Gemm+Reduce Fusion (#128)

* add gridwise gemm v4r1

* rename

* adding gemm+reduce

* adding gemm+reduce

* adding gemm+reduce

* adding gemm+reduce

* use sfc in shuffling

* remove hardcode

* remove hardcode

* refactor

* fix build

* adding gemm+reduce

* adding gemm+reduce

* adding gemm+reduce

* adding gemm+reduce

* adding gemm+reduce

* format

* clean

* adding gemm+reduce

* adding profiler for gemm+reduce

* adding gemm+reduce profiler

* fix build

* clean up

* gemm+reduce

* fix build

* update DeviceGemm_Xdl_CShuffle; update enum to enum class

* clean up

* add test for gemm+reduce

* clean up

* refactor

* fix build

* fix build

[ROCm/composable_kernel commit: f95267f166]

library/src/host_tensor/device.cpp

@@ -7,6 +7,8 @@ DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
 
 void* DeviceMem::GetDeviceBuffer() { return mpDeviceBuf; }
 
+std::size_t DeviceMem::GetBufferSize() { return mMemSize; }
+
 void DeviceMem::ToDevice(const void* p)
 {
     hipGetErrorString(
@@ -18,6 +20,8 @@ void DeviceMem::FromDevice(void* p)
     hipGetErrorString(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
 }
 
+void DeviceMem::SetZero() { hipGetErrorString(hipMemset(mpDeviceBuf, 0, mMemSize)); }
+
 DeviceMem::~DeviceMem() { hipGetErrorString(hipFree(mpDeviceBuf)); }
 
 struct KernelTimerImpl

library/src/host_tensor/host_tensor.cpp

@@ -64,6 +64,8 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream
     os << "}" << std::endl;
 }
 
+#if 1
+// FIXME: remove
 float bf16_to_f32_(ck::bhalf_t src_val)
 {
     union
@@ -74,8 +76,10 @@ float bf16_to_f32_(ck::bhalf_t src_val)
     return u.fp32;
 }
 
+// FIXME: remove
 void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst)
 {
     for(int i = 0; i < src.mData.size(); ++i)
         dst.mData[i] = bf16_to_f32_(src.mData[i]);
 }
+#endif