зеркало из https://github.com/mozilla/gecko-dev.git
Bug 1155211: Remove selectBits on Float types; r=nbp
--HG-- extra : rebase_source : d78859d4900403c74c3dc25bca70a98ae55d2f23 extra : histedit_source : e135f9d93e9d40d9107e50a19ff66d700be9379f
This commit is contained in:
Benjamin Bouvier
Родитель
44f39fa226
Коммит
995104a40e
|
|
@ -64,7 +64,6 @@
|
|||
V(clamp, Clamp<Float32x4>, 3) \
|
||||
V(replaceLane, (ReplaceLane<Float32x4>), 3) \
|
||||
V(select, (Select<Float32x4, Int32x4>), 3) \
|
||||
V(selectBits, (SelectBits<Float32x4, Int32x4>), 3) \
|
||||
V(store, (Store<Float32x4, 4>), 3) \
|
||||
V(store3, (Store<Float32x4, 3>), 3) \
|
||||
V(store2, (Store<Float32x4, 2>), 3) \
|
||||
|
|
@ -118,7 +117,6 @@
|
|||
V(clamp, Clamp<Float64x2>, 3) \
|
||||
V(replaceLane, (ReplaceLane<Float64x2>), 3) \
|
||||
V(select, (Select<Float64x2, Int32x4>), 3) \
|
||||
V(selectBits, (SelectBits<Float64x2, Int32x4>), 3) \
|
||||
V(store, (Store<Float64x2, 2>), 3) \
|
||||
V(store1, (Store<Float64x2, 1>), 3)
|
||||
|
||||
|
|
@ -292,6 +290,7 @@
|
|||
_(fromFloat32x4Bits)
|
||||
#define FOREACH_INT32X4_SIMD_OP(_) \
|
||||
CONVERSION_INT32X4_SIMD_OP(_) \
|
||||
_(selectBits) \
|
||||
_(shiftLeftByScalar) \
|
||||
_(shiftRightArithmeticByScalar) \
|
||||
_(shiftRightLogicalByScalar)
|
||||
|
|
@ -333,7 +332,6 @@
|
|||
_(extractLane) \
|
||||
_(replaceLane) \
|
||||
_(select) \
|
||||
_(selectBits) \
|
||||
_(splat) \
|
||||
_(not) \
|
||||
_(neg) \
|
||||
|
|
|
|||
|
|
@ -2,23 +2,12 @@ load(libdir + 'simd.js');
|
|||
|
||||
setJitCompilerOption("ion.warmup.trigger", 50);
|
||||
|
||||
function Int32x4FromTypeBits(type, vec) {
|
||||
if (type == SIMD.Int32x4)
|
||||
return vec;
|
||||
if (type == SIMD.Float32x4)
|
||||
return SIMD.Int32x4.fromFloat32x4Bits(vec);
|
||||
throw 'unimplemented';
|
||||
}
|
||||
|
||||
function selectBits(type, mask, ifTrue, ifFalse) {
|
||||
function selectBits(mask, ifTrue, ifFalse) {
|
||||
var Int32x4 = SIMD.Int32x4;
|
||||
var tv = Int32x4FromTypeBits(type, ifTrue);
|
||||
var fv = Int32x4FromTypeBits(type, ifFalse);
|
||||
var tr = Int32x4.and(mask, tv);
|
||||
var fr = Int32x4.and(Int32x4.not(mask), fv);
|
||||
var tr = Int32x4.and(mask, ifTrue);
|
||||
var fr = Int32x4.and(Int32x4.not(mask), ifFalse);
|
||||
var orApplied = Int32x4.or(tr, fr);
|
||||
var converted = type == Int32x4 ? orApplied : type.fromInt32x4Bits(orApplied);
|
||||
return simdToArray(converted);
|
||||
return simdToArray(orApplied);
|
||||
}
|
||||
|
||||
function select(type, mask, ifTrue, ifFalse) {
|
||||
|
|
@ -45,11 +34,11 @@ function f() {
|
|||
for (var i = 0; i < 150; i++) {
|
||||
assertEqX4(SIMD.Float32x4.select(TTFT, f1, f2), select(SIMD.Float32x4, TTFT, f1, f2));
|
||||
assertEqX4(SIMD.Float32x4.select(TFTF, f1, f2), select(SIMD.Float32x4, TFTF, f1, f2));
|
||||
