Merge branch 'develop' into aosewski/test_ggemm_splitk

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"

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

@@ -3,10 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
 #include <vector>
 #include <memory>
-
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"

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

@@ -3,10 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
 #include <vector>
 #include <memory>
-
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"

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

@@ -3,8 +3,7 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"

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

@@ -4,7 +4,7 @@
 #pragma once
 
 #include <cstdlib>
-
+#include <vector>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"

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

@@ -3,10 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
 #include <memory>
 #include <vector>
-
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm.hpp"

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

@@ -3,10 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
 #include <vector>
 #include <memory>
-
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"

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

@@ -3,10 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
 #include <vector>
 #include <memory>
-
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"

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

@@ -4,7 +4,7 @@
 #pragma once
 
 #include <vector>
-
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"

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

@@ -3,8 +3,8 @@
 
 #pragma once
 
-#include <cstdlib>
-
+#include <vector>
+#include <memory>
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_normalization.hpp"

library/include/ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto InOutRank  = 4;
+static constexpr auto WindowRank = 2;
+
+static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
+static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
+
+// FP16
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&);
+
+void add_device_pool2d_fwd_nhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&);
+
+// FP16 - return index
+void add_device_pool2d_fwd_nhwc_index_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&);
+
+// FP32
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&);
+
+void add_device_pool2d_fwd_nhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&);
+
+// FP32 - return index
+void add_device_pool2d_fwd_nhwc_index_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&);
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  IndexDataType,
+                                                                                  ReduceOpId,
+                                                                                  OutputIndex>>
+{
+    using DeviceOp = DevicePoolFwd<InOutRank,
+                                   WindowRank,
+                                   InDataType,
+                                   OutDataType,
+                                   IndexDataType,
+                                   ReduceOpId,
+                                   OutputIndex>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
+                     is_same_v<IndexDataType, I32>)
+        {
+            if constexpr(OutputIndex && ReduceOpId == MaxOp)
+            {
+                add_device_pool2d_fwd_nhwc_index_f16_instances(op_ptrs);
+            }
+            else
+            {
+                add_device_pool2d_fwd_nhwc_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
+                          is_same_v<IndexDataType, I32>)
+        {
+            if constexpr(OutputIndex && ReduceOpId == MaxOp)
+            {
+                add_device_pool2d_fwd_nhwc_index_f32_instances(op_ptrs);
+            }
+            else
+            {
+                add_device_pool2d_fwd_nhwc_f32_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto InOutRank  = 5;
+static constexpr auto WindowRank = 3;
+
+static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
+static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
+
+// FP16
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&);
+
+// FP16 - return index
+void add_device_pool3d_fwd_ndhwc_index_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&);
+
+// FP32
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&);
+
+// FP32 - return index
+void add_device_pool3d_fwd_ndhwc_index_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&);
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  IndexDataType,
+                                                                                  ReduceOpId,
+                                                                                  OutputIndex>>
+{
+    using DeviceOp = DevicePoolFwd<InOutRank,
+                                   WindowRank,
+                                   InDataType,
+                                   OutDataType,
+                                   IndexDataType,
+                                   ReduceOpId,
+                                   OutputIndex>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
+                     is_same_v<IndexDataType, I32>)
+        {
+            if constexpr(OutputIndex && ReduceOpId == MaxOp)
+            {
+                add_device_pool3d_fwd_ndhwc_index_f16_instances(op_ptrs);
+            }
+            else
+            {
+                add_device_pool3d_fwd_ndhwc_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
+                          is_same_v<IndexDataType, I32>)
+        {
+            if constexpr(OutputIndex && ReduceOpId == MaxOp)
+            {
+                add_device_pool3d_fwd_ndhwc_index_f32_instances(op_ptrs);
+            }
+            else
+            {
+                add_device_pool3d_fwd_ndhwc_f32_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck