AuroraMiddleware/SPIRV-Cross
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

MSL: Selectively enable fast-math in MSL code to match Vulkan CTS results.

Browse Source 
Based on CTS testing, math optimizations between MSL and Vulkan are inconsistent.
In some cases, enabling MSL's fast-math compilation option matches Vulkan's math
results. In other cases, disabling it does. Broadly enabling or disabling fast-math
across all shaders results in some CTS test failures either way.

To fix this, selectively enable/disable fast-math optimizations in the MSL code,
using metal::fast and metal::precise function namespaces, where supported, and
the [[clang::optnone]] function attribute otherwise.

Adjust SPIRV-Cross unit test reference shaders to accommodate these changes.
This commit is contained in:
Bill Hollings 2021-09-22 18:58:31 -04:00
parent b81334a513
commit 40141ffddf
28 changed files with 75 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
reference/opt/shaders-msl/flatten/struct.flatten.vert
View File

@ -34,7 +34,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _18 [[buffer(0)]]
out.gl_Position = _18.uMVP * in.aVertex;
out.vColor = float4(0.0);
float3 _39 = in.aVertex.xyz - float3(_18.light.Position);
out.vColor += ((_18.light.Color * fast::clamp(1.0 - (length(_39) / _18.light.Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(_39)));
out.vColor += ((_18.light.Color * fast::clamp(1.0 - (length(_39) / _18.light.Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(_39)));
return out;
}

2
reference/opt/shaders-msl/frag/in_mat.frag
View File

@ -27,7 +27,7 @@ fragment main0_out main0(main0_in in [[stage_in]], texturecube<float> samplerCol
inInvModelView[1] = in.inInvModelView_1;
inInvModelView[2] = in.inInvModelView_2;
inInvModelView[3] = in.inInvModelView_3;
float4 _31 = inInvModelView * float4(reflect(normalize(in.inPos), normalize(in.inNormal)), 0.0);
float4 _31 = inInvModelView * float4(reflect(fast::normalize(in.inPos), fast::normalize(in.inNormal)), 0.0);
float _33 = _31.x;
float3 _59 = float3(_33, _31.yz);
_59.x = _33 * (-1.0);

2
reference/opt/shaders-msl/frag/scalar-refract-reflect.frag
View File

@ -16,7 +16,7 @@ struct main0_in
};
template<typename T>
inline T spvReflect(T i, T n)
[[clang::optnone]] T spvReflect(T i, T n)
{
return i - T(2) * i * n * n;
}

2
reference/opt/shaders-msl/vert/copy.flatten.vert
View File

@ -36,7 +36,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _21 [[buffer(0)]]
for (int _96 = 0; _96 < 4; )
{
float3 _68 = in.aVertex.xyz - float3(_21.lights[_96].Position);
out.vColor += ((_21.lights[_96].Color * fast::clamp(1.0 - (length(_68) / _21.lights[_96].Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(_68)));
out.vColor += ((_21.lights[_96].Color * fast::clamp(1.0 - (length(_68) / _21.lights[_96].Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(_68)));
_96++;
continue;
}

2
reference/opt/shaders-msl/vert/dynamic.flatten.vert
View File

@ -36,7 +36,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _21 [[buffer(0)]]
for (int _82 = 0; _82 < 4; )
{
float3 _54 = in.aVertex.xyz - float3(_21.lights[_82].Position);
out.vColor += ((_21.lights[_82].Color * fast::clamp(1.0 - (length(_54) / _21.lights[_82].Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(_54)));
out.vColor += ((_21.lights[_82].Color * fast::clamp(1.0 - (length(_54) / _21.lights[_82].Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(_54)));
_82++;
continue;
}

2
reference/opt/shaders-msl/vert/packed_matrix.vert
View File

@ -39,7 +39,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant _RESERVED_IDENTIFIER_F
{
main0_out out = {};
float4 _70 = _RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m0 * float4(float3(_RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m10) + (in._RESERVED_IDENTIFIER_FIXUP_5275.xyz * (_RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m17 + _RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m18)), 1.0);
out._RESERVED_IDENTIFIER_FIXUP_3976 = normalize(float4(in._RESERVED_IDENTIFIER_FIXUP_5275.xyz, 0.0) * _RESERVED_IDENTIFIER_FIXUP_18812._RESERVED_IDENTIFIER_FIXUP_m1);
out._RESERVED_IDENTIFIER_FIXUP_3976 = fast::normalize(float4(in._RESERVED_IDENTIFIER_FIXUP_5275.xyz, 0.0) * _RESERVED_IDENTIFIER_FIXUP_18812._RESERVED_IDENTIFIER_FIXUP_m1);
float4 _94 = _70;
_94.y = -_70.y;
out.gl_Position = _94;

