Merge remote-tracking branch 'origin/develop' into migx-flash-attn

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp

@@ -16,7 +16,6 @@ template <index_t Rank, index_t Reduce>
 using device_softmax_f16_f16_instances = std::tuple<
     // clang-format off
     //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
-    // fallback kernel
     DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,              1>,
     DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               8,              8>,
     DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               8,              8>,
@@ -33,6 +32,13 @@ using device_softmax_f16_f16_instances = std::tuple<
     // clang-format on
     >;
 
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_generic_instance = std::tuple<
+    // clang-format off
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,        64,                  8,                  8,                1,                1,              1,               1,              1>
+    // clang-format on
+    >;
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_f32_f32_rank3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
-void add_device_softmax_f32_f32_rank4_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank3_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank3_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank3_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce4_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances);
 
 } // namespace instance
 } // namespace device

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp

@@ -16,7 +16,7 @@ template <index_t Rank, index_t Reduce>
 using device_softmax_f32_f32_instances = std::tuple<
     // clang-format off
     //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
-    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,               1>, // fallback kernel
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,               1>,
     DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               4,               4>,
     DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               4,               4>,
     DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               4,               4>,
@@ -32,6 +32,13 @@ using device_softmax_f32_f32_instances = std::tuple<
     // clang-format on
     >;
 
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_generic_instance = std::tuple<
+    // clang-format off
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,        64,                  8,                  8,                1,                1,              1,               1,               1>
+    // clang-format on
+    >;
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank3_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
-void add_device_softmax_i8_i8_rank4_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank3_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank3_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank4_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank4_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank4_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_i8_i8_rank4_reduce4_instances(
-    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#include <tuple>
-
-#include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
-#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-template <index_t Rank, index_t Reduce>
-using device_softmax_i8_i8_instances = std::tuple<
-    // clang-format off
-    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
-    // fallback kernel
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,               1,              1>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,               16,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               64,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,              16,             16>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               64,              1,              16,             16>,
-    // Reduction on middle dimensions
-    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                8,              0,               1,              1>,
-    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                 32,                  8,               32,                8,              0,              16,              8>
-    // clang-format on
-    >;
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp

@@ -3,6 +3,17 @@
 
 #pragma once
 
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"

library/include/ck/library/utility/device_memory.hpp

@@ -20,8 +20,9 @@ __global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
  */
 struct DeviceMem
 {
-    DeviceMem() = delete;
+    DeviceMem() : mpDeviceBuf(nullptr), mMemSize(0) {}
     DeviceMem(std::size_t mem_size);
+    void Realloc(std::size_t mem_size);
     void* GetDeviceBuffer() const;
     std::size_t GetBufferSize() const;
     void ToDevice(const void* p) const;