|
||||
assertEqX4(SIMD.Int32x4.select(TFTF, i1, i2), select(SIMD.Int32x4, TFTF, i1, i2));
|
||||
assertEqX4(SIMD.Int32x4.select(TTFT, i1, i2), select(SIMD.Int32x4, TTFT, i1, i2));
|
||||
|
||||
assertEqX4(SIMD.Float32x4.selectBits(mask, f1, f2), selectBits(SIMD.Float32x4, mask, f1, f2));
|
||||
assertEqX4(SIMD.Int32x4.selectBits(mask, i1, i2), selectBits(SIMD.Int32x4, mask, i1, i2));
|
||||
assertEqX4(SIMD.Int32x4.selectBits(mask, i1, i2), selectBits(mask, i1, i2));
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -877,7 +877,6 @@ for (var i = 1; i < 64; i++) {
|
|||
const I32SEL = 'var i4sel = i4.select;'
|
||||
const F32SEL = 'var f4sel = f4.select;'
|
||||
const I32BSEL = 'var i4sel = i4.selectBits;'
|
||||
const F32BSEL = 'var f4sel = f4.selectBits;'
|
||||
|
||||
assertAsmTypeFail('glob', USE_ASM + I32 + F32 + CI32 + I32SEL + "function f() {var x=f4(1,2,3,4); return ci4(i4sel(x,x,x));} return f");
|
||||
assertAsmTypeFail('glob', USE_ASM + I32 + F32 + CI32 + I32SEL + "function f() {var m=f4(1,2,3,4); var x=i4(1,2,3,4); return ci4(i4sel(m,x,x));} return f");
|
||||
|
|
@ -920,11 +919,6 @@ assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + CI32 + I32BSEL + "function
|
|||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + CI32 + I32BSEL + "function f() {var m=i4(0,0xffffffff,0,0xffffffff); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return ci4(i4sel(m,x,y)); } return f"), this)(), [5, 2, 7, 4]);
|
||||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + CI32 + I32BSEL + "function f() {var m=i4(0,0,0xffffffff,0xffffffff); var x=i4(1,2,3,4); var y=i4(5,6,7,8); return ci4(i4sel(m,x,y)); } return f"), this)(), [5, 6, 3, 4]);
|
||||
|
||||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + CF32 + F32BSEL + "function f() {var m=i4(0,0,0,0); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return cf4(f4sel(m,x,y)); } return f"), this)(), [5, 6, 7, 8]);
|
||||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + CF32 + F32BSEL + "function f() {var m=i4(0xffffffff,0xffffffff,0xffffffff,0xffffffff); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return cf4(f4sel(m,x,y)); } return f"), this)(), [1, 2, 3, 4]);
|
||||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + CF32 + F32BSEL + "function f() {var m=i4(0,0xffffffff,0,0xffffffff); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return cf4(f4sel(m,x,y)); } return f"), this)(), [5, 2, 7, 4]);
|
||||
assertEqX4(asmLink(asmCompile('glob', USE_ASM + I32 + F32 + CF32 + F32BSEL + "function f() {var m=i4(0,0,0xffffffff,0xffffffff); var x=f4(1,2,3,4); var y=f4(5,6,7,8); return cf4(f4sel(m,x,y)); } return f"), this)(), [5, 6, 3, 4]);
|
||||
|
||||
// Specific selectBits tests
|
||||
var masks = [
|
||||
SIMD.Int32x4(1337, 0x1337, 0x42, 42),
|
||||
|
|
@ -945,18 +939,6 @@ for (var mask of masks) {
