[hipdnn] Warn when conv wgrad/dgrad infers output dims assuming groups=1 (#7263)

Closes #5259.

The two `infer_properties_node()` implementations silently assume
`groups = 1` when computing `dw[1]` / `dx[1]`, which gives the wrong
channel count for grouped convolutions unless the caller passes `dw` /
`dx` shapes explicitly.

This PR makes the assumption visible. Each `infer_properties_node()` now
emits a `HIPDNN_FE_LOG_WARN` describing what was assumed and how to
override it, and the inline comments are rewritten to match. The same
caveat is added as a `@note` on `conv_dgrad` and `conv_wgrad` in
`Graph.hpp` so the limitation shows up in the public docs.

No behavior change for non-grouped convolutions or for callers that pass
explicit `dw` / `dx` dimensions. Grouped convolutions that previously
relied on inference now log a warning; the inferred channel count itself
is unchanged by this PR, since fixing it would require a separate API
discussion.

---------

Co-authored-by: BrianHarrisonAMD <169072757+BrianHarrisonAMD@users.noreply.github.com>

Author: 1wos

projects/hipdnn/frontend/include/hipdnn_frontend/Graph.hpp

@@ -2890,6 +2890,12 @@ class Graph : public INode
      *        (must match forward pass)
      * @return dx: Gradient w.r.t. input (same shape as forward input)
      *
+     * @note If `dx` dimensions are not provided, the channel count is
+     *       inferred assuming `groups = 1`. For grouped convolutions,
+     *       set dimensions on the returned `dx` tensor before graph
+     *       validation/finalization to avoid an incorrect channel count
+     *       on the inferred input-gradient tensor.
+     *
      * @see hipdnn_frontend::graph::ConvDgradAttributes
      */
     // NOLINTBEGIN(readability-identifier-naming)
@@ -2942,6 +2948,12 @@ class Graph : public INode
      *        (must match forward pass)
      * @return dw: Gradient w.r.t. filter weights (same shape as forward weights)
      *
+     * @note If `dw` dimensions are not provided, the channel count is
+     *       inferred assuming `groups = 1`. For grouped convolutions,
+     *       set dimensions on the returned `dw` tensor before graph
+     *       validation/finalization to avoid an incorrect channel count
+     *       on the inferred weight tensor.
+     *
      * @see hipdnn_frontend::graph::ConvWgradAttributes
      */
     // NOLINTBEGIN(readability-identifier-naming)

projects/hipdnn/frontend/include/hipdnn_frontend/node/ConvolutionDgradNode.hpp

@@ -5,6 +5,7 @@
 #include "Node.hpp"
 #include <hipdnn_data_sdk/utilities/ShapeUtilities.hpp>
 #include <hipdnn_frontend/Error.hpp>
+#include <hipdnn_frontend/Logging.hpp>
 #include <hipdnn_frontend/attributes/ConvolutionDgradAttributes.hpp>
 #include <hipdnn_frontend/attributes/GraphAttributes.hpp>
 #include <hipdnn_frontend/detail/ConvolutionDgradPacker.hpp>
@@ -258,9 +259,15 @@ class ConvolutionDgradNode : public BaseNode<ConvolutionDgradNode, NodeType::CON
 
             dxDims[0] = dyDims[0]; // N (batch) matches dy
 
-            // Impossible to infer group count without dx dimensions.
-            // Therefore, assume groups = 1.
-            dxDims[1] = wDims[1]; // C (input channels)
+            // Group count cannot be inferred from dy and w alone, so the
+            // inferred dx[1] uses w[1] (i.e. assumes groups = 1). For
+            // grouped convolutions, callers should set dx dimensions
+            // explicitly to avoid an incorrect channel count on the
+            // inferred input-gradient tensor.
+            HIPDNN_FE_LOG_WARN("ConvolutionDgradNode: inferring dx dimensions without an "
+                               "explicit dx shape; assuming groups=1. For grouped "
+                               "convolutions, set dx dimensions explicitly.");
+            dxDims[1] = wDims[1]; // C (input channels), assuming groups=1
 
             // We calculate spatial dimensions (i_2, ..., i_n)
             for(size_t i = 2; i < dyDims.size(); ++i)

projects/hipdnn/frontend/include/hipdnn_frontend/node/ConvolutionWgradNode.hpp

@@ -5,6 +5,7 @@
 #include "Node.hpp"
 #include <hipdnn_data_sdk/utilities/ShapeUtilities.hpp>
 #include <hipdnn_frontend/Error.hpp>
+#include <hipdnn_frontend/Logging.hpp>
 #include <hipdnn_frontend/attributes/ConvolutionWgradAttributes.hpp>
 #include <hipdnn_frontend/attributes/GraphAttributes.hpp>
 #include <hipdnn_frontend/detail/ConvolutionWgradPacker.hpp>
@@ -254,9 +255,15 @@ class ConvolutionWgradNode : public BaseNode<ConvolutionWgradNode, NodeType::CON
 
             dwDims[0] = dyDims[1]; // Output channels match dy channels
 
-            // Impossible to infer group count without dw dimensions.
-            // Therefore, assume groups = 1.
-            dwDims[1] = xDims[1]; // Input channels (per group)
+            // Group count cannot be inferred from x and dy alone, so the
+            // inferred dw[1] uses x[1] (i.e. assumes groups = 1). For
+            // grouped convolutions, callers should set dw dimensions
+            // explicitly to avoid an incorrect channel count on the
+            // inferred weight tensor.
+            HIPDNN_FE_LOG_WARN("ConvolutionWgradNode: inferring dw dimensions without an "
+                               "explicit dw shape; assuming groups=1. For grouped "
+                               "convolutions, set dw dimensions explicitly.");
+            dwDims[1] = xDims[1]; // Input channels (per group), assuming groups=1
 
             // Calculate kernel spatial dimensions (k_2, ..., k_n)
             for(size_t i = 2; i < dyDims.size(); ++i)