Note that this is more or less a continuation of the Sparse Representation discussion, but now focused specifically on the progress of MLIR support for sparse tensors. In a distant past, I pioneered the idea of letting a compiler automatically convert annotated dense linear algebra Fortran to seman…

Before actual MLIR code, I found it easier to get the code generation structure right by simply emitting pseudo-code on the debugging output. The kernel: x(i) = a(i) + b(i) * c(i) looks as generic Linalg operation as follows: func @generic_op_vec1d(%arga: tensor<?xf32>, %a…

I am still very enthusiastic, with lots of progress this week! Most part of the actual MLIR codegen is done now too, pending while-loops for true co-iteration. But anything expressible with for-loops works already. Given a kernel “x(i) = a(i) * b(i)”, expressed in linalg as follows, and b annotated …

Thanks to @ftynse , we now have while-loops in the SCF dialect as well, which greatly sped up development of the sparse compiler part! I expect to send out the first CL any day now. Given a kernel like x(i) = a(i) + b(i) with both a and b annotated as sparse vectors, MLIR now lowers the linalg ope…

For those interested, the first CL is out for review ( https://reviews.llvm.org/D90994 ). Note that a lot still needs to be done. Running and testing the generated code is next, then measuring and improving the performance. Also, proliferating the sparse bufferization needs to be addressed.

The first prototype CL has landed, so it was time for a quick sanity check. Starting with a “dense” Linalg expression for column-wise matrix x vector, I let MLIR automatically generate sparse code. Then I replaced the “localized buffers” with parameters of a surrounding method and hand-coded initial…

A pending CL ( https://reviews.llvm.org/D91978 ) adds some rudimentary parallelization strategies for the generated for-loops to the sparse compiler (outer/all, dense-only/all). This work will combine nicely with the work @ezhulenev is doing on executing parallel loops with a thread pool on CPU.

And right before the break, the sparse compiler also has a vectorization strategy, at least for the innermost for-loops. #trait_s = { indexing_maps = [ affine_map<(i) -> (i)>, // a affine_map<(i) -> (i)> // x (out) ], sparse = [ [ "S" ], // a [ "D" ] // x ], iterato…

[image] aartbik: And right before the break, the sparse compiler also has a vectorization strategy, at least for the innermost for-loops. Nice :), happy holidays!

For who is interested, the vectorization strategy in the sparse compiler is now available for review. The generated masked memory operations interact nicely with subsequent vector dialect optimizations, such as folding constant masks. Note, however, for masks that remain, we will have to add more lo…

Note that the strategies can also be combined. For example, a kernel like "“x(i) = SUM A(i,j) * B(i,j)”, with A’s innermost dimension marked as sparse can yield the following parallel-vector-reduction loop nest: scf.parallel (%arg3) = (%c0) to (%c512) step (%c1) { %6 = load %0[%arg3] : memref<?xi…

MLIR Support for Sparse Tensors

MLIR

nicolasvasilache December 18, 2020, 10:43am 10

Nice :), happy holidays!

MLIR News, 25th edition (1/22/2021)

Topic		Replies	Views
Sparse Representation MLIR	46	5386	December 3, 2020
MLIR Sparse Compiler Progress MLIR	24	5469	February 9, 2026
Sparse Tensors in MLIR MLIR	62	7434	March 25, 2025
Sparse Compiler and GPU Code-Generation MLIR	8	3106	October 20, 2023
Paper: Compiler Support for Sparse Tensor Computations in MLIR Announcements	0	519	February 10, 2022

MLIR Support for Sparse Tensors

Related topics