update the DCU compile (#5563)

A-006 · Qianruipku · web-flow · commit 0fd832eb9001 · 2024-11-26T10:00:42.000+08:00
* update the duc compile

* update the FFT_CUDA in the fft_bundle.cpp

* update the rcom to rocm

---------

Co-authored-by: Qianrui Liu &lt;76200646+Qianruipku@users.noreply.github.com&gt;
diff --git a/source/module_base/module_device/rocm/memory_op.hip.cu b/source/module_base/module_device/rocm/memory_op.hip.cu
@@ -219,6 +219,8 @@ template struct cast_memory_op<std::complex<double>,
                                std::complex<float>,
                                base_device::DEVICE_GPU,
                                base_device::DEVICE_GPU>;
+template struct cast_memory_op<std::complex<float>, float, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
+template struct cast_memory_op<std::complex<double>, double, base_device::DEVICE_GPU, base_device::DEVICE_GPU>;
 template struct cast_memory_op<float, float, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
 template struct cast_memory_op<double, double, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
 template struct cast_memory_op<float, double, base_device::DEVICE_GPU, base_device::DEVICE_CPU>;
diff --git a/source/module_basis/module_pw/CMakeLists.txt b/source/module_basis/module_pw/CMakeLists.txt
@@ -10,7 +10,7 @@ if (USE_CUDA)
 endif()
 if (USE_ROCM)
   list (APPEND FFT_SRC
-    module_fft/fft_rcom.cpp
+    module_fft/fft_rocm.cpp
   )
 endif()
 
diff --git a/source/module_basis/module_pw/module_fft/fft_bundle.cpp b/source/module_basis/module_pw/module_fft/fft_bundle.cpp
@@ -6,7 +6,7 @@
 #include "fft_cuda.h"
 #endif
 #if defined(__ROCM)
-#include "fft_rcom.h"
+#include "fft_rocm.h"
 #endif
 
 template<typename FFT_BASE, typename... Args>
@@ -89,9 +89,9 @@ void FFT_Bundle::initfft(int nx_in,
     if (device=="gpu")
     {
         #if defined(__ROCM)
-            fft_float = new FFT_RCOM<float>();
+            fft_float = make_unique<FFT_ROCM<float>>();
             fft_float->initfft(nx_in,ny_in,nz_in);
-            fft_double = new FFT_RCOM<double>();
+            fft_double = make_unique<FFT_ROCM<double>>();
             fft_double->initfft(nx_in,ny_in,nz_in);
         #elif defined(__CUDA)
             fft_float = make_unique<FFT_CUDA<float>>();
diff --git a/source/module_basis/module_pw/module_fft/fft_rocm.cpp b/source/module_basis/module_pw/module_fft/fft_rocm.cpp
@@ -1,10 +1,10 @@
-#include "fft_rcom.h"
+#include "fft_rocm.h"
 #include "module_base/module_device/memory_op.h"
 #include "module_hamilt_pw/hamilt_pwdft/global.h"
 namespace ModulePW
 {
 template <typename FPTYPE>
-void FFT_RCOM<FPTYPE>::initfft(int nx_in, 
+void FFT_ROCM<FPTYPE>::initfft(int nx_in, 
                                int ny_in, 
                                int nz_in)
 {
@@ -13,20 +13,20 @@ void FFT_RCOM<FPTYPE>::initfft(int nx_in,
     this->nz = nz_in;
 }
 template <>
-void FFT_RCOM<float>::setupFFT()
+void FFT_ROCM<float>::setupFFT()
 {
     hipfftPlan3d(&c_handle, this->nx, this->ny, this->nz, HIPFFT_C2C);
     resmem_cd_op()(gpu_ctx, this->c_auxr_3d, this->nx * this->ny * this->nz);
         