|
|||
}
|
||||
}
|
||||
|
||||
inputs = [
|
||||
[SIMD.Float32x4(0.125,4.25,9.75,16.125), SIMD.Float32x4(1.5,2.75,3.25,4.5)],
|
||||
[SIMD.Float32x4(-1.5,-0,NaN,-Infinity), SIMD.Float32x4(1,-2,13.37,3.13)],
|
||||
[SIMD.Float32x4(1.5,2.75,NaN,Infinity), SIMD.Float32x4(-NaN,-Infinity,9.75,16.125)]
|
||||
];
|
||||
|
||||
var f32bsel = asmLink(asmCompile('glob', USE_ASM + I32 + F32 + CI32 + CF32 + F32BSEL + "function f(mask, ifTrue, ifFalse) {mask=ci4(mask); ifTrue=cf4(ifTrue); ifFalse=cf4(ifFalse); return cf4(f4sel(mask,ifTrue,ifFalse)); } return f"), this)
|
||||
|
||||
for (var mask of masks)
|
||||
for (var [x, y] of inputs)
|
||||
assertEqX4(f32bsel(mask, x, y), simdToArray(SIMD.Float32x4.selectBits(mask, x, y)));
|
||||
|
||||
// Splat
|
||||
const I32SPLAT = 'var splat=i4.splat;'
|
||||
const F32SPLAT = 'var splat=f4.splat;'
|
||||
|
|
|
|||
|
|
@ -368,8 +368,6 @@ IonBuilder::inlineNativeCall(CallInfo& callInfo, JSFunction* target)
|
|||
return inlineSimdSelect(callInfo, native, IsElementWise(false), SimdTypeDescr::Int32x4);
|
||||
if (native == js::simd_float32x4_select)
|
||||
return inlineSimdSelect(callInfo, native, IsElementWise(true), SimdTypeDescr::Float32x4);
|
||||
if (native == js::simd_float32x4_selectBits)
|
||||
return inlineSimdSelect(callInfo, native, IsElementWise(false), SimdTypeDescr::Float32x4);
|
||||
|
||||
if (native == js::simd_int32x4_swizzle)
|
||||
return inlineSimdShuffle(callInfo, native, SimdTypeDescr::Int32x4, 1, 4);
|
||||
|
|
|
|||
|
|
@ -24,12 +24,6 @@ function getMask(i, maskLength) {
|
|||
throw new Error("Invalid mask length.");
|
||||
}
|
||||
|
||||
function selectMaskType(type) {
|
||||
if (type == Int32x4 || type == Float32x4 || type == Float64x2)
|
||||
return Int32x4;
|
||||
return type;
|
||||
}
|
||||
|
||||
function select(mask, ifTrue, ifFalse) {
|
||||
var m = simdToArray(mask);
|
||||
var tv = simdToArray(ifTrue);
|
||||
|
|
@ -57,43 +51,11 @@ function testSelect(type, inputs) {
|
|||
}
|
||||
}
|
||||
|
||||
function intFromTypeBits(type, vec) {
|
||||
switch (type) {
|
||||
case Float32x4:
|
||||
return Int32x4.fromFloat32x4Bits(vec);
|
||||
case Float64x2:
|
||||
return Int32x4.fromFloat64x2Bits(vec);
|
||||
case Int8x16:
|
||||
return vec;
|
||||
case Int16x8:
|
||||
return vec;
|
||||
case Int32x4:
|
||||
return vec;
|
||||
default:
|
||||
throw new TypeError("Unknown SIMD type.");
|
||||
}
|
||||
}
|
||||
|
||||
function selectBits(type, mask, ifTrue, ifFalse) {
|
||||
var maskType = selectMaskType(type);
|
||||
var tv = intFromTypeBits(type, ifTrue);
|
||||
var fv = intFromTypeBits(type, ifFalse);
|
||||
var tr = maskType.and(mask, tv);
|
||||
var fr = maskType.and(maskType.not(mask), fv);
|
||||
var orApplied = maskType.or(tr, fr);
|
||||
var converted = type == maskType ? orApplied : type.fromInt32x4Bits(orApplied);
|
||||
return simdToArray(converted);
|
||||
}
|
||||
|
||||
function findCorrespondingScalarTypedArray(type) {
|
||||
switch (type) {
|
||||
case Int8x16: return Int8Array;
|
||||
case Int16x8: return Int16Array;
|
||||
case Int32x4: return Int32Array;
|
||||
case Float32x4: return Float32Array;
|
||||
case Float64x2: return Float64Array;
|
||||
default: throw new Error("undefined scalar typed array");
|
||||
}
|
||||
var tr = type.and(mask, ifTrue);
|
||||
var fr = type.and(type.not(mask), ifFalse);
|
||||
var orApplied = type.or(tr, fr);
|
||||
return simdToArray(orApplied);
|
||||
}
|
||||
|
||||
/**
|
||||
|
|
@ -103,12 +65,11 @@ function findCorrespondingScalarTypedArray(type) {
|
|||
function testSelectBitsSimple(type, inputs) {
|
||||
var x, y;
|
||||
var maskLength = simdLengthType(type);
|
||||
maskLength = maskLength != 2 ? maskLength : 4;
|
||||
var ScalarTypedArray = findCorrespondingScalarTypedArray(type);
|
||||
for (var i = 0; i < Math.pow(maskLength, 2); i++) {
|
||||
var mask = getMask(i, maskLength);
|
||||
for ([x, y] of inputs)
|
||||
assertEqVec(type.selectBits(mask, x, y), selectBits(type, mask, x, y));
|
||||
assertEqVec(type.selectBits(mask, x, y), simdToArray(type.select(mask, x, y)));
|
||||
}
|
||||
}
|
||||
|
||||
|
|
@ -129,19 +90,18 @@ function testSelectBitsComplex(type, inputs) {
|
|||
Int32x4(0x00FF1CE, 0xBAADF00D, 0xDEADBEEF, 0xCAFED00D),
|
||||
Int32x4(0xD15EA5E, 0xDEADC0DE, 0xFACEB00C, 0x4B1D4B1D)
|
||||
];
|
||||
var masks = [];
|
||||
var maskType = selectMaskType(type);
|
||||
if (maskType == SIMD.Int8x16)
|
||||
|
||||
var masks;
|
||||
if (type == SIMD.Int8x16)
|
||||
masks = masks8;
|
||||
else if (maskType == SIMD.Int16x8)
|
||||
else if (type == SIMD.Int16x8)
|
||||
masks = masks16;
|
||||
else if (maskType == SIMD.Int32x4)
|
||||
else if (type == SIMD.Int32x4)
|
||||
masks = masks32;
|
||||
else
|
||||
throw new Error("Unknown mask type.");
|
||||
|
||||
var x, y;
|
||||
var ScalarTypedArray = findCorrespondingScalarTypedArray(type);
|
||||
for (var mask of masks) {
|
||||
for ([x, y] of inputs)
|
||||
assertEqVec(type.selectBits(mask, x, y), selectBits(type, mask, x, y));
|
||||
|
|
@ -185,8 +145,6 @@ function test() {
|
|||
];
|
||||
|
||||
testSelect(Float32x4, inputs);
|
||||
testSelectBitsSimple(Float32x4, inputs);
|
||||
testSelectBitsComplex(Float32x4, inputs);
|
||||
|
||||
inputs = [
|
||||
[Float64x2(0.125,4.25), Float64x2(9.75,16.125)],
|
||||
|
|
@ -198,8 +156,6 @@ function test() {
|
|||
];
|
||||
|
||||
testSelect(Float64x2, inputs);
|
||||
testSelectBitsSimple(Float64x2, inputs);
|
||||
testSelectBitsComplex(Float64x2, inputs);
|
||||
|
||||
if (typeof reportCompare === "function")
|
||||
reportCompare(true, true);
|
||||
|
|
|
|||
Загрузка…
Ссылка в новой задаче