4
reference/opt/shaders-ue4/asm/frag/depth-compare.asm.frag
View File

@ -246,7 +246,7 @@ fragment main0_out main0(constant type_View& View [[buffer(0)]], constant type_G
float _218 = sqrt(mix(1.0, _213 * _213, _Globals.ShadowFadeFraction));
float4 _219 = _453;
_219.z = _218;
float3 _236 = normalize((SceneTexturesStruct_GBufferATexture.sample(SceneTexturesStruct_GBufferATextureSampler, _114, level(0.0)).xyz * float3(2.0)) - float3(1.0));
float3 _236 = fast::normalize((SceneTexturesStruct_GBufferATexture.sample(SceneTexturesStruct_GBufferATextureSampler, _114, level(0.0)).xyz * float3(2.0)) - float3(1.0));
uint _240 = uint(round(SceneTexturesStruct_GBufferBTexture.sample(SceneTexturesStruct_GBufferBTextureSampler, _114, level(0.0)).w * 255.0));
bool _248 = (_240 & 15u) == 5u;
float _448;
@ -259,7 +259,7 @@ fragment main0_out main0(constant type_View& View [[buffer(0)]], constant type_G
if (_160)
{
float3 _278 = _152 / float3(_158);
float3 _280 = normalize(cross(_278, float3(0.0, 0.0, 1.0)));
float3 _280 = fast::normalize(cross(_278, float3(0.0, 0.0, 1.0)));
float3 _284 = float3(_Globals.InvShadowmapResolution);
float3 _285 = _280 * _284;
float3 _286 = cross(_280, _278) * _284;

8
reference/opt/shaders-ue4/asm/frag/global-constant-arrays.asm.frag
View File

@ -176,7 +176,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
float _644 = ((0.317398726940155029296875 + (4.25 * _Globals.WhiteTemp)) + ((4.2048167614439080352894961833954e-08 * _Globals.WhiteTemp) * _Globals.WhiteTemp)) / ((1.0 - (2.8974181986995972692966461181641e-05 * _Globals.WhiteTemp)) + ((1.6145605741257895715534687042236e-07 * _Globals.WhiteTemp) * _Globals.WhiteTemp));
float _649 = ((2.0 * _633) - (8.0 * _644)) + 4.0;
float2 _653 = float2((3.0 * _633) / _649, (2.0 * _644) / _649);
float2 _660 = normalize(float2(_633, _644));
float2 _660 = fast::normalize(float2(_633, _644));
float _665 = _633 + (((-_660.y) * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _669 = _644 + ((_660.x * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _674 = ((2.0 * _665) - (8.0 * _669)) + 4.0;
@ -272,7 +272,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_990 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_974 - _976), ((2.0 * _973) - _974) - _976);
_990 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_974 - _976), ((2.0 * _973) - _974) - _976);
}
float _995;
if (_990 < 0.0)
@ -434,7 +434,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_2174 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_2158 - _2160), ((2.0 * _2157) - _2158) - _2160);
_2174 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_2158 - _2160), ((2.0 * _2157) - _2158) - _2160);
}
float _2179;
if (_2174 < 0.0)
@ -796,7 +796,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1337 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_1321 - _1323), ((2.0 * _1320) - _1321) - _1323);
_1337 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_1321 - _1323), ((2.0 * _1320) - _1321) - _1323);
}
float _1342;
if (_1337 < 0.0)

8
reference/opt/shaders-ue4/asm/frag/padded-float-array-member-defef.asm.frag
View File

@ -178,7 +178,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
float _670 = ((0.317398726940155029296875 + (4.25 * _Globals.WhiteTemp)) + ((4.2048167614439080352894961833954e-08 * _Globals.WhiteTemp) * _Globals.WhiteTemp)) / ((1.0 - (2.8974181986995972692966461181641e-05 * _Globals.WhiteTemp)) + ((1.6145605741257895715534687042236e-07 * _Globals.WhiteTemp) * _Globals.WhiteTemp));
float _675 = ((2.0 * _659) - (8.0 * _670)) + 4.0;
float2 _679 = float2((3.0 * _659) / _675, (2.0 * _670) / _675);
float2 _686 = normalize(float2(_659, _670));
float2 _686 = fast::normalize(float2(_659, _670));
float _691 = _659 + (((-_686.y) * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _695 = _670 + ((_686.x * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _700 = ((2.0 * _691) - (8.0 * _695)) + 4.0;
@ -274,7 +274,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1016 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_1000 - _1002), ((2.0 * _999) - _1000) - _1002);
_1016 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_1000 - _1002), ((2.0 * _999) - _1000) - _1002);
}
float _1021;
if (_1016 < 0.0)
@ -483,7 +483,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_2278 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_2262 - _2264), ((2.0 * _2261) - _2262) - _2264);
_2278 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_2262 - _2264), ((2.0 * _2261) - _2262) - _2264);
}
float _2283;
if (_2278 < 0.0)
@ -845,7 +845,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1443 = 57.2957763671875 * atan2(1.73205077648162841796875 * (_1427 - _1429), ((2.0 * _1426) - _1427) - _1429);
_1443 = 57.2957763671875 * precise::atan2(1.73205077648162841796875 * (_1427 - _1429), ((2.0 * _1426) - _1427) - _1429);
}
float _1448;
if (_1443 < 0.0)

