Implement the `fma` HLSL Function

[farzonl]

• Implement fma clang builtin,
• Link fma clang builtin with hlsl_alias_intrinsics.h
• Add sema checks for fma to CheckHLSLBuiltinFunctionCall in SemaHLSL.cpp
• Add codegen for fma to EmitHLSLBuiltinExpr in CGHLSLBuiltins.cpp
• Add codegen tests to clang/test/CodeGenHLSL/builtins/fma.hlsl
• Add sema tests to clang/test/SemaHLSL/BuiltIns/fma-errors.hlsl
• Create the int_dx_fma intrinsic in IntrinsicsDirectX.td
• Create the DXILOpMapping of int_dx_fma to 47 in DXIL.td
• Create the fma.ll and fma_errors.ll tests in llvm/test/CodeGen/DirectX/
• Create the int_spv_fma intrinsic in IntrinsicsSPIRV.td
• In SPIRVInstructionSelector.cpp create the fma lowering and map it to int_spv_fma in SPIRVInstructionSelector::selectIntrinsic.
• Create SPIR-V backend test case in llvm/test/CodeGen/SPIRV/hlsl-intrinsics/fma.ll
• Set the DX11_1_DoubleExtensions shader flag to true in DXILShaderFlags.cpp when fma is used
• Add an fma case to the shader flag test llvm/test/CodeGen/DirectX/ShaderFlags/double-extensions.ll

DirectX

DXIL Opcode	DXIL OpName	Shader Model	Shader Stages
47	Fma	6.0	()

SPIR-V

Fma:

Description:

Fma

Computes a * b + c. In uses where this operation is decorated
with NoContraction:

• fma is considered a single operation, whereas the expression a
  * b + c is considered two operations.

- The precision of fma can differ from the precision of the
expression a * b + c.

- fma will be computed with the same precision as any other fma
decorated with NoContraction, giving invariant results for the same
input values of a, b, and c.

Otherwise, in the absence of a NoContraction decoration, there are
no special constraints on the number of operations or difference in
precision between fma and the expression a * b +c.

The operands must all be a scalar or vector whose component type is
floating-point.

Result Type and the type of all operands must be the same type.
Results are computed per component.

Number	Operand 1	Operand 2	Operand 3	Operand 4
50	<id> a	<id> b	<id> c

Test Case(s)

Example 1

//dxc fma_test.hlsl -T lib_6_8 -enable-16bit-types -O0

export double4 fn(double4 p1, double4 p2, double4 p3) {
    return fma(p1, p2, p3);
}

HLSL:

Returns the double-precision fused multiply-addition of a * b + c.

ret fma(double a, b, c);

Parameters

a

[in] The first value in the fused multiply-addition.

b

[in] The second value in the fused multiply-addition.

c

[in] The third value in the fused multiply-addition.

Return Value

The double-precision fused multiply-addition of parameters a * b + c. The returned value must be accurate to 0.5 units of least precision (ULP).

Remarks

The fma intrinsic must support NaNs, INFs, and Denorms.

To use the fma intrinsic in your shader code, call the ID3D11Device::CheckFeatureSupport method with D3D11_FEATURE_D3D11_OPTIONS to verify that the Direct3D device supports the ExtendedDoublesShaderInstructions feature option. The fma intrinsic requires a WDDM 1.2 display driver, and all WDDM 1.2 display drivers must support fma. If your app creates a rendering device with feature level 11.0 or 11.1 and the compilation target is shader model 5 or later, the HLSL source code can use the fma intrinsic.

Type Description

Name	Template Type	Component Type	Size
a	scalar, vector, or matrix	double	any
b	same as input a	double	same dimensions as input a
c	same as input a	double	same dimensions as input a
ret	same as input a	double	same dimensions as input a

Minimum Shader Model

This function is supported in the following shader models.

Shader Model	Supported
Shader model 5 or later	yes

Requirements

Requirement	Value
Minimum supported client	Windows 8 [desktop apps | UWP apps]
Minimum supported server	Windows Server 2012 [desktop apps | UWP apps]
Header	Corecrt_math.h

See also

Intrinsic Functions