Add host API (#220)

* Add host API * manually rebase on develop * clean * manually rebase on develop * exclude tests from all target * address review comments * update client app name * fix missing lib name * clang-format update * refactor * refactor * refactor * refactor * refactor * fix test issue * refactor * refactor * refactor * upate cmake and readme Co-authored-by: Chao Liu <chao.liu2@amd.com>
2026-05-12 09:16:52 +00:00 · 2022-05-12 09:21:01 -05:00
parent 0f912e205e
commit cec69bc3bc
131 changed files with 1666 additions and 1089 deletions
--- a/library/include/ck/library/host/host_interface.hpp
+++ b/library/include/ck/library/host/host_interface.hpp
@@ -0,0 +1,54 @@
+#pragma once
+
+#include <memory>
+#include <string>
+
+#include "stream_config.hpp"
+#include "config.hpp"
+#include "device_base.hpp"
+
+struct DeviceConvFwdPtr_t
+{
+    using BaseArgument = ck::tensor_operation::device::BaseArgument;
+    using BaseInvoker  = ck::tensor_operation::device::BaseInvoker;
+
+    struct DeviceConvFwdPtrImpl;
+    std::unique_ptr<DeviceConvFwdPtrImpl> pImpl;
+    DeviceConvFwdPtr_t();
+    ~DeviceConvFwdPtr_t();
+    DeviceConvFwdPtr_t(DeviceConvFwdPtr_t&&);
+    DeviceConvFwdPtr_t(DeviceConvFwdPtrImpl&);
+    DeviceConvFwdPtr_t& operator=(DeviceConvFwdPtr_t&) = delete;
+    DeviceConvFwdPtr_t& operator=(const DeviceConvFwdPtr_t&) = delete;
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* in_ptr,
+                        void* wei_ptr,
+                        void* out_ptr,
+                        size_t N,
+                        size_t K,
+                        size_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads)
+        const; // in,wei and out element ops are ignored for now since even if we change them, they
+               // cant be linked
+    std::unique_ptr<BaseInvoker>
+    MakeInvokerPointer() const; // requires including BaseInvoker headers
+    std::string GetTypeString();
+    bool IsSupportedArgument(const BaseArgument* arg_ptr);
+};
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances);
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances_t(
+    std::vector<DeviceConvFwdPtr_t>& instances);
--- a/library/include/ck/library/host_tensor/device.hpp
+++ b/library/include/ck/library/host_tensor/device.hpp
@@ -1,12 +1,25 @@
-#ifndef DEVICE_HPP
-#define DEVICE_HPP
+#pragma once

 #include <memory>
 #include <functional>
 #include <thread>
 #include <chrono>
-#include "hip/hip_runtime.h"
-#include "hip/hip_fp16.h"
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+
+#include "stream_config.hpp"
+#include "ck/options.hpp"
+
+inline void hip_check_error(hipError_t x)
+{
+    if(x != hipSuccess)
+    {
+        std::ostringstream ss;
+        ss << "HIP runtime error: " << hipGetErrorString(x) << ". " << __FILE__ << ": " << __LINE__
+           << "in function: " << __func__;
+        throw std::runtime_error(ss.str());
+    }
+}

 struct DeviceMem
 {
@@ -36,49 +49,59 @@ struct KernelTimer
    std::unique_ptr<KernelTimerImpl> impl;
 };

