Merge branch 'develop' into gemm_layernorm_instance

include/ck/utility/reduction_operator.hpp

@@ -251,27 +251,27 @@ constexpr T GetIdentityValueForInMemoryDataOperation(InMemoryDataOperationEnum o
 };
 
 template <InMemoryDataOperationEnum Operation, typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType
+struct InMemoryDataOperationSupportedOnDataType
 {
     static constexpr bool value = false;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value;
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value ||
@@ -280,7 +280,7 @@ struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Set, D
 };
 
 template <typename DataType>
-struct InMemoryDataOperatonSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
 {
     static constexpr bool value =
         is_same<DataType, float>::value || is_same<DataType, double>::value ||

include/ck/utility/synchronization.hpp

@@ -18,6 +18,7 @@ __device__ void block_sync_lds()
     __syncthreads();
 #endif
 }
+
 __device__ void s_nop()
 {
 #if 1

library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp

@@ -24,11 +24,14 @@ struct ReferenceSoftmax : public device::BaseOperator
     {
         Argument(const Tensor<InDataType>& in,
                  Tensor<OutDataType>& out,
-                 AccDataType alpha,
-                 AccDataType beta,
+                 double alpha,
+                 double beta,
                  const std::vector<index_t> sm_reduce_dims)
-            : in_(in), out_(out), alpha_(alpha), beta_(beta), sm_reduce_dims_(sm_reduce_dims)
+            : in_(in), out_(out), sm_reduce_dims_(sm_reduce_dims)
         {
+            alpha_ = static_cast<AccDataType>(alpha);
+            beta_  = static_cast<AccDataType>(beta);
+
             // std::cout << "debug: scalar dims: ";
             for(size_t i = 0; i < in.mDesc.GetNumOfDimension(); i++)
             {
@@ -143,8 +146,8 @@ struct ReferenceSoftmax : public device::BaseOperator
 
     static auto MakeArgument(const Tensor<InDataType>& in,
                              Tensor<OutDataType>& out,
-                             AccDataType alpha,
-                             AccDataType beta,
+                             double alpha,
+                             double beta,
                              const std::vector<index_t> sm_reduce_dims)
     {
         return Argument{in, out, alpha, beta, sm_reduce_dims};

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -91,6 +91,7 @@ using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
 using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
 using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
+using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp

@@ -7,7 +7,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -18,11 +18,8 @@ namespace device {
 namespace instance {
 
 using Normalize                            = ck::tensor_operation::element_wise::Normalize;
-using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::DeviceElementwiseBasePtr<
-    Tuple<half_t, float, float, half_t, half_t>,
-    Tuple<half_t>,
-    Normalize,
-    2>;
+using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::
+    DeviceElementwisePtr<Tuple<half_t, float, float, half_t, half_t>, Tuple<half_t>, Normalize, 2>;
 
 void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
     std::vector<DeviceNormalizeFromMeanMeanSquarePtr>& instances);

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp

@@ -76,8 +76,16 @@ template <typename InDataType,
           bool PropagateNan,
           bool OutputIndex>
 void add_device_reduce_instance_blockwise(
-    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
-        device_op_instances)
+    std::vector<DeviceReducePtr<InDataType,
+                                AccDataType,
+                                OutDataType,
+                                Rank,
+                                NumReduceDim,
+                                ReduceOperation,
+                                InElementwiseOp,
+                                AccElementwiseOp,
+                                PropagateNan,
+                                OutputIndex>>& device_op_instances)
 {
     static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
         [&](auto i) {

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance