Merge branch 'lib_gemm_softmax_gemm_type' of...

Merge branch 'lib_gemm_softmax_gemm_type' of https://github.com/ROCmSoftwarePlatform/composable_kernel into lib_gemm_softmax_gemm_type

example/40_conv2d_fwd_quantization/CMakeLists.txt

@@ -14,3 +14,8 @@ add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_in
 add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
 add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
 
+# Conv + bias + tanh perlayer quantization
+add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
+
+# Conv + bias + tanh perchannel quantization
+add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp

@@ -76,6 +76,10 @@ using DeviceGroupedConvNDFwdInstance =
         5,                   // CThreadTransferSrcDstVectorDim
         4>;                  // CThreadTransferDstScalarPerVector
 
-#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
 
-int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp

@@ -74,6 +74,11 @@ using DeviceGroupedConvNDFwdInstance =
         5,                   // CThreadTransferSrcDstVectorDim
         4>;                  // CThreadTransferDstScalarPerVector
 
-#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
 
-int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
+
+int main()
+{
+    float scale_z_inv         = 0.5f;
+    const auto out_element_op = OutElementOp{scale_z_inv, ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using AccDataType  = int32_t;
+using OutDataType  = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
+
+int main()
+{
+    float scale_acc           = 0.5f;
+    float scale_z_inv         = 0.5f;
+    const auto out_element_op = OutElementOp{scale_z_inv, scale_acc, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp

@@ -76,4 +76,8 @@ using DeviceGroupedConvNDFwdInstance =
 
 #include "run_conv2d_fwd_perchannel_quantization_example.inc"
 
-int main() { run_conv2d_fwd_perchannel_quantization_example(); }
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_perchannel_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp

@@ -71,4 +71,9 @@ using DeviceGroupedConvNDFwdInstance =
 
 #include "run_conv2d_fwd_perlayer_quantization_example.inc"
 
-int main() { run_conv2d_fwd_perlayer_quantization_example(); }
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_perlayer_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp

@@ -80,6 +80,10 @@ using DeviceGroupedConvNDFwdInstance =
         S<1, 64, 1, 4>,
         8>;
 
-#include "run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc"
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
 
-int main() { run_conv2d_fwd_bias_relu_perchannel_quantization_example(); };
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};

example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp

@@ -78,6 +78,11 @@ using DeviceGroupedConvNDFwdInstance =
         S<1, 64, 1, 4>,
         8>;
 
-#include "run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc"
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
 
-int main() { run_conv2d_fwd_bias_relu_perlayer_quantization_example(); }
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp

@@ -80,4 +80,8 @@ using DeviceGroupedConvNDFwdInstance =
 
 #include "run_conv2d_fwd_perchannel_quantization_example.inc"
 
-int main() { run_conv2d_fwd_perchannel_quantization_example(); }
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_perchannel_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp

@@ -75,4 +75,9 @@ using DeviceGroupedConvNDFwdInstance =
 
 #include "run_conv2d_fwd_perlayer_quantization_example.inc"
 
-int main() { run_conv2d_fwd_perlayer_quantization_example(); }
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_perlayer_quantization_example(out_element_op);
+}

example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_relu_perchannel_quantization_example.inc → example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc

@@ -167,7 +167,7 @@ bool run_grouped_conv_fwd(bool do_verification,
     return (pass ? 0 : 1);
 }
 
-int run_conv2d_fwd_bias_relu_perchannel_quantization_example()
+int run_conv2d_fwd_bias_perchannel_quantization_example(const OutElementOp& out_element_op)
 {
     bool do_verification           = true;
     bool time_kernel               = true;
@@ -189,7 +189,6 @@ int run_conv2d_fwd_bias_relu_perchannel_quantization_example()
 
     const auto in_element_op  = InElementOp{};
     const auto wei_element_op = WeiElementOp{};
-    const auto out_element_op = OutElementOp{ActivationOp{}};
 
     using InLayout           = ck::tensor_layout::convolution::GNHWC;
     using WeiLayout          = ck::tensor_layout::convolution::GKYXC;

example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_relu_perlayer_quantization_example.inc → example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc

@@ -155,7 +155,7 @@ bool run_grouped_conv_fwd(bool do_verification,
     return (pass ? 0 : 1);
 }
 
-int run_conv2d_fwd_bias_relu_perlayer_quantization_example()
+int run_conv2d_fwd_bias_perlayer_quantization_example(const OutElementOp& out_element_op)
 {
     bool do_verification           = true;
     bool time_kernel               = true;
@@ -177,7 +177,6 @@ int run_conv2d_fwd_bias_relu_perlayer_quantization_example()
 
     const auto in_element_op  = InElementOp{};
     const auto wei_element_op = WeiElementOp{};
-    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
 
     using InLayout   = ck::tensor_layout::convolution::GNHWC;
     using WeiLayout  = ck::tensor_layout::convolution::GKYXC;

example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc

@@ -157,7 +157,7 @@ bool run_grouped_conv_fwd(bool do_verification,
     return (pass ? 0 : 1);
 }
 
-int run_conv2d_fwd_perchannel_quantization_example()
+int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_element_op)
 {
     bool do_verification           = true;
     bool time_kernel               = true;
@@ -179,7 +179,6 @@ int run_conv2d_fwd_perchannel_quantization_example()
 
     const auto in_element_op  = InElementOp{};
     const auto wei_element_op = WeiElementOp{};
-    const auto out_element_op = OutElementOp{ActivationOp{}};
 
     using InLayout           = ck::tensor_layout::convolution::GNHWC;
     using WeiLayout          = ck::tensor_layout::convolution::GKYXC;

example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc

@@ -139,7 +139,7 @@ bool run_grouped_conv_fwd(bool do_verification,
     return (pass ? 0 : 1);
 }
 
-int run_conv2d_fwd_perlayer_quantization_example()
+int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element_op)
 {
     bool do_verification           = true;
     bool time_kernel               = false;
@@ -161,7 +161,6 @@ int run_conv2d_fwd_perlayer_quantization_example()
 
     const auto in_element_op  = InElementOp{};
     const auto wei_element_op = WeiElementOp{};
-    const auto out_element_op = OutElementOp{0.5f, ActivationOp{}};
 
     using InLayout  = ck::tensor_layout::convolution::GNHWC;
     using WeiLayout = ck::tensor_layout::convolution::GKYXC;

include/ck/ck.hpp

@@ -36,7 +36,7 @@
 #elif defined(__gfx1030__) // for GPU code
 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
 #elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
-#define CK_BUFFER_RESOURCE_3RD_DWORD 0x10020000
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
 #endif
 
 // FMA instruction

include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

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

@@ -840,17 +840,8 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
             << KPerBlock << ", "
             << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
             << K1 << ", "
-            << MPerXDL << ", "
-            << NPerXDL << ", "
-            << MXdlPerWave << ", "
-            << NXdlPerWave << ", "
             << ABlockTransferSrcScalarPerVector << ", "
-            << ABlockTransferDstScalarPerVector_K1 << ", "
-            << BBlockTransferSrcScalarPerVector << ", "
-            << BBlockTransferDstScalarPerVector_K1 << ", "
-            << CShuffleMXdlPerWavePerShuffle << ", "
-            << CShuffleNXdlPerWavePerShuffle << ", "
-            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << BBlockTransferSrcScalarPerVector
             << ">";
         // clang-format on
 

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once