2
reference/opt/shaders-ue4/asm/frag/sample-mask-not-array.asm.frag
View File

@ -462,7 +462,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
{
_215 = in.in_var_TEXCOORD7;
}
float3 _216 = fast::max(Material.Material_VectorExpressions[1].xyz * float3(((1.0 + dot(float3(-0.2857142984867095947265625, -0.4285714328289031982421875, 0.857142865657806396484375), normalize(float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz) * normalize((float3(0.0, 0.0, 1.0) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz)))) * 0.5) + 0.20000000298023223876953125), float3(0.0));
float3 _216 = fast::max(Material.Material_VectorExpressions[1].xyz * float3(((1.0 + dot(float3(-0.2857142984867095947265625, -0.4285714328289031982421875, 0.857142865657806396484375), fast::normalize(float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz) * fast::normalize((float3(0.0, 0.0, 1.0) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz)))) * 0.5) + 0.20000000298023223876953125), float3(0.0));
float3 _246;
if (View.View_OutOfBoundsMask > 0.0)
{

6
reference/opt/shaders-ue4/asm/vert/texture-buffer.asm.vert
View File

@ -296,7 +296,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant type_View& View [[buff
float4 _145 = AttributesTexture.sample(AttributesTextureSampler, _133, level(0.0));
float _146 = _137.w;
float3 _158 = float3x3(Primitive.Primitive_LocalToWorld[0].xyz, Primitive.Primitive_LocalToWorld[1].xyz, Primitive.Primitive_LocalToWorld[2].xyz) * VelocityTexture.sample(VelocityTextureSampler, _133, level(0.0)).xyz;
float3 _160 = normalize(_158 + float3(0.0, 0.0, 9.9999997473787516355514526367188e-05));
float3 _160 = fast::normalize(_158 + float3(0.0, 0.0, 9.9999997473787516355514526367188e-05));
float2 _204 = ((((_145.xy + float2((_145.x < 0.5) ? 0.0 : (-0.5), (_145.y < 0.5) ? 0.0 : (-0.5))) * float2(2.0)) * (((CurveTexture.sample(CurveTextureSampler, (EmitterUniforms.EmitterUniforms_MiscCurve.xy + (EmitterUniforms.EmitterUniforms_MiscCurve.zw * float2(_146))), level(0.0)) * EmitterUniforms.EmitterUniforms_MiscScale) + EmitterUniforms.EmitterUniforms_MiscBias).xy * EmitterDynamicUniforms.EmitterDynamicUniforms_LocalToWorldScale)) * fast::min(fast::max(EmitterUniforms.EmitterUniforms_SizeBySpeed.xy * float2(length(_158)), float2(1.0)), EmitterUniforms.EmitterUniforms_SizeBySpeed.zw)) * float2(step(_146, 1.0));
float3 _239 = float4((((Primitive.Primitive_LocalToWorld[0u].xyz * _137.xxx) + (Primitive.Primitive_LocalToWorld[1u].xyz * _137.yyy)) + (Primitive.Primitive_LocalToWorld[2u].xyz * _137.zzz)) + (Primitive.Primitive_LocalToWorld[3u].xyz + float3(View.View_PreViewTranslation)), 1.0).xyz;
float3 _242 = float3(EmitterUniforms.EmitterUniforms_RemoveHMDRoll);
@ -312,8 +312,8 @@ vertex main0_out main0(main0_in in [[stage_in]], constant type_View& View [[buff
float3 _279 = cross(_265, float3(0.0, 0.0, 1.0));
float3 _284 = _279 / float3(sqrt(fast::max(dot(_279, _279), 0.00999999977648258209228515625)));
float3 _286 = float3(fast::clamp((_261 * EmitterUniforms.EmitterUniforms_CameraFacingBlend[1]) - EmitterUniforms.EmitterUniforms_CameraFacingBlend[2], 0.0, 1.0));
_335 = normalize(mix(_251, _284, _286));
_336 = normalize(mix(_259, cross(_265, _284), _286));
_335 = fast::normalize(mix(_251, _284, _286));
_336 = fast::normalize(mix(_259, cross(_265, _284), _286));
}
else
{

