Blit directly from GPU object to texture

[athas]

We need to find a way to avoid the copy to CPU memory.

[athas]

I tried disabling the roundtrip entirely (thereby disabling rendering), and for some reason it did not become that much faster. Maybe this is not the bottleneck we think it is.

[nqpz]

The line

FUT_CHECK(ctx->fut, futhark_values_i32_2d(ctx->fut, out_arr, ctx->data));

takes ~1.5 ms on my laptop for every frame. Wouldn't somehow blitting the cl_mem directly get rid of this overhead, or am I missing some details?

[athas]

It would, but 1.5ms is not a lot, is all I'm saying.

[nqpz]

Well... 1.5 ms is not enough for me to do anything about it, but it would still be nice to get rid of it.