 }
 template <>  
-void FFT_RCOM<double>::setupFFT()
+void FFT_ROCM<double>::setupFFT()
 {
     hipfftPlan3d(&z_handle, this->nx, this->ny, this->nz, HIPFFT_Z2Z);
     resmem_zd_op()(gpu_ctx, this->z_auxr_3d, this->nx * this->ny * this->nz);
 }
 template <>
-void FFT_RCOM<float>::cleanFFT()
+void FFT_ROCM<float>::cleanFFT()
 {
     if (c_handle)
     {
@@ -35,7 +35,7 @@ void FFT_RCOM<float>::cleanFFT()
     }
 }
 template <>
-void FFT_RCOM<double>::cleanFFT()
+void FFT_ROCM<double>::cleanFFT()
 {
     if (z_handle)
     {
@@ -44,7 +44,7 @@ void FFT_RCOM<double>::cleanFFT()
     }
 }
 template <>
-void FFT_RCOM<float>::clear()
+void FFT_ROCM<float>::clear()
 {
     this->cleanFFT();
     if (c_auxr_3d != nullptr)
@@ -54,7 +54,7 @@ void FFT_RCOM<float>::clear()
     }
 }
 template <>
-void FFT_RCOM<double>::clear()
+void FFT_ROCM<double>::clear()
 {
     this->cleanFFT();
     if (z_auxr_3d != nullptr)
@@ -64,7 +64,7 @@ void FFT_RCOM<double>::clear()
     }
 }
 template <>
-void FFT_RCOM<float>::fft3D_forward(std::complex<float>* in, 
+void FFT_ROCM<float>::fft3D_forward(std::complex<float>* in, 
                                     std::complex<float>* out) const
 {
     CHECK_CUFFT(hipfftExecC2C(this->c_handle, 
@@ -73,7 +73,7 @@ void FFT_RCOM<float>::fft3D_forward(std::complex<float>* in,
                               HIPFFT_FORWARD));
 }
 template <>
-void FFT_RCOM<double>::fft3D_forward(std::complex<double>* in, 
+void FFT_ROCM<double>::fft3D_forward(std::complex<double>* in, 
                                      std::complex<double>* out) const
 {
     CHECK_CUFFT(hipfftExecZ2Z(this->z_handle, 
@@ -82,7 +82,7 @@ void FFT_RCOM<double>::fft3D_forward(std::complex<double>* in,
                               HIPFFT_FORWARD));
 }
 template <>
-void FFT_RCOM<float>::fft3D_backward(std::complex<float>* in, 
+void FFT_ROCM<float>::fft3D_backward(std::complex<float>* in, 
                                      std::complex<float>* out) const
 {
     CHECK_CUFFT(hipfftExecC2C(this->c_handle, 
@@ -91,7 +91,7 @@ void FFT_RCOM<float>::fft3D_backward(std::complex<float>* in,
                               HIPFFT_BACKWARD));
 }
 template <>
-void FFT_RCOM<double>::fft3D_backward(std::complex<double>* in, 
+void FFT_ROCM<double>::fft3D_backward(std::complex<double>* in, 
                                       std::complex<double>* out) const
 {
     CHECK_CUFFT(hipfftExecZ2Z(this->z_handle, 
@@ -100,7 +100,11 @@ void FFT_RCOM<double>::fft3D_backward(std::complex<double>* in,
                               HIPFFT_BACKWARD));
 }
 template <> std::complex<float>* 
-FFT_RCOM<float>::get_auxr_3d_data()  const {return this->c_auxr_3d;}
+FFT_ROCM<float>::get_auxr_3d_data()  const {return this->c_auxr_3d;}
 template <> std::complex<double>* 
-FFT_RCOM<double>::get_auxr_3d_data() const {return this->z_auxr_3d;}
+FFT_ROCM<double>::get_auxr_3d_data() const {return this->z_auxr_3d;}
+template FFT_ROCM<float>::FFT_ROCM();
+template FFT_ROCM<float>::~FFT_ROCM();
+template FFT_ROCM<double>::FFT_ROCM();
+template FFT_ROCM<double>::~FFT_ROCM();
 }// namespace ModulePW
diff --git a/source/module_basis/module_pw/module_fft/fft_rocm.h b/source/module_basis/module_pw/module_fft/fft_rocm.h
@@ -57,9 +57,5 @@ class FFT_ROCM : public FFT_BASE<FPTYPE>
         mutable std::complex<double>* z_auxr_3d = nullptr; // fft space
 
