Improve external interface for GEMM and GEMM+add+add+fastgelu (#311)

* interface for GEMM and GEMM+add+add+fastgelu

* rename namespace

* instance factory

* fix build

* fix build; add GEMM client example

* clean

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -24,7 +24,7 @@ ADD_BLOCKWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -48,7 +48,7 @@ ADD_BLOCKWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -24,7 +24,7 @@ ADD_BLOCKWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -48,7 +48,7 @@ ADD_BLOCKWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_BLOCKWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -36,7 +36,7 @@ ADD_BLOCKWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
 ADD_BLOCKWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -20,7 +20,7 @@ ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -49,7 +49,7 @@ ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
 ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -36,7 +36,7 @@ ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -24,7 +24,7 @@ ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -48,7 +48,7 @@ ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -24,7 +24,7 @@ ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -48,7 +48,7 @@ ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -21,7 +21,7 @@ ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
 // clang-format on
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.cpp

@@ -6,7 +6,7 @@
 namespace ck {
 namespace tensor_operation {
 namespace device {
-namespace device_reduce_instance {
+namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
@@ -36,7 +36,7 @@ ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
 ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
 // clang-format on
 
-} // namespace device_reduce_instance
+} // namespace instance
 } // namespace device
 } // namespace tensor_operation
 

profiler/include/profile_batched_gemm_impl.hpp

@@ -10,7 +10,7 @@
 #include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/tensor_operation_instance/gpu/device_batched_gemm_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/host_tensor/device_memory.hpp"
@@ -116,19 +116,21 @@ bool profile_batched_gemm_impl(int do_verification,
     b_device_buf.ToDevice(b_g_k_n.mData.data());
     c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
 
-    // add device op instances
-    const auto op_ptrs = ck::tensor_operation::device::device_batched_gemm_instance::
-        get_device_batched_gemm_instances<ADataType,
-                                          BDataType,
-                                          CDataType,
-                                          ALayout,
-                                          BLayout,
-                                          CLayout>();
-
-    if(op_ptrs.size() <= 0)
-    {
-        throw std::runtime_error("wrong! no device GEMM instance found");
-    }
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm<ALayout,
+                                                                     BLayout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
     std::string best_op_name;
     float best_ave_time   = 0;