12
reference/shaders-msl-no-opt/comp/glsl.std450.comp
View File

@ -170,7 +170,7 @@ float2x2 spvInverse2x2(float2x2 m)
}
template<typename T>
inline T spvReflect(T i, T n)
[[clang::optnone]] T spvReflect(T i, T n)
{
return i - T(2) * i * n * n;
}
@ -217,13 +217,13 @@ kernel void main0(device SSBO& _19 [[buffer(0)]])
_19.res = asin(((device float*)&_19.f32)[0u]);
_19.res = acos(((device float*)&_19.f32)[0u]);
_19.res = atan(((device float*)&_19.f32)[0u]);
_19.res = sinh(((device float*)&_19.f32)[0u]);
_19.res = cosh(((device float*)&_19.f32)[0u]);
_19.res = tanh(((device float*)&_19.f32)[0u]);
_19.res = fast::sinh(((device float*)&_19.f32)[0u]);
_19.res = fast::cosh(((device float*)&_19.f32)[0u]);
_19.res = precise::tanh(((device float*)&_19.f32)[0u]);
_19.res = asinh(((device float*)&_19.f32)[0u]);
_19.res = acosh(((device float*)&_19.f32)[0u]);
_19.res = atanh(((device float*)&_19.f32)[0u]);
_19.res = atan2(((device float*)&_19.f32)[0u], ((device float*)&_19.f32)[1u]);
_19.res = precise::atan2(((device float*)&_19.f32)[0u], ((device float*)&_19.f32)[1u]);
_19.res = pow(((device float*)&_19.f32)[0u], ((device float*)&_19.f32)[1u]);
_19.res = exp(((device float*)&_19.f32)[0u]);
_19.res = log(((device float*)&_19.f32)[0u]);
@ -239,7 +239,7 @@ kernel void main0(device SSBO& _19 [[buffer(0)]])
_19.res = spvRefract(((device float*)&_19.f32)[0u], ((device float*)&_19.f32)[1u], ((device float*)&_19.f32)[2u]);
_19.res = length(_19.f32.xy);
_19.res = distance(_19.f32.xy, _19.f32.zw);
float2 v2 = normalize(_19.f32.xy);
float2 v2 = fast::normalize(_19.f32.xy);
v2 = faceforward(_19.f32.xy, _19.f32.yz, _19.f32.zw);
v2 = reflect(_19.f32.xy, _19.f32.zw);
v2 = refract(_19.f32.xy, _19.f32.yz, ((device float*)&_19.f32)[3u]);

10
reference/shaders-msl-no-opt/frag/fp16.desktop.invalid.frag
View File

@ -94,11 +94,11 @@ void test_builtins(thread half4& v4, thread half3& v3, thread half& v1)
res = cos(v4);
res = tan(v4);
res = asin(v4);
res = atan2(v4, v3.xyzz);
res = precise::atan2(v4, v3.xyzz);
res = atan(v4);
res = sinh(v4);
res = cosh(v4);
res = tanh(v4);
res = fast::sinh(v4);
res = fast::cosh(v4);
res = precise::tanh(v4);
res = asinh(v4);
res = acosh(v4);
res = atanh(v4);
@ -143,7 +143,7 @@ void test_builtins(thread half4& v4, thread half3& v3, thread half& v1)
t0 = distance(v4, v4);
t0 = dot(v4, v4);
half3 res3 = cross(v3, v3);
res = normalize(v4);
res = fast::normalize(v4);
res = faceforward(v4, v4, v4);
res = reflect(v4, v4);
res = refract(v4, v4, v1);

2
reference/shaders-msl/flatten/struct.flatten.vert
View File

@ -34,7 +34,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _18 [[buffer(0)]]
out.gl_Position = _18.uMVP * in.aVertex;
out.vColor = float4(0.0);
float3 L = in.aVertex.xyz - float3(_18.light.Position);
out.vColor += ((_18.light.Color * fast::clamp(1.0 - (length(L) / _18.light.Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(L)));
out.vColor += ((_18.light.Color * fast::clamp(1.0 - (length(L) / _18.light.Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(L)));
return out;
}

4
reference/shaders-msl/frag/in_mat.frag
View File

@ -27,8 +27,8 @@ fragment main0_out main0(main0_in in [[stage_in]], texturecube<float> samplerCol
inInvModelView[1] = in.inInvModelView_1;
inInvModelView[2] = in.inInvModelView_2;
inInvModelView[3] = in.inInvModelView_3;
float3 cI = normalize(in.inPos);
float3 cR = reflect(cI, normalize(in.inNormal));
float3 cI = fast::normalize(in.inPos);
float3 cR = reflect(cI, fast::normalize(in.inNormal));
cR = float3((inInvModelView * float4(cR, 0.0)).xyz);
cR.x *= (-1.0);
out.outFragColor = samplerColor.sample(samplerColorSmplr, cR, bias(in.inLodBias));

2
reference/shaders-msl/frag/scalar-refract-reflect.frag
View File

@ -16,7 +16,7 @@ struct main0_in
};
template<typename T>
inline T spvReflect(T i, T n)
[[clang::optnone]] T spvReflect(T i, T n)
{
return i - T(2) * i * n * n;
}

2
reference/shaders-msl/vert/copy.flatten.vert
View File

@ -47,7 +47,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _21 [[buffer(0)]]
light.Radius = _21.lights[i].Radius;
light.Color = _21.lights[i].Color;
float3 L = in.aVertex.xyz - light.Position;
out.vColor += ((_21.lights[i].Color * fast::clamp(1.0 - (length(L) / light.Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(L)));
out.vColor += ((_21.lights[i].Color * fast::clamp(1.0 - (length(L) / light.Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(L)));
}
return out;
}

