=%d: got %#02x, want %#02x", prefix, i, got[i], want[i])
+			return
 		}
 	}
 }
+
+func TestBasicPathDstAlpha(t *testing.T) {
+	for _, background := range []uint8{0x00, 0x80} {
+		for _, op := range []draw.Op{draw.Over, draw.Src} {
+			for _, xPadding := range []int{0, 7} {
+				bounds := image.Rect(0, 0, 16+xPadding, 16)
+				dst := image.NewAlpha(bounds)
+				for i := range dst.Pix {
+					dst.Pix[i] = background
+				}
+
+				want := make([]byte, len(dst.Pix))
+				copy(want, dst.Pix)
+
+				if op == draw.Over && background == 0x80 {
+					for y := 0; y < 16; y++ {
+						for x := 0; x < 16; x++ {
+							ma := basicMask[16*y+x]
+							i := dst.PixOffset(x, y)
+							want[i] = 0xff - (0xff-ma)/2
+						}
+					}
+				} else {
+					for y := 0; y < 16; y++ {
+						for x := 0; x < 16; x++ {
+							ma := basicMask[16*y+x]
+							i := dst.PixOffset(x, y)
+							want[i] = ma
+						}
+					}
+				}
+
+				prefix := fmt.Sprintf("background=%#02x, op=%v, xPadding=%d", background, op, xPadding)
+				testBasicPath(t, prefix, dst, image.Opaque, op, want)
+			}
+		}
+	}
+}
+
+func TestBasicPathDstRGBA(t *testing.T) {
+	blue := image.NewUniform(color.RGBA{0x00, 0x00, 0xff, 0xff})
+
+	for _, op := range []draw.Op{draw.Over, draw.Src} {
+		for _, xPadding := range []int{0, 7} {
+			bounds := image.Rect(0, 0, 16+xPadding, 16)
+			dst := image.NewRGBA(bounds)
+			for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
+				for x := bounds.Min.X; x < bounds.Max.X; x++ {
+					dst.SetRGBA(x, y, color.RGBA{
+						R: uint8(y * 0x07),
+						G: uint8(x * 0x05),
+						B: 0x00,
+						A: 0x80,
+					})
+				}
+			}
+
+			want := make([]byte, len(dst.Pix))
+			copy(want, dst.Pix)
+
+			if op == draw.Over {
+				for y := 0; y < 16; y++ {
+					for x := 0; x < 16; x++ {
+						ma := basicMask[16*y+x]
+						i := dst.PixOffset(x, y)
+						want[i+0] = uint8((uint32(0xff-ma) * uint32(y*0x07)) / 0xff)
+						want[i+1] = uint8((uint32(0xff-ma) * uint32(x*0x05)) / 0xff)
+						want[i+2] = ma
+						want[i+3] = ma/2 + 0x80
+					}
+				}
+			} else {
+				for y := 0; y < 16; y++ {
+					for x := 0; x < 16; x++ {
+						ma := basicMask[16*y+x]
+						i := dst.PixOffset(x, y)
+						want[i+0] = 0x00
+						want[i+1] = 0x00
+						want[i+2] = ma
+						want[i+3] = ma
+					}
+				}
+			}
+
+			prefix := fmt.Sprintf("op=%v, xPadding=%d", op, xPadding)
+			testBasicPath(t, prefix, dst, blue, op, want)
+		}
+	}
+}
+
+const (
+	benchmarkGlyphWidth  = 893
+	benchmarkGlyphHeight = 1122
+)
+
+type benchmarkGlyphDatum struct {
+	// n being 0, 1 or 2 means moveTo, lineTo or quadTo.
+	n  uint32
+	px float32
+	py float32
+	qx float32
+	qy float32
+}
+
+// benchmarkGlyphData is the 'a' glyph from the Roboto Regular font, translated
+// so that its top left corner is (0, 0).
+var benchmarkGlyphData = []benchmarkGlyphDatum{
+	{0, 699, 1102, 0, 0},
+	{2, 683, 1070, 673, 988},
+	{2, 544, 1122, 365, 1122},
+	{2, 205, 1122, 102.5, 1031.5},
+	{2, 0, 941, 0, 802},
+	{2, 0, 633, 128.5, 539.5},
+	{2, 257, 446, 490, 446},
+	{1, 670, 446, 0, 0},
+	{1, 670, 361, 0, 0},
+	{2, 670, 264, 612, 206.5},
+	{2, 554, 149, 441, 149},
+	{2, 342, 149, 275, 199},
+	{2, 208, 249, 208, 320},
+	{1, 22, 320, 0, 0},
+	{2, 22, 239, 79.5, 163.5},
+	{2, 137, 88, 235.5, 44},
+	{2, 334, 0, 452, 0},
+	{2, 639, 0, 745, 93.5},
+	{2, 851, 187, 855, 351},
+	{1, 855, 849, 0, 0},
+	{2, 855, 998, 893, 1086},
+	{1, 893, 1102, 0, 0},
+	{1, 699, 1102, 0, 0},
+	{0, 392, 961, 0, 0},
+	{2, 479, 961, 557, 916},
+	{2, 635, 871, 670, 799},
+	{1, 670, 577, 0, 0},
+	{1, 525, 577, 0, 0},
+	{2, 185, 577, 185, 776},
+	{2, 185, 863, 243, 912},
+	{2, 301, 961, 392, 961},
+}
+
+func scaledBenchmarkGlyphData(height int) (width int, data []benchmarkGlyphDatum) {
+	scale := float32(height) / benchmarkGlyphHeight
+
+	// Clone the benchmarkGlyphData slice and scale its coordinates.
+	data = append(data, benchmarkGlyphData...)
+	for i := range data {
+		data[i].px *= scale
+		data[i].py *= scale
+		data[i].qx *= scale
+		data[i].qy *= scale
+	}
+
+	return int(math.Ceil(float64(benchmarkGlyphWidth * scale))), data
+}
+
+// benchGlyph benchmarks rasterizing a TrueType glyph.
+//
+// Note that, compared to the github.com/google/font-go prototype, the height
+// here is the height of the bounding box, not the pixels per em used to scale
+// a glyph's vectors. A height of 64 corresponds to a ppem greater than 64.
+func benchGlyph(b *testing.B, colorModel byte, loose bool, height int, op draw.Op) {
+	width, data := scaledBenchmarkGlyphData(height)
+	z := NewRasterizer(width, height)
+
+	bounds := z.Bounds()
+	if loose {
+		bounds.Max.X++
+	}
+	dst, src := draw.Image(nil), image.Image(nil)
+	switch colorModel {
+	case 'A':
+		dst = image.NewAlpha(bounds)
+		src = image.Opaque
+	case 'N':
+		dst = image.NewNRGBA(bounds)
+		src = image.NewUniform(color.NRGBA{0x40, 0x80, 0xc0, 0xff})
+	case 'R':
+		dst = image.NewRGBA(bounds)
+		src = image.NewUniform(color.RGBA{0x40, 0x80, 0xc0, 0xff})
+	default:
+		b.Fatal("unsupported color model")
+	}
+	bounds = z.Bounds()
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		z.Reset(width, height)
+		z.DrawOp = op
+		for _, d := range data {
+			switch d.n {
+			case 0:
+				z.MoveTo(d.px, d.py)
+			case 1:
+				z.LineTo(d.px, d.py)
+			case 2:
+				z.QuadTo(d.px, d.py, d.qx, d.qy)
+			}
+		}
+		z.Draw(dst, bounds, src, image.Point{})
+	}
+}
+
+func BenchmarkGlyphAlpha16Over(b *testing.B)  { benchGlyph(b, 'A', false, 16, draw.Over) }
+func BenchmarkGlyphAlpha16Src(b *testing.B)   { benchGlyph(b, 'A', false, 16, draw.Src) }
+func BenchmarkGlyphAlpha32Over(b *testing.B)  { benchGlyph(b, 'A', false, 32, draw.Over) }
+func BenchmarkGlyphAlpha32Src(b *testing.B)   { benchGlyph(b, 'A', false, 32, draw.Src) }
+func BenchmarkGlyphAlpha64Over(b *testing.B)  { benchGlyph(b, 'A', false, 64, draw.Over) }
+func BenchmarkGlyphAlpha64Src(b *testing.B)   { benchGlyph(b, 'A', false, 64, draw.Src) }
