4 ms·

"Idiomatic" depends on language and domain and era and purpose. Consider the dgemm method in BLAS (double general matrix multiply, powers all sorts of modern co

"Idiomatic" depends on language and domain and era and purpose. Consider the dgemm method in BLAS (double general matrix multiply, powers all sorts of modern code like numpy). It's such a workhorse, used by everything under the sun. But it does one thing. It multiplies two matrices of doubles. You're going to want to optimize the hell out of it and ensure it's correct, then never revisit it. It's been around (in the exact same form as far as I can tell) nearly as long as I've been alive. The right, idiomatic code for that is going to look different from the right, idiomatic code for something at the business logic level that changes every 6 months. It does it a disservice to shit on it like you are.

aswanson 7y ago
You're going to want to optimize the hell out of it and ensure it's correct, then never revisit it. Balderdash. Claptrap. Codswallop. Poppycock. What if you want to revisit it to make it run on parallel processors? Optimize it for a new CUDA architecture or caching scheme? A new instruction set that handles sparse matrices better? Code lives forever, at every level.
- 6gvONxR4sf7o 7y ago
  For this kind of linear algebra, you're going to have a different routine for a sparse matrix (dcoomm) or a symmetric one (dsymm) or a triangular one (dtrmm) or symmetric banded matrix vector product (dsbmv) or whatever. Hopefully those are separate functions from your dense one (dgemm), which is what I'm talking about. You'll also have a different function for CUDA (which exists in cuBLAS) than for multi-core (in ScaLAPACK? PLASMA?) than for single-CPU (in BLAS), which is what I'm talking about. General matrix multiply in this sense isn't "all the different matrix multiplies." It's a function with a very specific purpose, and yeah you don't really need to revisit it every decade. Either way, it's not that much horror: http://www.netlib.org/lapack/lapack-3.1.1/html/dgemm.f.html http://www.netlib.org/lapack/lapack-3.1.1/html/dgemm.f.html

aswanson 7y ago

You're going to want to optimize the hell out of it and ensure it's correct, then never revisit it. Balderdash. Claptrap. Codswallop. Poppycock. What if you want to revisit it to make it run on parallel processors? Optimize it for a new CUDA architecture or caching scheme? A new instruction set that handles sparse matrices better? Code lives forever, at every level.

6gvONxR4sf7o 7y ago
For this kind of linear algebra, you're going to have a different routine for a sparse matrix (dcoomm) or a symmetric one (dsymm) or a triangular one (dtrmm) or symmetric banded matrix vector product (dsbmv) or whatever. Hopefully those are separate functions from your dense one (dgemm), which is what I'm talking about. You'll also have a different function for CUDA (which exists in cuBLAS) than for multi-core (in ScaLAPACK? PLASMA?) than for single-CPU (in BLAS), which is what I'm talking about. General matrix multiply in this sense isn't "all the different matrix multiplies." It's a function with a very specific purpose, and yeah you don't really need to revisit it every decade. Either way, it's not that much horror: http://www.netlib.org/lapack/lapack-3.1.1/html/dgemm.f.html http://www.netlib.org/lapack/lapack-3.1.1/html/dgemm.f.html