2
reference/shaders-msl/vert/dynamic.flatten.vert
View File

@ -36,7 +36,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant UBO& _21 [[buffer(0)]]
for (int i = 0; i < 4; i++)
{
float3 L = in.aVertex.xyz - float3(_21.lights[i].Position);
out.vColor += ((_21.lights[i].Color * fast::clamp(1.0 - (length(L) / _21.lights[i].Radius), 0.0, 1.0)) * dot(in.aNormal, normalize(L)));
out.vColor += ((_21.lights[i].Color * fast::clamp(1.0 - (length(L) / _21.lights[i].Radius), 0.0, 1.0)) * dot(in.aNormal, fast::normalize(L)));
}
return out;
}

2
reference/shaders-msl/vert/packed_matrix.vert
View File

@ -42,7 +42,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant _RESERVED_IDENTIFIER_F
float3 _RESERVED_IDENTIFIER_FIXUP_23783;
for (;;)
{
_RESERVED_IDENTIFIER_FIXUP_23783 = normalize(float4(in._RESERVED_IDENTIFIER_FIXUP_5275.xyz, 0.0) * _RESERVED_IDENTIFIER_FIXUP_18812._RESERVED_IDENTIFIER_FIXUP_m1);
_RESERVED_IDENTIFIER_FIXUP_23783 = fast::normalize(float4(in._RESERVED_IDENTIFIER_FIXUP_5275.xyz, 0.0) * _RESERVED_IDENTIFIER_FIXUP_18812._RESERVED_IDENTIFIER_FIXUP_m1);
break;
}
float4 _RESERVED_IDENTIFIER_FIXUP_14995 = _RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m0 * float4(float3(_RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m10) + (in._RESERVED_IDENTIFIER_FIXUP_5275.xyz * (_RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m17 + _RESERVED_IDENTIFIER_FIXUP_22044._RESERVED_IDENTIFIER_FIXUP_m18)), 1.0);

8
reference/shaders-ue4-no-opt/asm/frag/accesschain-invalid-expression.asm.invalid.frag
View File

@ -217,10 +217,10 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
main0_out out = {};
float4 _177 = float4((((gl_FragCoord.xy - View.View_ViewRectMin.xy) * View.View_ViewSizeAndInvSize.zw) - float2(0.5)) * float2(2.0, -2.0), _138, 1.0) * float4(gl_FragCoord.w);
float3 _179 = in.in_var_TEXCOORD8.xyz - float3(View.View_PreViewTranslation);
float3 _181 = normalize(-in.in_var_TEXCOORD8.xyz);
float3 _181 = fast::normalize(-in.in_var_TEXCOORD8.xyz);
float4 _187 = Material_Texture2D_0.sample(Material_Texture2D_0Sampler, (in.in_var_TEXCOORD0 * float2(10.0)));
float2 _190 = (_187.xy * float2(2.0)) - float2(1.0);
float3 _206 = normalize(float3x3(float3(1.0, 0.0, 0.0), float3(0.0, 1.0, 0.0), float3(0.0, 0.0, 1.0)) * (((float4(_190, sqrt(fast::clamp(1.0 - dot(_190, _190), 0.0, 1.0)), 1.0).xyz * float3(0.300000011920928955078125, 0.300000011920928955078125, 1.0)) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz));
float3 _206 = fast::normalize(float3x3(float3(1.0, 0.0, 0.0), float3(0.0, 1.0, 0.0), float3(0.0, 0.0, 1.0)) * (((float4(_190, sqrt(fast::clamp(1.0 - dot(_190, _190), 0.0, 1.0)), 1.0).xyz * float3(0.300000011920928955078125, 0.300000011920928955078125, 1.0)) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz));
float _208 = dot(_206, _181);
float4 _217 = Material_Texture2D_1.sample(Material_Texture2D_1Sampler, (in.in_var_TEXCOORD0 * float2(20.0)));
float _219 = mix(0.4000000059604644775390625, 1.0, _217.x);
@ -301,7 +301,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
{
_423 = 1.0;
}
float3 _429 = normalize(_181 + MobileDirectionalLight.MobileDirectionalLight_DirectionalLightDirectionAndShadowTransition.xyz);
float3 _429 = fast::normalize(_181 + MobileDirectionalLight.MobileDirectionalLight_DirectionalLightDirectionAndShadowTransition.xyz);
float _439 = (_253 * 0.25) + 0.25;
float3 _440 = cross(_206, _429);
float _442 = _253 * _253;
@ -334,7 +334,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
float3 _501 = _Globals.LightPositionAndInvRadius[_491].xyz - _179;
float _502 = dot(_501, _501);
float3 _505 = _501 * float3(rsqrt(_502));
_507 = normalize(_181 + _505);
_507 = fast::normalize(_181 + _505);
_509 = fast::max(0.0, dot(_206, _505));
_511 = fast::max(0.0, dot(_206, _507));
if (_Globals.LightColorAndFalloffExponent[_491].w == 0.0)

