Merge remote-tracking branch 'origin/develop' into aosewski/gemm_tile_loop

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -198,7 +198,9 @@ template <index_t NDimSpatial,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
-          LoopScheduler LoopSched = make_default_loop_scheduler()>
+          LoopScheduler LoopSched = make_default_loop_scheduler(),
+          typename AComputeType   = ADataType,
+          typename BComputeType   = AComputeType>
 struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
     : public DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
                                                ALayout,    // output image
@@ -211,7 +213,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                                EDataType,  // input image
                                                AElementwiseOp,
                                                BElementwiseOp,
-                                               CDEElementwiseOp>
+                                               CDEElementwiseOp,
+                                               AComputeType,
+                                               BComputeType>
 {
     // TODO: Extend support for more spatial dimensions.
     static_assert(NDimSpatial == 2 || NDimSpatial == 3,
@@ -312,9 +316,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
-        ABDataType, // TODO: distinguish A/B datatype
-        ABDataType, // TODO: distinguish A/B datatype
-        ABDataType, // TODO: distinguish A/B datatype
+        ABDataType,
+        ABDataType,
+        AComputeType,
         AccDataType,
         CShuffleDataType,
         DsDataType,
@@ -354,7 +358,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
         CShuffleNXdlPerWavePerShuffle,
         CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CDEBlockTransferScalarPerVector_NPerBlock,
-        LoopSched>;
+        LoopSched,
+        PipelineVersion::v1,
+        BComputeType>;
 
     template <typename Desc_K0_M_K1>
     static auto transform_k0_m_k1_to_m_k(const Desc_K0_M_K1& desc_k0_m_k1)

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp

@@ -211,6 +211,7 @@ template <index_t NDimSpatial,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          typename ComputeDataType = ADataType,
           LoopScheduler LoopSched  = make_default_loop_scheduler()>
 struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
     : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
@@ -224,7 +225,8 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
                                            EDataType,
                                            AElementwiseOperation,
                                            BElementwiseOperation,
-                                           CDEElementwiseOperation>
+                                           CDEElementwiseOperation,
+                                           ComputeDataType>
 {
     using DeviceOp = DeviceGroupedConvFwdMultipleD_Xdl_CShuffle;
 
@@ -323,8 +325,6 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
     using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
 
-    using ComputeDataType = ADataType;
-
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -186,6 +186,25 @@ struct Bilinear
         y = type_convert<half_t>(alpha_ * x0 + beta_ * ck::type_convert<float>(x1));
     };
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x0_tmp = type_convert<float>(x0);
+        const float x1_tmp = type_convert<float>(x1);
+        const float y_tmp  = alpha_ * x0_tmp + beta_ * x1_tmp;
+        y                  = type_convert<bhalf_t>(y_tmp);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, float, bhalf_t>(bhalf_t& y, const float& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = alpha_ * x0 + beta_ * x1_tmp;
+        y                  = y_tmp;
+    };
+
     template <>
     __host__ __device__ constexpr void operator()<std::int8_t, std::int32_t, std::int8_t>(
         std::int8_t& y, const std::int32_t& x0, const std::int8_t& x1) const

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp

@@ -522,6 +522,7 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),