refactor build, clean up

src/include/gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded_lds_pipeline.cuh

@@ -28,9 +28,9 @@ template <unsigned GridSize,
           unsigned WeiBlockCopyThreadPerDim0,
           unsigned WeiBlockCopyThreadPerDim1>
 __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded_lds_pipeline(
-    Float* const __restrict__ p_in_global,
-    Float* const __restrict__ p_wei_global,
-    Float* __restrict__ p_out_global)
+    const Float* const __restrict__ p_in_global,
+    const Float* const __restrict__ p_wei_global,
+    Float* const __restrict__ p_out_global)
 {
     // NPerThread == NPerBlock, because the format of input in LDS [C,Hi,Wi,N]
     //   for GEMM trans([C,K]) * [C,Wo*N], we need a thread to do all the "N"
@@ -220,7 +220,7 @@ __global__ void gridwise_implicit_gemm_convolution_1_chwn_csrk_khwn_padded_lds_p
     // set threadwise output tensor to 0
     threadwise_4d_tensor_set_zero(out_hkwn_thread_desc, p_out_thread);
 
-    Float* p_wei_global_block_begin =
+    const Float* p_wei_global_block_begin =
         p_wei_global + wei_ek_global_desc.Get1dIndex(0, k_block_data_begin);
 
     // prelog: load data

src/include/gridwise_implicit_gemm_convolution_1_nchw_kcsr.cuh → src/include/gridwise_implicit_gemm_convolution_1_nchw_kcsr_nkhw.cuh

@@ -22,12 +22,9 @@ template <unsigned GridSize,
           unsigned HoPerThread,
           unsigned WoPerThread>
 __global__ void
-gridwise_implicit_gemm_convolution_1_nchw_kcsr(InGlobalDesc,
-                                               Float* const __restrict__ p_in_global,
-                                               WeiGlobalDesc,
-                                               Float* const __restrict__ p_wei_global,
-                                               OutGlobalDesc,
-                                               Float* __restrict__ p_out_global)
+gridwise_implicit_gemm_convolution_1_nchw_kcsr_nkhw(const Float* const __restrict__ p_in_global,
+                                                    const Float* const __restrict__ p_wei_global,
+                                                    Float* const __restrict__ p_out_global)
 {
     // NPerThread == NPerBlock, because the format of input in LDS [C,Hi,Wi,N]
     //   for GEMM trans([C,K]) * [C,Wo*N], we need a thread to do all the "N"

src/include/gridwise_implicit_gemm_convolution_1_nchw_srck_nkhw.cuh

@@ -23,12 +23,9 @@ template <unsigned GridSize,
           unsigned HoPerThread,
           unsigned WoPerThread>
 __global__ void
-gridwise_implicit_gemm_convolution_1_nchw_srck_nkhw(InGlobalDesc,
-                                                    Float* const __restrict__ p_in_global,
-                                                    WeiGlobalDesc,
-                                                    Float* const __restrict__ p_wei_global,
-                                                    OutGlobalDesc,
-                                                    Float* __restrict__ p_out_global)
+gridwise_implicit_gemm_convolution_1_nchw_srck_nkhw(const Float* const __restrict__ p_in_global,
+                                                    const Float* const __restrict__ p_wei_global,
+                                                    Float* const __restrict__ p_out_global)
 {
     // NPerThread == NPerBlock, because the format of input in LDS [C,Hi,Wi,N]
     //   for GEMM trans([C,K]) * [C,Wo*N], we need a thread to do all the "N"

src/include/gridwise_implicit_gemm_convolution_2_cnhw_csrk_knhw.cuh

@@ -35,11 +35,8 @@ template <unsigned GridSize,
           unsigned InBlockCopyDataPerRead,
           unsigned WeiBlockCopyDataPerRead>
 __global__ void
-gridwise_implicit_gemm_convolution_2_cnhw_csrk_knhw(InGlobalDesc,
-                                                    const Float* const __restrict__ p_in_global,
-                                                    WeiGlobalDesc,
+gridwise_implicit_gemm_convolution_2_cnhw_csrk_knhw(const Float* const __restrict__ p_in_global,
                                                     const Float* const __restrict__ p_wei_global,
-                                                    OutGlobalDesc,
                                                     Float* const __restrict__ p_out_global)
 {
     constexpr auto I0 = Number<0>{};

src/include/gridwise_implicit_gemm_convolution_2_cnhw_csrk_knhw_lds_double_buffer.cuh