2
reference/shaders-ue4-no-opt/asm/frag/array-copy-error.asm.invalid.frag
View File

@ -271,7 +271,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
float4 _144 = View.View_SVPositionToTranslatedWorld * float4(gl_FragCoord.xyz, 1.0);
float3 _148 = _144.xyz / float3(_144.w);
float3 _149 = _148 - float3(View.View_PreViewTranslation);
float3 _151 = normalize(-_148);
float3 _151 = fast::normalize(-_148);
float3 _152 = _151 * float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz);
float _170 = mix(Material.Material_ScalarExpressions[0].y, Material.Material_ScalarExpressions[0].z, fast::min(fast::max(abs(dot(_151, in.in_var_TEXCOORD11_centroid.xyz)), 0.0), 1.0));
float _171 = floor(_170);

2
reference/shaders-ue4-no-opt/asm/frag/phi-variable-declaration.asm.invalid.frag
View File

@ -271,7 +271,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
float4 _144 = View.View_SVPositionToTranslatedWorld * float4(gl_FragCoord.xyz, 1.0);
float3 _148 = _144.xyz / float3(_144.w);
float3 _149 = _148 - float3(View.View_PreViewTranslation);
float3 _151 = normalize(-_148);
float3 _151 = fast::normalize(-_148);
float3 _152 = _151 * float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz);
float _170 = mix(Material.Material_ScalarExpressions[0].y, Material.Material_ScalarExpressions[0].z, fast::min(fast::max(abs(dot(_151, in.in_var_TEXCOORD11_centroid.xyz)), 0.0), 1.0));
float _171 = floor(_170);

4
reference/shaders-ue4/asm/frag/depth-compare.asm.frag
View File

@ -246,7 +246,7 @@ fragment main0_out main0(constant type_View& View [[buffer(0)]], constant type_G
float _218 = sqrt(mix(1.0, _213 * _213, _Globals.ShadowFadeFraction));
float4 _219 = _107;
_219.z = _218;
float3 _236 = normalize((SceneTexturesStruct_GBufferATexture.sample(SceneTexturesStruct_GBufferATextureSampler, _114, level(0.0)).xyz * float3(2.0)) - float3(1.0));
float3 _236 = fast::normalize((SceneTexturesStruct_GBufferATexture.sample(SceneTexturesStruct_GBufferATextureSampler, _114, level(0.0)).xyz * float3(2.0)) - float3(1.0));
uint _240 = uint(round(SceneTexturesStruct_GBufferBTexture.sample(SceneTexturesStruct_GBufferBTextureSampler, _114, level(0.0)).w * 255.0));
bool _248 = (_240 & 15u) == 5u;
float _448;
@ -259,7 +259,7 @@ fragment main0_out main0(constant type_View& View [[buffer(0)]], constant type_G
if (_160)
{
float3 _278 = _152 / float3(_158);
float3 _280 = normalize(cross(_278, float3(0.0, 0.0, 1.0)));
float3 _280 = fast::normalize(cross(_278, float3(0.0, 0.0, 1.0)));
float3 _284 = float3(_Globals.InvShadowmapResolution);
float3 _285 = _280 * _284;
float3 _286 = cross(_280, _278) * _284;

8
reference/shaders-ue4/asm/frag/global-constant-arrays.asm.frag
View File

@ -176,7 +176,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
float _644 = ((0.317398726940155029296875 + (4.25 * _Globals.WhiteTemp)) + ((4.2048167614439080352894961833954e-08 * _Globals.WhiteTemp) * _Globals.WhiteTemp)) / ((1.0 - (2.8974181986995972692966461181641e-05 * _Globals.WhiteTemp)) + ((1.6145605741257895715534687042236e-07 * _Globals.WhiteTemp) * _Globals.WhiteTemp));
float _649 = ((2.0 * _633) - (8.0 * _644)) + 4.0;
float2 _653 = float2((3.0 * _633) / _649, (2.0 * _644) / _649);
float2 _660 = normalize(float2(_633, _644));
float2 _660 = fast::normalize(float2(_633, _644));
float _665 = _633 + (((-_660.y) * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _669 = _644 + ((_660.x * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _674 = ((2.0 * _665) - (8.0 * _669)) + 4.0;
@ -273,7 +273,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_990 = 57.2957763671875 * atan2(sqrt(3.0) * (_974 - _976), ((2.0 * _973) - _974) - _976);
_990 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_974 - _976), ((2.0 * _973) - _974) - _976);
}
float _995;
if (_990 < 0.0)
@ -436,7 +436,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_2174 = 57.2957763671875 * atan2(sqrt(3.0) * (_2158 - _2160), ((2.0 * _2157) - _2158) - _2160);
_2174 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_2158 - _2160), ((2.0 * _2157) - _2158) - _2160);
}
float _2179;
if (_2174 < 0.0)
@ -909,7 +909,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1337 = 57.2957763671875 * atan2(sqrt(3.0) * (_1321 - _1323), ((2.0 * _1320) - _1321) - _1323);
_1337 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_1321 - _1323), ((2.0 * _1320) - _1321) - _1323);
}
float _1342;
if (_1337 < 0.0)

