No asserts emitted for strides based on other dims extent.

[mcourteaux]

When promising dense data:

      int bsx = 16; // block size x
      output.dim(0).extent() = output.dim(0).extent() / bsx * bsx;
      output.dim(1).stride() = output.dim(0).extent();
      output.dim(2).stride() = output.dim(0).stride() * output.dim(1).extent();

Not a single assert is emitted that checks this.

I validated that this both happens on main and on @alexreinking/fix-extent-alignment

[alexreinking]

Try this:

int bsx = 16; // block size x
output.dim(0).set_extent(output.dim(0).extent() / bsx * bsx);
output.dim(1).set_stride(output.dim(0).extent());
output.dim(2).set_stride(output.dim(0).stride() * output.dim(1).extent());

[mcourteaux]

Dang. 💀 These should probably return const & or const to avoid usage bugs like this.

[mcourteaux]

Yes that does emit the asserts for the first two:

assert(((output$2.extent.0 % 16) == 0), halide_error_constraint_violated("output$2.extent.0", output$2.extent.0, "((output$2.extent.0/16)*16)", ((output$2.extent.0 / 16) * 16)))
assert((output$2.extent.0 == output$2.stride.1), halide_error_constraint_violated("output$2.stride.1", output$2.stride.1, "output$2.extent.0", output$2.extent.0))

However, the third one, regarding output.stride.2 is missing.

On a side-note: I do spot this weird let:

let output$2.total_extent.1 = (int64(output$2.stride.2) * int64(output$2.stride.1))

I cannot think of anything where multiplying strides together makes any sense. Comparing the same Let, but with add_requirements() as described in #9328 generates a reasonable Let:

let output$2.total_extent.1 = (int64(output$2.extent.1) * int64(output$2.extent.0))

[alexreinking]

On a side-note: I do spot this weird let:

let output$2.total_extent.1 = (int64(output$2.stride.2) * int64(output$2.stride.1))

Well, you did say that output$2.stride.1 and output$2.extent.0 are equal. 🙂

[mcourteaux]

Yes, but I didn't specify that stride.2 equals extent.1. That doesn't make any sense.