@@ -35,12 +35,9 @@ template <unsigned GridSize,
           unsigned InBlockCopyDataPerRead,
           unsigned WeiBlockCopyDataPerRead>
 __global__ void gridwise_implicit_gemm_convolution_2_cnhw_csrk_knhw_lds_double_buffer(
-    InGlobalDesc,
-    Float* const __restrict__ p_in_global,
-    WeiGlobalDesc,
-    Float* const __restrict__ p_wei_global,
-    OutGlobalDesc,
-    Float* __restrict__ p_out_global)
+    const Float* const __restrict__ p_in_global,
+    const Float* const __restrict__ p_wei_global,
+    Float* const __restrict__ p_out_global)
 {
     constexpr auto I0 = Number<0>{};
     constexpr auto I1 = Number<1>{};

src/include/gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw.cuh

@@ -25,12 +25,9 @@ template <unsigned GridSize,
           unsigned InBlockCopyThreadPerDim0,
           unsigned InBlockCopyThreadPerDim1>
 __global__ void
-gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw(InGlobalDesc,
-                                                    Float* const __restrict__ p_in_global,
-                                                    WeiGlobalDesc,
-                                                    Float* const __restrict__ p_wei_global,
-                                                    OutGlobalDesc,
-                                                    Float* __restrict__ p_out_global)
+gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw(const Float* const __restrict__ p_in_global,
+                                                    const Float* const __restrict__ p_wei_global,
+                                                    Float* const __restrict__ p_out_global)
 {
     constexpr auto I0 = Number<0>{};
     constexpr auto I1 = Number<1>{};
@@ -174,10 +171,10 @@ gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw(InGlobalDesc,
     // set threadwise output tensor to 0
     threadwise_2d_tensor_set_zero(out_kb_thread_desc, p_out_thread);
 
-    Float* p_in_global_block_offset =
+    const Float* p_in_global_block_offset =
         p_in_global + in_cb_global_desc.Get1dIndex(0, b_block_data_begin);
 
-    Float* p_wei_global_block_offset =
+    const Float* p_wei_global_block_offset =
         p_wei_global + wei_srck_global_desc.Get1dIndex(0, 0, 0, k_block_data_begin);
 
     for(unsigned c_block_data_begin = 0; c_block_data_begin < C; c_block_data_begin += CPerBlock,

src/include/gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw_lds_pipeline.cuh

@@ -25,12 +25,9 @@ template <unsigned GridSize,
           unsigned InBlockCopyThreadPerDim0,
           unsigned InBlockCopyThreadPerDim1>
 __global__ void gridwise_implicit_gemm_convolution_2_cnhw_srck_knhw_lds_pipeline(
-    InGlobalDesc,
-    Float* const __restrict__ p_in_global,
-    WeiGlobalDesc,
-    Float* const __restrict__ p_wei_global,
-    OutGlobalDesc,
-    Float* __restrict__ p_out_global)
+    const Float* const __restrict__ p_in_global,
+    const Float* const __restrict__ p_wei_global,
+    Float* const __restrict__ p_out_global)
 {
     constexpr auto I0 = Number<0>{};
     constexpr auto I1 = Number<1>{};

src/include/gridwise_winograd_convolution.cuh

@@ -19,12 +19,9 @@ template <class Float,
           unsigned CPerThread,
           unsigned BlockSize,
           unsigned GridSize>
-__global__ void gridwise_winograd_convolution(InGlobalDesc,
-                                              Float* const __restrict__ p_in_global,
-                                              WeiGlobalDesc,
-                                              Float* const __restrict__ p_wei_global,
-                                              OutGlobalDesc,
-                                              Float* __restrict__ p_out_global)
+__global__ void gridwise_winograd_convolution(const Float* const __restrict__ p_in_global,
+                                              const Float* const __restrict__ p_wei_global,
+                                              Float* const __restrict__ p_out_global)
 {
     constexpr auto I0 = Number<0>{};
     constexpr auto I1 = Number<1>{};

src/include/tensor.hpp

@@ -6,8 +6,6 @@
 #include <utility>
 #include <cassert>
 #include <iostream>
-#include "cuda_runtime.h"
-#include "helper_cuda.h"
 
 template <class Range>
 std::ostream& LogRange(std::ostream& os, Range&& r, std::string delim)
@@ -108,33 +106,6 @@ struct TensorDescriptor
     std::vector<std::size_t> mStrides;
 };
 
-struct DeviceMem
-{
-    DeviceMem() = delete;
-    DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
-    {
-        cudaMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize);
-    }
-
-    void* GetDeviceBuffer() { return mpDeviceBuf; }
-
-    int ToDevice(const void* p)
-    {
-        return static_cast<int>(
-            cudaMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, cudaMemcpyHostToDevice));
-    }
-
-    int FromDevice(void* p)
-    {
-        return static_cast<int>(cudaMemcpy(p, mpDeviceBuf, mMemSize, cudaMemcpyDeviceToHost));
-    }
-
-    ~DeviceMem() { cudaFree(mpDeviceBuf); }
-
-    void* mpDeviceBuf;
-    std::size_t mMemSize;
-};
-
 struct joinable_thread : std::thread
 {
     template <class... Xs>