tiny.simd.perBlock2x2MatMul

1pub fn perBlock2x2MatMul(2    comptime D: type,3    a: anytype,4    b: @TypeOf(a),5    c: D.Vector,6) D.Vector {7    validateDescriptor(D);8    if (comptime D.Lane == i32) {9        validateVector(@TypeOf(a), i8, D.lane_count * 4);10        return int8MatMul(D, a, b, c);11    }12    if (comptime D.Lane == f32) {13        validateVector(@TypeOf(a), u16, D.lane_count * 2);14        return bfloatMatMul(D, a, b, c);15    }16    @compileError("2x2 block matrix multiplication requires i32 or f32 accumulators");17}

1pub const perBlock2x2MatMul = matmul.perBlock2x2MatMul;