Optimizing Complex Matrix Multiplication using Karatsuba’s Algorithm (#2220)

* Implementing Complex Matmul using Karatsuba Algorithm

* Implemented Karatsuba's Algorithm for complex matmul and pre-commit them

* fix

---------

Co-authored-by: Awni Hannun <awni@apple.com>
This commit is contained in:
Suryash Malviya
2025-06-02 18:58:46 -04:00
committed by GitHub
parent cbad6c3093
commit 0408ba0a76
2 changed files with 24 additions and 15 deletions

View File

@@ -2862,21 +2862,30 @@ array matmul(
<< " second input with shape " << b.shape() << ".";
throw std::invalid_argument(msg.str());
}
// Type promotion
auto out_type = promote_types(a.dtype(), b.dtype());
// Complex matmul in terms of real matmuls
if (out_type == complex64) {
// complex matmul using Karatsuba's Algorithm
if (a.dtype() == complex64 || b.dtype() == complex64) {
// Extract real and imaginary parts
auto a_real = real(a, s);
auto b_real = real(b, s);
auto a_imag = imag(a, s);
auto b_real = real(b, s);
auto b_imag = imag(b, s);
auto c_real =
subtract(matmul(a_real, b_real, s), matmul(a_imag, b_imag, s), s);
auto c_imag = add(matmul(a_real, b_imag, s), matmul(a_imag, b_real, s), s);
// Compute real and imaginary components of the result
auto m1 = matmul(a_real, b_real, s);
auto m2 = matmul(a_imag, b_imag, s);
auto m3 = matmul(add(a_real, a_imag, s), add(b_real, b_imag, s), s);
auto c_real = subtract(m1, m2, s);
auto c_imag = subtract(m3, add(m1, m2, s), s);
return add(
c_real, multiply(array(complex64_t{0, 1}, complex64), c_imag, s), s);
}
// Type promotion
auto out_type = promote_types(a.dtype(), b.dtype());
if (!issubdtype(out_type, floating)) {
std::ostringstream msg;
msg << "[matmul] Only real floating point types are supported but "