golang-image/vector/raster_floating.go
Michael Munday e20db36d77 vector: prevent fused multiply-add being used in floatingLineTo
The test code now produces bit-identical results across platforms.

Fixes golang/go#21460.

Change-Id: I58b72dacf3c7930e2804b4e37bd3ccd9ef3dc2f0
Reviewed-on: https://go-review.googlesource.com/61024
Run-TryBot: Michael Munday <mike.munday@ibm.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Nigel Tao <nigeltao@golang.org>
2017-09-07 23:27:51 +00:00

221 lines
5.4 KiB
Go

// 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 floating point math implementation of the vector
// graphics rasterizer.
import (
"math"
)
func floatingMax(x, y float32) float32 {
if x > y {
return x
}
return y
}
func floatingMin(x, y float32) float32 {
if x < y {
return x
}
return y
}
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(bx, by float32) {
ax, ay := z.penX, z.penY
z.penX, z.penY = bx, by
dir := float32(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 floating point
// math, so we treat the segment as if it was perfectly horizontal.
if by-ay <= 0.000001 {
return
}
dxdy := (bx - ax) / (by - ay)
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), by) - floatingMax(float32(y), ay)
// The "float32" in expressions like "float32(foo*bar)" here and below
// look redundant, since foo and bar already have type float32, but are
// explicit in order to disable the compiler's Fused Multiply Add (FMA)
// instruction selection, which can improve performance but can result
// in different rounding errors in floating point computations.
//
// This package aims to have bit-exact identical results across all
// GOARCHes, and across pure Go code and assembly, so it disables FMA.
//
// See the discussion at
// https://groups.google.com/d/topic/golang-dev/Sti0bl2xUXQ/discussion
xNext := x + float32(dy*dxdy)
if y < 0 {
x = xNext
continue
}
buf := z.bufF32[y*width:]
d := float32(dy * dir)
x0, x1 := x, xNext
if x > xNext {
x0, x1 = x1, x0
}
x0i := floatingFloor(x0)
x0Floor := float32(x0i)
x1i := floatingCeil(x1)
x1Ceil := float32(x1i)
if x1i <= x0i+1 {
xmf := float32(0.5*(x+xNext)) - x0Floor
if i := clamp(x0i+0, width); i < uint(len(buf)) {
buf[i] += d - float32(d*xmf)
}
if i := clamp(x0i+1, width); i < uint(len(buf)) {
buf[i] += float32(d * xmf)
}
} else {
s := 1 / (x1 - x0)
x0f := x0 - x0Floor
oneMinusX0f := 1 - x0f
a0 := float32(0.5 * s * oneMinusX0f * oneMinusX0f)
x1f := x1 - x1Ceil + 1
am := float32(0.5 * s * x1f * x1f)
if i := clamp(x0i, width); i < uint(len(buf)) {
buf[i] += float32(d * a0)
}
if x1i == x0i+2 {
if i := clamp(x0i+1, width); i < uint(len(buf)) {
buf[i] += float32(d * (1 - a0 - am))
}
} else {
a1 := float32(s * (1.5 - x0f))
if i := clamp(x0i+1, width); i < uint(len(buf)) {
buf[i] += float32(d * (a1 - a0))
}
dTimesS := float32(d * s)
for xi := x0i + 2; xi < x1i-1; xi++ {
if i := clamp(xi, width); i < uint(len(buf)) {
buf[i] += dTimesS
}
}
a2 := a1 + float32(s*float32(x1i-x0i-3))
if i := clamp(x1i-1, width); i < uint(len(buf)) {
buf[i] += float32(d * (1 - a2 - am))
}
}
if i := clamp(x1i, width); i < uint(len(buf)) {
buf[i] += float32(d * am)
}
}
x = xNext
}
}
const (
// almost256 scales a floating point value in the range [0, 1] to a uint8
// value in the range [0x00, 0xff].
//
// 255 is too small. Floating point math accumulates rounding errors, so a
// fully covered src value that would in ideal math be float32(1) might be
// float32(1-ε), and uint8(255 * (1-ε)) would be 0xfe instead of 0xff. The
// uint8 conversion rounds to zero, not to nearest.
//
// 256 is too big. If we multiplied by 256, below, then a fully covered src
// value of float32(1) would translate to uint8(256 * 1), which can be 0x00
// instead of the maximal value 0xff.
//
// math.Float32bits(almost256) is 0x437fffff.
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 {
acc += v
a := acc
if a < 0 {
a = -a
}
if a > 1 {
a = 1
}
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)
}
}