CUTLASS 2.6.1 - functional and performance enhancements to strided DGRAD, fixes, and tuning

* cutlass 2.6 update

* remove debug prints

* cutlass 2.6.1 (minor update)

* Updated CHANGELOG.

* Minor edit to readme to indicate patch version.

* Minor edit to readme.

Co-authored-by:  Haicheng Wu <haichengw@nvidia.com>, Andrew Kerr <akerr@nvidia.com>

include/cutlass/conv/threadblock/conv2d_dgrad_output_gradient_tile_access_iterator_analytic.h

@@ -130,7 +130,6 @@ private:
   int offset_p_[ThreadMap::Iterations::kStrided];
   int offset_q_[ThreadMap::Iterations::kStrided];
 
-
 public:
 
   CUTLASS_HOST_DEVICE
@@ -139,6 +138,7 @@ public:
     Conv2dProblemSize const &problem_size,
     Element const *ptr,
     int thread_idx,
+    FastDivmod const &stride_h_divmod, FastDivmod const &stride_w_divmod,
     int start_r, int start_s,
     MatrixCoord const &threadblock_offset = MatrixCoord()     // threadblock offset - units are whole CTA tiles
   ):
@@ -164,9 +164,12 @@ public:
     }
 
     // Starting h, w positions for filter position in gemm_k=0
-    int start_h = std::abs((problem_size_.pad_h - filter_r) % problem_size_.stride_h);
-    int start_w = std::abs((problem_size_.pad_w - filter_s) % problem_size_.stride_w);
-
+    int start_h, start_w;
+    strided_dgrad_starting_coords(
+      problem_size_, 
+      stride_h_divmod, stride_w_divmod, 
+      filter_r, filter_s, 
+      start_h, start_w);
 
     // Effective P and Q for filter position required for remapping NHW rows
     int P = (problem_size_.H - start_h + problem_size_.stride_h - 1) / problem_size_.stride_h;

include/cutlass/conv/threadblock/conv2d_tile_iterator.h

@@ -200,7 +200,27 @@ private:
 
 public:
 
-  /// Constructor
+  /// Constructor (output gradient (Dy) OperandA ctor)
+  CUTLASS_HOST_DEVICE
+  TileIteratorStridedDgrad(
+    Params const &params,
+    ConvProblemSize const &problem_size,
+    Element const *ptr,
+    int thread_idx,
+    FastDivmod const &stride_h_divmod, FastDivmod const &stride_w_divmod,
+    int start_r, int start_s,
+    MatrixCoord const &threadblock_offset = MatrixCoord()
+  ):
+    tile_access_iterator_(
+      params, 
+      problem_size, 
+      ptr, 
+      thread_idx, 
+      stride_h_divmod, stride_w_divmod, 
+      start_r, start_s, 
+      threadblock_offset) { }
+
+  /// Constructor (filter (w) OperandB ctor)
   CUTLASS_HOST_DEVICE
   TileIteratorStridedDgrad(
     Params const &params,
@@ -210,7 +230,12 @@ public:
     int start_r, int start_s,
     MatrixCoord const &threadblock_offset = MatrixCoord()
   ):
-    tile_access_iterator_(params, problem_size, ptr, thread_idx, start_r, start_s, threadblock_offset) { }
+    tile_access_iterator_(params, 
+      problem_size, 
+      ptr, 
+      thread_idx, 
+      start_r, start_s, 
+      threadblock_offset) { }
 
   CUTLASS_HOST_DEVICE
   static Params getParams(ConvProblemSize const &problem_size, Layout const &layout) {