From e5be038a5e8e64ec6eb56369b910bdccb8df6fb1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Ville=20Pietil=C3=A4?= <>
Date: Fri, 16 Jan 2026 05:32:29 -0500
Subject: [PATCH] Fix GEMM pipeline concept usage.

---
 .../include/ck_tile/builder/factory/README.md       |  7 +++----
 .../ck_tile/builder/factory/conv_algorithms.hpp     | 13 ++++++-------
 .../builder/test/impl/conv_algorithm_types.hpp      |  4 ----
 3 files changed, 9 insertions(+), 15 deletions(-)

diff --git a/experimental/builder/include/ck_tile/builder/factory/README.md b/experimental/builder/include/ck_tile/builder/factory/README.md
index 4c30348f9a..f82b5f9ad6 100644
--- a/experimental/builder/include/ck_tile/builder/factory/README.md
+++ b/experimental/builder/include/ck_tile/builder/factory/README.md
@@ -49,14 +49,13 @@ XDL and WMMA algorithms share a common base, while DL algorithms have their own
 Both XDL and WMMA algorithms share the following foundational concepts:
 
 ```
-ConvAlgorithm (Base Concept)
+ConvWarpGemmAlgorithm (Base Concept)
 │
 │  Requirements:
 │  • ConvAlgorithmDescriptor
 │  • SpecifiesThreadBlock
 │  • SpecifiesTileTransferParameters (ThreadClusters, LdsTransfer, AccessOrders)
 │  • SpecifiesWarpGemm
-│  • SpecifiesGemmPipeline
 │
 ├─── FwdAlgorithm (Forward Convolution)
 │    │
@@ -64,7 +63,7 @@ ConvAlgorithm (Base Concept)
 │    │
 │    └─── FwdAlgorithmV3
 │         │
-│         │  Additional: SpecifiesPipelineV3
+│         │  Additional: SpecifiesPipelineV3 + SpecifiesGemmPipeline
 │         │
 │
 └─── BwdAlgorithm (Backward Weight Convolution)
@@ -73,7 +72,7 @@ ConvAlgorithm (Base Concept)
      │
      └─── BwdAlgorithmV3
           │
-          │  Additional: SpecifiesPipelineV3
+          │  Additional: SpecifiesPipelineV3 + SpecifiesGemmPipeline
           │
 ```
 
diff --git a/experimental/builder/include/ck_tile/builder/factory/conv_algorithms.hpp b/experimental/builder/include/ck_tile/builder/factory/conv_algorithms.hpp
index ad72609f84..0c4d766fff 100644
--- a/experimental/builder/include/ck_tile/builder/factory/conv_algorithms.hpp
+++ b/experimental/builder/include/ck_tile/builder/factory/conv_algorithms.hpp
@@ -14,23 +14,22 @@ concept SpecifiesTileTransferParameters =
 
 // Base algorithm concepts
 template <typename T, size_t ThreadClusterRank = 3>
-concept ConvAlgorithm = ConvAlgorithmDescriptor<T> && SpecifiesThreadBlock<T> &&
+concept ConvWarpGemmAlgorithm = ConvAlgorithmDescriptor<T> && SpecifiesThreadBlock<T> &&
                         SpecifiesTileTransferParameters<T, ThreadClusterRank> &&
-                        SpecifiesWarpGemm<T> && SpecifiesGemmPipeline<T>;
-;
+                        SpecifiesWarpGemm<T>;
 
 template <typename T>
-concept FwdAlgorithm = ConvAlgorithm<T, 3> && SpecifiesFwdConvSpecialization<T>;
+concept FwdAlgorithm = ConvWarpGemmAlgorithm<T, 3> && SpecifiesFwdConvSpecialization<T>;
 
 template <typename T>
-concept FwdAlgorithmV3 = FwdAlgorithm<T> && SpecifiesPipelineV3<T>;
+concept FwdAlgorithmV3 = FwdAlgorithm<T> && SpecifiesPipelineV3<T> && SpecifiesGemmPipeline<T>;
 
 template <typename T, size_t ThreadClusterRank = 3>
 concept BwdAlgorithm =
-    ConvAlgorithm<T, ThreadClusterRank> && SpecifiesBwdWeightConvSpecialization<T>;
+    ConvWarpGemmAlgorithm<T, ThreadClusterRank> && SpecifiesBwdWeightConvSpecialization<T>;
 
 template <typename T>
-concept BwdAlgorithmV3 = BwdAlgorithm<T, 3> && SpecifiesPipelineV3<T>;
+concept BwdAlgorithmV3 = BwdAlgorithm<T, 3> && SpecifiesPipelineV3<T> && SpecifiesGemmPipeline<T>;
 
 template <typename T>
 concept DlAlgorithm =
diff --git a/experimental/builder/test/impl/conv_algorithm_types.hpp b/experimental/builder/test/impl/conv_algorithm_types.hpp
index cf4bc25c0f..6989887a26 100644
--- a/experimental/builder/test/impl/conv_algorithm_types.hpp
+++ b/experimental/builder/test/impl/conv_algorithm_types.hpp
@@ -543,8 +543,6 @@ using ConvAlgorithm_DeviceGroupedConvBwdWeight_Xdl_CShuffle =
                           WarpGemm_,
                           InputOutputTileTransfer_<4>,
                           ConvSpecializationBwdWeight_,
-                          GemmPipeline_, // Not needed, but we need this to meet the ConvAlgorithm
-                                         // concept.
                           TransposeParams_,
                           AlgorithmSpecialization_<>>;
 
@@ -594,8 +592,6 @@ using ConvAlgorithm_DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle =
     ConvAlgorithmTemplate<ThreadBlock_,
                           WarpGemm_,
                           InputOutputTileTransfer_<4>,
-                          GemmPipeline_, // Not needed, but we need this to meet the ConvAlgorithm
-                                         // concept.
                           ConvSpecializationBwdWeight_,
                           AlgorithmSpecialization_<MULTIPLE_D>>;