8
reference/shaders-ue4/asm/frag/padded-float-array-member-defef.asm.frag
View File

@ -178,7 +178,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
float _670 = ((0.317398726940155029296875 + (4.25 * _Globals.WhiteTemp)) + ((4.2048167614439080352894961833954e-08 * _Globals.WhiteTemp) * _Globals.WhiteTemp)) / ((1.0 - (2.8974181986995972692966461181641e-05 * _Globals.WhiteTemp)) + ((1.6145605741257895715534687042236e-07 * _Globals.WhiteTemp) * _Globals.WhiteTemp));
float _675 = ((2.0 * _659) - (8.0 * _670)) + 4.0;
float2 _679 = float2((3.0 * _659) / _675, (2.0 * _670) / _675);
float2 _686 = normalize(float2(_659, _670));
float2 _686 = fast::normalize(float2(_659, _670));
float _691 = _659 + (((-_686.y) * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _695 = _670 + ((_686.x * _Globals.WhiteTint) * 0.0500000007450580596923828125);
float _700 = ((2.0 * _691) - (8.0 * _695)) + 4.0;
@ -275,7 +275,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1016 = 57.2957763671875 * atan2(sqrt(3.0) * (_1000 - _1002), ((2.0 * _999) - _1000) - _1002);
_1016 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_1000 - _1002), ((2.0 * _999) - _1000) - _1002);
}
float _1021;
if (_1016 < 0.0)
@ -485,7 +485,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_2278 = 57.2957763671875 * atan2(sqrt(3.0) * (_2262 - _2264), ((2.0 * _2261) - _2262) - _2264);
_2278 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_2262 - _2264), ((2.0 * _2261) - _2262) - _2264);
}
float _2283;
if (_2278 < 0.0)
@ -958,7 +958,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_Globals& _Globa
}
else
{
_1443 = 57.2957763671875 * atan2(sqrt(3.0) * (_1427 - _1429), ((2.0 * _1426) - _1427) - _1429);
_1443 = 57.2957763671875 * precise::atan2(sqrt(3.0) * (_1427 - _1429), ((2.0 * _1426) - _1427) - _1429);
}
float _1448;
if (_1443 < 0.0)

2
reference/shaders-ue4/asm/frag/sample-mask-not-array.asm.frag
View File

@ -462,7 +462,7 @@ fragment main0_out main0(main0_in in [[stage_in]], constant type_View& View [[bu
{
_215 = in.in_var_TEXCOORD7;
}
float3 _216 = fast::max(Material.Material_VectorExpressions[1].xyz * float3(((1.0 + dot(float3(-1.0, -1.5, 3.0) / float3(sqrt(12.25)), normalize(float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz) * normalize((float3(0.0, 0.0, 1.0) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz)))) * 0.5) + 0.20000000298023223876953125), float3(0.0));
float3 _216 = fast::max(Material.Material_VectorExpressions[1].xyz * float3(((1.0 + dot(float3(-1.0, -1.5, 3.0) / float3(sqrt(12.25)), fast::normalize(float3x3(in.in_var_TEXCOORD10_centroid.xyz, cross(in.in_var_TEXCOORD11_centroid.xyz, in.in_var_TEXCOORD10_centroid.xyz) * float3(in.in_var_TEXCOORD11_centroid.w), in.in_var_TEXCOORD11_centroid.xyz) * fast::normalize((float3(0.0, 0.0, 1.0) * float3(View.View_NormalOverrideParameter.w)) + View.View_NormalOverrideParameter.xyz)))) * 0.5) + 0.20000000298023223876953125), float3(0.0));
float3 _246;
if (View.View_OutOfBoundsMask > 0.0)
{