+func BenchmarkGlyphAlpha128Over(b *testing.B) { benchGlyph(b, 'A', false, 128, draw.Over) }
+func BenchmarkGlyphAlpha128Src(b *testing.B)  { benchGlyph(b, 'A', false, 128, draw.Src) }
+func BenchmarkGlyphAlpha256Over(b *testing.B) { benchGlyph(b, 'A', false, 256, draw.Over) }
+func BenchmarkGlyphAlpha256Src(b *testing.B)  { benchGlyph(b, 'A', false, 256, draw.Src) }
+
+func BenchmarkGlyphAlphaLoose16Over(b *testing.B)  { benchGlyph(b, 'A', true, 16, draw.Over) }
+func BenchmarkGlyphAlphaLoose16Src(b *testing.B)   { benchGlyph(b, 'A', true, 16, draw.Src) }
+func BenchmarkGlyphAlphaLoose32Over(b *testing.B)  { benchGlyph(b, 'A', true, 32, draw.Over) }
+func BenchmarkGlyphAlphaLoose32Src(b *testing.B)   { benchGlyph(b, 'A', true, 32, draw.Src) }
+func BenchmarkGlyphAlphaLoose64Over(b *testing.B)  { benchGlyph(b, 'A', true, 64, draw.Over) }
+func BenchmarkGlyphAlphaLoose64Src(b *testing.B)   { benchGlyph(b, 'A', true, 64, draw.Src) }
+func BenchmarkGlyphAlphaLoose128Over(b *testing.B) { benchGlyph(b, 'A', true, 128, draw.Over) }
+func BenchmarkGlyphAlphaLoose128Src(b *testing.B)  { benchGlyph(b, 'A', true, 128, draw.Src) }
+func BenchmarkGlyphAlphaLoose256Over(b *testing.B) { benchGlyph(b, 'A', true, 256, draw.Over) }
+func BenchmarkGlyphAlphaLoose256Src(b *testing.B)  { benchGlyph(b, 'A', true, 256, draw.Src) }
+
+func BenchmarkGlyphRGBA16Over(b *testing.B)  { benchGlyph(b, 'R', false, 16, draw.Over) }
+func BenchmarkGlyphRGBA16Src(b *testing.B)   { benchGlyph(b, 'R', false, 16, draw.Src) }
+func BenchmarkGlyphRGBA32Over(b *testing.B)  { benchGlyph(b, 'R', false, 32, draw.Over) }
+func BenchmarkGlyphRGBA32Src(b *testing.B)   { benchGlyph(b, 'R', false, 32, draw.Src) }
+func BenchmarkGlyphRGBA64Over(b *testing.B)  { benchGlyph(b, 'R', false, 64, draw.Over) }
+func BenchmarkGlyphRGBA64Src(b *testing.B)   { benchGlyph(b, 'R', false, 64, draw.Src) }
+func BenchmarkGlyphRGBA128Over(b *testing.B) { benchGlyph(b, 'R', false, 128, draw.Over) }
+func BenchmarkGlyphRGBA128Src(b *testing.B)  { benchGlyph(b, 'R', false, 128, draw.Src) }
+func BenchmarkGlyphRGBA256Over(b *testing.B) { benchGlyph(b, 'R', false, 256, draw.Over) }
+func BenchmarkGlyphRGBA256Src(b *testing.B)  { benchGlyph(b, 'R', false, 256, draw.Src) }
+
+func BenchmarkGlyphNRGBA16Over(b *testing.B)  { benchGlyph(b, 'N', false, 16, draw.Over) }
+func BenchmarkGlyphNRGBA16Src(b *testing.B)   { benchGlyph(b, 'N', false, 16, draw.Src) }
+func BenchmarkGlyphNRGBA32Over(b *testing.B)  { benchGlyph(b, 'N', false, 32, draw.Over) }
+func BenchmarkGlyphNRGBA32Src(b *testing.B)   { benchGlyph(b, 'N', false, 32, draw.Src) }
+func BenchmarkGlyphNRGBA64Over(b *testing.B)  { benchGlyph(b, 'N', false, 64, draw.Over) }
+func BenchmarkGlyphNRGBA64Src(b *testing.B)   { benchGlyph(b, 'N', false, 64, draw.Src) }
+func BenchmarkGlyphNRGBA128Over(b *testing.B) { benchGlyph(b, 'N', false, 128, draw.Over) }
+func BenchmarkGlyphNRGBA128Src(b *testing.B)  { benchGlyph(b, 'N', false, 128, draw.Src) }
+func BenchmarkGlyphNRGBA256Over(b *testing.B) { benchGlyph(b, 'N', false, 256, draw.Over) }
+func BenchmarkGlyphNRGBA256Src(b *testing.B)  { benchGlyph(b, 'N', false, 256, draw.Src) }