mlir

We are missing an optimization to build reduction ops when possible. Consider the following which is doing b[0] | b[1]:

module {
  hw.module @Foo(%a: i1, %b: i2) -> (c: i1) {
    %0 = comb.extract %b from 1 : (i2) -> i1
    %1 = comb.extract %b from 0 : (i2) -> i1
    %2 = comb.or %0, %1 {sv.namehint = "_b"} : i1
    hw.output %2 : i1
  }
}

This produces:

The existing code assumes the result tensor type is the same as input type for a few aten ops like log, exp, erf. See https://github.com/llvm/torch-mlir/blob/486f95e84f587d020ba789b071b12f890510f1a1/lib/Dialect/Torch/Transforms/RefineTypes.cpp#L221-L235
This incorrect. The result tensor of these ops should always have the default dtype rather than the same as the input type. E2E tests fo

@honglh added optimized kernels for ARM32 in #432. It would be great if we could add support for bitpacked activations to them to match the AArch64 optimized bgemm kernels.

In test/create-cores/test_dma1.mlir, -aie-lower-memcpy convert

  AIE.memcpy @token0(1, 2) (%t11 : <%buf0, 0, 256>, %t22 : <%buf1, 0, 256>) : (memref<256xi32>, memref<256xi32>)
  AIE.memcpy @token1(1, 2) (%t11 : <%buf0, 0, 256>, %t33 : <%buf2, 0, 256>) : (memref<256xi32>, memref<256xi32>)

to (only shows the %t11 side)

  %2 = AIE.mem(%0) {
    %15 = AIE.dmaStart(MM2S0, ^bb1