-using device_stream_t = hipStream_t;
-
 template <typename... Args, typename F>
-void launch_kernel(F kernel, dim3 grid_dim, dim3 block_dim, std::size_t lds_byte, Args... args)
+float launch_and_time_kernel(const StreamConfig& stream_config,
+                             F kernel,
+                             dim3 grid_dim,
+                             dim3 block_dim,
+                             std::size_t lds_byte,
+                             Args... args)
 {
-    hipStream_t stream_id = nullptr;
-
-    hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_id, args...);
-}
-
-template <typename... Args, typename F>
-float launch_and_time_kernel(
-    F kernel, int nrepeat, dim3 grid_dim, dim3 block_dim, std::size_t lds_byte, Args... args)
-{
-    KernelTimer timer;
-
-    printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
-           __func__,
-           grid_dim.x,
-           grid_dim.y,
-           grid_dim.z,
-           block_dim.x,
-           block_dim.y,
-           block_dim.z);
-
-    printf("Warm up\n");
-
-    hipStream_t stream_id = nullptr;
-
-    // warm up
-    hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_id, args...);
-
-    printf("Start running %d times...\n", nrepeat);
-
-    timer.Start();
-
-    for(int i = 0; i < nrepeat; ++i)
+#if CK_TIME_KERNEL
+    if(stream_config.time_kernel_)
    {
-        hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_id, args...);
+        printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
+               __func__,
+               grid_dim.x,
+               grid_dim.y,
+               grid_dim.z,
+               block_dim.x,
+               block_dim.y,
+               block_dim.z);
+
+        const int nrepeat = 10;
+
+        printf("Warm up 1 time\n");
+
+        // warm up
+        hipLaunchKernelGGL(
+            kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);
+
+        printf("Start running %d times...\n", nrepeat);
+
+        KernelTimer timer;
+        timer.Start();
+
+        for(int i = 0; i < nrepeat; ++i)
+        {
+            hipLaunchKernelGGL(
+                kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);
+        }
+
+        timer.End();
+
+        return timer.GetElapsedTime() / nrepeat;
    }
+    else
+    {
+        hipLaunchKernelGGL(
+            kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);

-    timer.End();
+        return 0;
+    }
+#else
+    hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...);

-    return timer.GetElapsedTime() / nrepeat;
-}
+    return 0;
 #endif
+}
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
@@ -84,7 +84,8 @@ struct ReferenceBatchedGemm : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_backward_weight.hpp
@@ -121,7 +121,8 @@ struct ReferenceConvBwdWeight : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
@@ -291,7 +291,8 @@ struct ReferenceConvBwdData : public device::BaseOperator
            }
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
@@ -1,9 +1,10 @@
-#ifndef REFERENCE_CONV_FWD_HPP
-#define REFERENCE_CONV_FWD_HPP
+#pragma once

 #include <iostream>
 #include <type_traits>
 #include <sstream>
+
+#include "stream_config.hpp"
 #include "device_base.hpp"
 #include "host_tensor.hpp"

@@ -251,7 +252,8 @@ struct ReferenceConvFwd : public device::BaseOperator
            }
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
@@ -311,4 +313,3 @@ struct ReferenceConvFwd : public device::BaseOperator
 } // namespace host
 } // namespace tensor_operation
 } // namespace ck
-#endif
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
@@ -124,7 +124,8 @@ struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
@@ -130,7 +130,8 @@ struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
@@ -80,7 +80,8 @@ struct ReferenceGemm : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
@@ -82,7 +82,8 @@ struct ReferenceGemmBias2D : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp
@@ -85,7 +85,8 @@ struct ReferenceGemmBiasActivation : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp
@@ -91,7 +91,8 @@ struct ReferenceGemmBiasActivationAdd : public device::BaseOperator
            return 0;
        }

-        float Run(const device::BaseArgument* p_arg, int) override
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
--- a/library/include/ck/library/utility/op_instance_engine.hpp
+++ b/library/include/ck/library/utility/op_instance_engine.hpp
@@ -128,7 +128,7 @@ class OpInstanceRunEngine

    template <typename OpInstancePtr>
    ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
-                              int nrepeat          = 100,
+                              bool time_kernel     = false,
                              bool do_verification = false,
                              bool do_log          = false)
    {
@@ -143,7 +143,7 @@ class OpInstanceRunEngine
            if(op_ptr->IsSupportedArgument(argument.get()))
            {
                std::string op_name = op_ptr->GetTypeString();
-                float avg_time      = invoker->Run(argument.get(), nrepeat);
+                float avg_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});

                std::size_t flops     = op_instance_.GetFlops();
                std::size_t num_btype = op_instance_.GetBtype();