[SM6.10] Thread scope Multiply output K Vec

[xiaolin-ji]

Description
From

DirectXShaderCompiler/tools/clang/lib/Headers/hlsl/dx/linalg.h

Line 468 in 8cb7673

typename hlsl::enable_if<hlsl::is_arithmetic<InputElTy>::value, vector<OutputElTy, K> >::type

typename hlsl::enable_if<hlsl::is_arithmetic::value, vector<OutputElTy, K> >::type
// clang-format on
Multiply(Matrix<MatrixDT, M, K, MatrixUse::A, MatrixScope::Thread> MatrixA,
vector<InputElTy, K> Vec) {
vector<OutputElTy, M> Result;
__builtin_LinAlg_MatrixVectorMultiply(Result, MatrixA.__handle,
hlsl::is_signed::value, Vec,
MatrixDT);
return Result;
}

the output vector length is M or K?

Steps to Reproduce

Actual Behavior

Environment

• DXC version
• Host Operating System <!--- Host operating system and version --->

[xiaolin-ji]

The input is [MxK] matrix A, it is left multiply of K vector, the output should be M vector

[tex3d]

Good catch! Note that the bias vector size is also wrong here, along with the output vector you pointed out (and in the spec):

DirectXShaderCompiler/tools/clang/lib/Headers/hlsl/dx/linalg.h

Lines 479 to 491 in 475e3cc

	template <typename OutputElTy, typename InputElTy, typename BiasElTy,
	SIZE_TYPE M, SIZE_TYPE K, ComponentEnum MatrixDT>
	// clang-format off
	typename hlsl::enable_if<hlsl::is_arithmetic<InputElTy>::value, vector<OutputElTy, K> >::type
	// clang-format on
	MultiplyAdd(Matrix<MatrixDT, M, K, MatrixUse::A, MatrixScope::Thread> MatrixA,
	vector<InputElTy, M> Vec, vector<BiasElTy, K> Bias) {
	vector<OutputElTy, K> Result;
	__builtin_LinAlg_MatrixVectorMultiplyAdd(Result, MatrixA.__handle,
	hlsl::is_signed<OutputElTy>::value,
	Vec, MatrixDT, Bias, MatrixDT);
	return Result;
	}

It should be based on M instead of K.

[hekota]

Good catch! Note that the bias vector size is also wrong here, along with the output vector you pointed out.

It looks like most of the vector dimensions are mixed up. Sorry about that! I'll put up a fix for the spec and the code.