6
reference/shaders-ue4/asm/vert/texture-buffer.asm.vert
View File

@ -296,7 +296,7 @@ vertex main0_out main0(main0_in in [[stage_in]], constant type_View& View [[buff
float4 _145 = AttributesTexture.sample(AttributesTextureSampler, _133, level(0.0));
float _146 = _137.w;
float3 _158 = float3x3(Primitive.Primitive_LocalToWorld[0].xyz, Primitive.Primitive_LocalToWorld[1].xyz, Primitive.Primitive_LocalToWorld[2].xyz) * VelocityTexture.sample(VelocityTextureSampler, _133, level(0.0)).xyz;
float3 _160 = normalize(_158 + float3(0.0, 0.0, 9.9999997473787516355514526367188e-05));
float3 _160 = fast::normalize(_158 + float3(0.0, 0.0, 9.9999997473787516355514526367188e-05));
float2 _204 = ((((_145.xy + float2((_145.x < 0.5) ? 0.0 : (-0.5), (_145.y < 0.5) ? 0.0 : (-0.5))) * float2(2.0)) * (((CurveTexture.sample(CurveTextureSampler, (EmitterUniforms.EmitterUniforms_MiscCurve.xy + (EmitterUniforms.EmitterUniforms_MiscCurve.zw * float2(_146))), level(0.0)) * EmitterUniforms.EmitterUniforms_MiscScale) + EmitterUniforms.EmitterUniforms_MiscBias).xy * EmitterDynamicUniforms.EmitterDynamicUniforms_LocalToWorldScale)) * fast::min(fast::max(EmitterUniforms.EmitterUniforms_SizeBySpeed.xy * float2(length(_158)), float2(1.0)), EmitterUniforms.EmitterUniforms_SizeBySpeed.zw)) * float2(step(_146, 1.0));
float3 _239 = float4((((Primitive.Primitive_LocalToWorld[0u].xyz * _137.xxx) + (Primitive.Primitive_LocalToWorld[1u].xyz * _137.yyy)) + (Primitive.Primitive_LocalToWorld[2u].xyz * _137.zzz)) + (Primitive.Primitive_LocalToWorld[3u].xyz + float3(View.View_PreViewTranslation)), 1.0).xyz;
float3 _242 = float3(EmitterUniforms.EmitterUniforms_RemoveHMDRoll);
@ -312,8 +312,8 @@ vertex main0_out main0(main0_in in [[stage_in]], constant type_View& View [[buff
float3 _279 = cross(_265, float3(0.0, 0.0, 1.0));
float3 _284 = _279 / float3(sqrt(fast::max(dot(_279, _279), 0.00999999977648258209228515625)));
float3 _286 = float3(fast::clamp((_261 * EmitterUniforms.EmitterUniforms_CameraFacingBlend[1]) - EmitterUniforms.EmitterUniforms_CameraFacingBlend[2], 0.0, 1.0));
_335 = normalize(mix(_251, _284, _286));
_336 = normalize(mix(_259, cross(_265, _284), _286));
_335 = fast::normalize(mix(_251, _284, _286));
_336 = fast::normalize(mix(_259, cross(_265, _284), _286));
}
else
{

28
spirv_msl.cpp
View File

@ -5719,8 +5719,9 @@ void CompilerMSL::emit_custom_functions()
case SPVFuncImplReflectScalar:
// Metal does not support scalar versions of these functions.
// Ensure fast-math is disabled to match Vulkan results.
statement("template<typename T>");
statement("inline T spvReflect(T i, T n)");
statement("[[clang::optnone]] T spvReflect(T i, T n)");
begin_scope();
statement("return i - T(2) * i * n * n;");
end_scope();
@ -8781,7 +8782,8 @@ const char *CompilerMSL::get_memory_order(uint32_t)
return "memory_order_relaxed";
}
// Override for MSL-specific extension syntax instructions
// Override for MSL-specific extension syntax instructions.
// In some cases, deliberately select either the fast or precise versions of the MSL functions to match Vulkan math precision results.
void CompilerMSL::emit_glsl_op(uint32_t result_type, uint32_t id, uint32_t eop, const uint32_t *args, uint32_t count)
{
auto op = static_cast<GLSLstd450>(eop);
@ -8793,8 +8795,17 @@ void CompilerMSL::emit_glsl_op(uint32_t result_type, uint32_t id, uint32_t eop,
switch (op)
{
case GLSLstd450Sinh:
emit_unary_func_op(result_type, id, args[0], "fast::sinh");
break;
case GLSLstd450Cosh:
emit_unary_func_op(result_type, id, args[0], "fast::cosh");
break;
case GLSLstd450Tanh:
emit_unary_func_op(result_type, id, args[0], "precise::tanh");
break;
case GLSLstd450Atan2:
emit_binary_func_op(result_type, id, args[0], args[1], "atan2");
emit_binary_func_op(result_type, id, args[0], args[1], "precise::atan2");
break;
case GLSLstd450InverseSqrt:
emit_unary_func_op(result_type, id, args[0], "rsqrt");
@ -9018,25 +9029,20 @@ void CompilerMSL::emit_glsl_op(uint32_t result_type, uint32_t id, uint32_t eop,
break;
case GLSLstd450Length:
// MSL does not support scalar versions here.
// MSL does not support scalar versions, so use abs().
if (expression_type(args[0]).vecsize == 1)
{
// Equivalent to abs().
emit_unary_func_op(result_type, id, args[0], "abs");
}
else
CompilerGLSL::emit_glsl_op(result_type, id, eop, args, count);
break;
case GLSLstd450Normalize:
// MSL does not support scalar versions here.
// Returns -1 or 1 for valid input, sign() does the job.
if (expression_type(args[0]).vecsize == 1)
{
// Returns -1 or 1 for valid input, sign() does the job.
emit_unary_func_op(result_type, id, args[0], "sign");
}
else
CompilerGLSL::emit_glsl_op(result_type, id, eop, args, count);
emit_unary_func_op(result_type, id, args[0], "fast::normalize");
break;
case GLSLstd450Reflect: