Generic region to loops

[nsmithtt]

This change adds a few new passes:

linalg.generic to affine passes

Simply calls the upstream pass for converting a linalg.generic into an affine loop nest.

Lower affine pass

Again, just uses an upstream pass for converting affine into SCF and arith.

Linearize memref accesses pass

A custom TTIR pass that takes a nested loop structure over n-dimensional memrefs and linearizes them into a single dimension. This is a useful because circular buffers in metal are only one-dimensional.

Example, this pass will convert the following code:

  affine.for %arg5 = 0 to 2 {        
    affine.for %arg6 = 0 to 4 {        
      %0 = affine.load %arg2[%arg5, %arg6] : memref<2x4x!tt.tile<32x32, f32>, #l1_>        
      %1 = affine.load %arg3[%arg5, %arg6] : memref<2x4x!tt.tile<32x32, f32>, #l1_>        
      %2 = "ttir.tile_maximum"(%0, %1) : (!tt.tile<32x32, f32>, !tt.tile<32x32, f32>) -> !tt.tile<32x32, f32>        
      affine.store %2, %arg4[%arg5, %arg6] : memref<2x4x!tt.tile<32x32, f32>, #l1_>                                                       
    }        
  }                                                                              

Into:

  %collapse_shape = memref.collapse_shape %arg2 [[0, 1]] : memref<2x4x!tt.tile<32x32, f32>, #l1_> into memref<8x!tt.tile<32x32, f32>, #l1_>        
  %collapse_shape_0 = memref.collapse_shape %arg3 [[0, 1]] : memref<2x4x!tt.tile<32x32, f32>, #l1_> into memref<8x!tt.tile<32x32, f32>, #l1_>        
  %collapse_shape_1 = memref.collapse_shape %arg4 [[0, 1]] : memref<2x4x!tt.tile<32x32, f32>, #l1_> into memref<8x!tt.tile<32x32, f32>, #l1_>        
  affine.for %arg5 = 0 to 2 {                                    
    affine.for %arg6 = 0 to 4 {                                                       
      %0 = affine.load %collapse_shape[%arg5 * 4 + %arg6] : memref<8x!tt.tile<32x32, f32>, #l1_>        
      %1 = affine.load %collapse_shape_0[%arg5 * 4 + %arg6] : memref<8x!tt.tile<32x32, f32>, #l1_>                                                                         
      %2 = "ttir.tile_maximum"(%0, %1) : (!tt.tile<32x32, f32>, !tt.tile<32x32, f32>) -> !tt.tile<32x32, f32>        
      affine.store %2, %collapse_shape_1[%arg5 * 4 + %arg6] : memref<8x!tt.tile<32x32, f32>, #l1_>        
    }                                                         
  }                                     

Closes #1910
Closes #1911

[github-actions]

⚠️ Clang-Tidy found issue(s) with the introduced code (1/1)

[github-actions]

⚠️ llvm-qualified-auto ⚠️
auto user can be declared as auto *user

Suggested change

	for (auto user : linearizedArg->getUsers()) {
	for (auto *user : linearizedArg->getUsers()) {

[azecevicTT]

Great pass documentation!