 };
-template FFT_RCOM<float>::FFT_RCOM();
-template FFT_ROCM<float>::~FFT_ROCM();
-template FFT_RCOM<double>::FFT_RCOM();
-template FFT_ROCM<double>::~FFT_ROCM();
 }// namespace ModulePW
 #endif

Original file line number	Diff line number	Diff line change
`@@ -10,7 +10,7 @@ if (USE_CUDA)`
`10`	`10`	`endif()`
`11`	`11`	`if (USE_ROCM)`
`12`	`12`	`list (APPEND FFT_SRC`
`13`		`- module_fft/fft_rcom.cpp`
	`13`	`+ module_fft/fft_rocm.cpp`
`14`	`14`	`)`
`15`	`15`	`endif()`
`16`	`16`
Original file line number	Diff line number	Diff line change
`@@ -1,10 +1,10 @@`
`1`		`-#include "fft_rcom.h"`
	`1`	`+#include "fft_rocm.h"`
`2`	`2`	`#include "module_base/module_device/memory_op.h"`
`3`	`3`	`#include "module_hamilt_pw/hamilt_pwdft/global.h"`
`4`	`4`	`namespace ModulePW`
`5`	`5`	`{`
`6`	`6`	`template <typename FPTYPE>`
`7`		`-void FFT_RCOM<FPTYPE>::initfft(int nx_in,`
	`7`	`+void FFT_ROCM<FPTYPE>::initfft(int nx_in,`
`8`	`8`	`int ny_in,`
`9`	`9`	`int nz_in)`
`10`	`10`	`{`
`@@ -13,20 +13,20 @@ void FFT_RCOM<FPTYPE>::initfft(int nx_in,`
`13`	`13`	`this->nz = nz_in;`
`14`	`14`	`}`
`15`	`15`	`template <>`
`16`		`-void FFT_RCOM<float>::setupFFT()`
	`16`	`+void FFT_ROCM<float>::setupFFT()`
`17`	`17`	`{`
`18`	`18`	`hipfftPlan3d(&c_handle, this->nx, this->ny, this->nz, HIPFFT_C2C);`
`19`	`19`	`resmem_cd_op()(gpu_ctx, this->c_auxr_3d, this->nx * this->ny * this->nz);`
`20`	`20`
`21`	`21`	`}`
`22`	`22`	`template <>`
`23`		`-void FFT_RCOM<double>::setupFFT()`
	`23`	`+void FFT_ROCM<double>::setupFFT()`
`24`	`24`	`{`
`25`	`25`	`hipfftPlan3d(&z_handle, this->nx, this->ny, this->nz, HIPFFT_Z2Z);`
`26`	`26`	`resmem_zd_op()(gpu_ctx, this->z_auxr_3d, this->nx * this->ny * this->nz);`
`27`	`27`	`}`
`28`	`28`	`template <>`
`29`		`-void FFT_RCOM<float>::cleanFFT()`
	`29`	`+void FFT_ROCM<float>::cleanFFT()`
`30`	`30`	`{`
`31`	`31`	`if (c_handle)`
`32`	`32`	`{`
`@@ -35,7 +35,7 @@ void FFT_RCOM<float>::cleanFFT()`
`35`	`35`	`}`
`36`	`36`	`}`
`37`	`37`	`template <>`
`38`		`-void FFT_RCOM<double>::cleanFFT()`
	`38`	`+void FFT_ROCM<double>::cleanFFT()`
`39`	`39`	`{`
`40`	`40`	`if (z_handle)`
`41`	`41`	`{`
`@@ -44,7 +44,7 @@ void FFT_RCOM<double>::cleanFFT()`
`44`	`44`	`}`
`45`	`45`	`}`
`46`	`46`	`template <>`
`47`		`-void FFT_RCOM<float>::clear()`
	`47`	`+void FFT_ROCM<float>::clear()`
`48`	`48`	`{`
`49`	`49`	`this->cleanFFT();`
`50`	`50`	`if (c_auxr_3d != nullptr)`
`@@ -54,7 +54,7 @@ void FFT_RCOM<float>::clear()`
`54`	`54`	`}`
`55`	`55`	`}`
`56`	`56`	`template <>`
`57`		`-void FFT_RCOM<double>::clear()`
