taco

taco copied to clipboard

Depending on the iteration space is partitioned, the GPU backend can emit CUDA kernel launches with a grid size of 0, which results in a cuda error. This is an...

rohany

Consider the code generated by http://tensor-compiler.org/codegen.html?expr=a(i)%20=%20b(i)%20*%20c(i)&format=a:d:0;b:s:0;c:s:0&sched=split:i:i0:i1:32. The operations: ``` int32_t ib0 = b1_crd[ib]; int32_t ic0 = c1_crd[ic]; int32_t i = TACO_MIN(ib0,ic0); int32_t i1 = i - i0 * 32; int32_t...

rohany

Assign new values to tensor coordinates

2

@stephenchouca Hi Stephen, I was looking for a way to assign a new value to a specific tensor coordinate after initialization of the tensor. Assume the tensor is A, and...

kakulo

TACO currently treats transposition as a user-space operation, but an important first step to making things more automatic is to add a construct to transpose on compute(). https://github.com/tensor-compiler/taco/issues/496 addresses allocation...

willow-ahrens

Scheduling commands require workspace tensors as inputs. For more automatic scheduling, it would be great if TACO could automatically allocate workspace tensors that show up in the concrete index notation.

willow-ahrens

This is a feature request. I would like to reduce runtime costs of calling TACO by caching the generated code in a file, and checking for that file the next...

willow-ahrens

Direct concrete index notation schedules breaks tensor serialization order

1

When specifying a schedule via direct concrete index notation, TACO assumes that the postorder traversal of tensors in the index expression will match the postorder traversal of tensors in the...

willow-ahrens

taco-tool: the benchmark utility of the taco tool no longer works

4

If you attempt to benchmark programs with the taco-tool, i.e. specifying formats, data and `-time=5`, the program fails with a generated code compilation error. pinging @Infinoid because I think you...

rohany

RFC: Tensor API

3

This issue presents a draft for the new API for Tensor objects in taco. This new API targets to address a number of issues with the current API, as well...

oicirtap

Reordering

1

Is there a way to convert an existing tensor to a new tensor with the same content, but with the modes ordered differently? This would be a generalization of matrix...

jcrbloch