Commits on Oct 11, 2024

1. Before, IR like …

   ```
   module attributes {
     llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128",
     llvm.target_triple = "x86_64-unknown-linux-gnu"} {
     func.func @forward(%arg0: tensor<1x255x40x40xi8> {func.orig_type = tensor<1x255x40x40xi8>, onnx.name = "inp3in"} loc(unknown), %arg1: tensor<1x255x20x20xi8> {func.orig_type = tensor<1x255x20x20xi8>, onnx.name = "inp2in"} loc(unknown), %arg2: tensor<1x255x80x80xi8> {func.orig_type = tensor<1x255x80x80xi8>, onnx.name = "inp1in"} loc(unknown)) -> (tensor<1x3x40x40x85xf32> {func.orig_type = tensor<1x3x40x40x85xf32>, onnx.name = "onnx::Sigmoid_564"}) {
       %cst = arith.constant dense<3.100000e+00> : tensor<1x1x1x1xf32> loc(#loc)
       %0 = tensor.empty() : tensor<1x255x40x40xf32> loc(#loc1)
       %1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x255x40x40xi8>) outs(%0 : tensor<1x255x40x40xf32>) {
       ^bb0(%in: i8 loc(unknown), %out: f32 loc("")):
         %5 = arith.sitofp %in : i8 to f32 loc(#loc1)
         linalg.yield %5 : f32 loc(#loc1)
       } -> tensor<1x255x40x40xf32> loc(#loc1)
       %2 = tensor.empty() : tensor<1x255x40x40xf32> loc(#loc1)
       %3 = linalg.generic {indexing_maps = [#map, #map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1, %cst : tensor<1x255x40x40xf32>, tensor<1x1x1x1xf32>) outs(%2 : tensor<1x255x40x40xf32>) {
       ^bb0(%in: f32 loc(""), %in_0: f32 loc(unknown), %out: f32 loc("")):
         %5 = arith.mulf %in, %in_0 : f32 loc(#loc1)
         linalg.yield %5 : f32 loc(#loc1)
       } -> tensor<1x255x40x40xf32> loc(#loc1)
       %expanded = tensor.expand_shape %3 [[0], [1, 2], [3], [4]] output_shape [1, 3, 85, 40, 40] : tensor<1x255x40x40xf32> into tensor<1x3x85x40x40xf32> loc(#loc2)
       %4 = tensor.empty() : tensor<1x3x40x40x85xf32> loc(#loc3)
       %transposed = linalg.transpose ins(%expanded : tensor<1x3x85x40x40xf32>) outs(%4 : tensor<1x3x40x40x85xf32>) permutation = [0, 1, 3, 4, 2]  loc(#loc3)
       return %transposed : tensor<1x3x40x40x85xf32> loc(#loc)
     } loc(#loc)
   } loc(#loc)
   ```
   
   would not move the tensor.expand_shape to the top because the check `isProjectedPermutation` on `#map`
   would fail. It passes if changing `#map` to the equivalent `affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>` (considering that the tensor size is 1 in that dimension).
   From my understanding, the transformation doesn't strictly require latter map,
   and there is already an option in isProjectedPermutation to allow zeros.

   

   mgehre-amd committed Oct 11, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for 86ab4d3
   • Browse repository at this point

   Copy the full SHA

   86ab4d3 View commit details

   Browse the repository at this point in the history