	`57`	`+void FFT_ROCM<double>::clear()`
`58`	`58`	`{`
`59`	`59`	`this->cleanFFT();`
`60`	`60`	`if (z_auxr_3d != nullptr)`
`@@ -64,7 +64,7 @@ void FFT_RCOM<double>::clear()`
`64`	`64`	`}`
`65`	`65`	`}`
`66`	`66`	`template <>`
`67`		`-void FFT_RCOM<float>::fft3D_forward(std::complex<float>* in,`
	`67`	`+void FFT_ROCM<float>::fft3D_forward(std::complex<float>* in,`
`68`	`68`	`std::complex<float>* out) const`
`69`	`69`	`{`
`70`	`70`	`CHECK_CUFFT(hipfftExecC2C(this->c_handle,`
`@@ -73,7 +73,7 @@ void FFT_RCOM<float>::fft3D_forward(std::complex<float>* in,`
`73`	`73`	`HIPFFT_FORWARD));`
`74`	`74`	`}`
`75`	`75`	`template <>`
`76`		`-void FFT_RCOM<double>::fft3D_forward(std::complex<double>* in,`
	`76`	`+void FFT_ROCM<double>::fft3D_forward(std::complex<double>* in,`
`77`	`77`	`std::complex<double>* out) const`
`78`	`78`	`{`
`79`	`79`	`CHECK_CUFFT(hipfftExecZ2Z(this->z_handle,`
`@@ -82,7 +82,7 @@ void FFT_RCOM<double>::fft3D_forward(std::complex<double>* in,`
`82`	`82`	`HIPFFT_FORWARD));`
`83`	`83`	`}`
`84`	`84`	`template <>`
`85`		`-void FFT_RCOM<float>::fft3D_backward(std::complex<float>* in,`
	`85`	`+void FFT_ROCM<float>::fft3D_backward(std::complex<float>* in,`
`86`	`86`	`std::complex<float>* out) const`
`87`	`87`	`{`
`88`	`88`	`CHECK_CUFFT(hipfftExecC2C(this->c_handle,`
`@@ -91,7 +91,7 @@ void FFT_RCOM<float>::fft3D_backward(std::complex<float>* in,`
`91`	`91`	`HIPFFT_BACKWARD));`
`92`	`92`	`}`
`93`	`93`	`template <>`
`94`		`-void FFT_RCOM<double>::fft3D_backward(std::complex<double>* in,`
	`94`	`+void FFT_ROCM<double>::fft3D_backward(std::complex<double>* in,`
`95`	`95`	`std::complex<double>* out) const`
`96`	`96`	`{`
`97`	`97`	`CHECK_CUFFT(hipfftExecZ2Z(this->z_handle,`
`@@ -100,7 +100,11 @@ void FFT_RCOM<double>::fft3D_backward(std::complex<double>* in,`
`100`	`100`	`HIPFFT_BACKWARD));`
`101`	`101`	`}`
`102`	`102`	`template <> std::complex<float>*`
`103`		`-FFT_RCOM<float>::get_auxr_3d_data() const {return this->c_auxr_3d;}`
	`103`	`+FFT_ROCM<float>::get_auxr_3d_data() const {return this->c_auxr_3d;}`
`104`	`104`	`template <> std::complex<double>*`
`105`		`-FFT_RCOM<double>::get_auxr_3d_data() const {return this->z_auxr_3d;}`
	`105`	`+FFT_ROCM<double>::get_auxr_3d_data() const {return this->z_auxr_3d;}`
	`106`	`+template FFT_ROCM<float>::FFT_ROCM();`
	`107`	`+template FFT_ROCM<float>::~FFT_ROCM();`
	`108`	`+template FFT_ROCM<double>::FFT_ROCM();`
	`109`	`+template FFT_ROCM<double>::~FFT_ROCM();`
`106`	`110`	`}// namespace ModulePW`