Reorganize CUDA dependencies

dirac-institute · Mar 11, 2024 · c134561 · c134561
1 parent 185d299
commit c134561
Show file tree

Hide file tree

Showing 8 changed files with 168 additions and 74 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -71,8 +71,9 @@ target_link_libraries(search PRIVATE
 if(HAVE_CUDA)
   message(STATUS "Building CUDA Libraries")
   add_library(searchcu STATIC
-      src/kbmod/search/image_kernels.cu
-      src/kbmod/search/kernels.cu
+      src/kbmod/search/kernels/image_kernels.cu
+      src/kbmod/search/kernels/kernel_memory.cu
+      src/kbmod/search/kernels/kernels.cu
   )
 
   set_target_properties(searchcu PROPERTIES

diff --git a/src/kbmod/search/cuda_errors.h → src/kbmod/search/kernels/cuda_errors.h b/src/kbmod/search/cuda_errors.h → src/kbmod/search/kernels/cuda_errors.h
diff --git a/src/kbmod/search/image_kernels.cu → src/kbmod/search/kernels/image_kernels.cu b/src/kbmod/search/image_kernels.cu → src/kbmod/search/kernels/image_kernels.cu
@@ -9,7 +9,7 @@
 #define IMAGE_KERNELS_CU_
 
 #include <assert.h>
-#include "common.h"
+#include "../common.h"
 #include "cuda_errors.h"
 #include <stdio.h>
 #include <float.h>

diff --git a/src/kbmod/search/kernels/kernel_memory.cu b/src/kbmod/search/kernels/kernel_memory.cu
@@ -0,0 +1,109 @@
+/*
+ * kernel_memory.cu
+ *
+ * Helper functions for transfering KBMOD data to/from GPU.
+ */
+
+#ifndef KERNELS_MEMORY_CU_
+#define KERNELS_MEMORY_CU_
+
+#include <stdexcept>
+#include <float.h>
+#include <vector>
+
+#include "cuda_errors.h"
+
+#include "../trajectory_list.h"
+
+namespace search {
+
+
+// ---------------------------------------
+// --- Basic Memory Functions ------------
+// ---------------------------------------
+
+extern "C" void* allocate_gpu_block(unsigned long memory_size) {
+    void *gpu_ptr;
+    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, memory_size));
+    if (gpu_ptr == nullptr) throw std::runtime_error("Unable to allocate GPU memory.");
+    return gpu_ptr;
+}
+
+extern "C" void free_gpu_block(void* gpu_ptr) {
+    if (gpu_ptr == nullptr) throw std::runtime_error("Trying to free nullptr.");
+    checkCudaErrors(cudaFree(gpu_ptr));
+}
+
+extern "C" void copy_block_to_gpu(void* cpu_ptr, void* gpu_ptr, unsigned long memory_size) {
+    if (cpu_ptr == nullptr) throw std::runtime_error("Invalid CPU pointer");
+    if (gpu_ptr == nullptr) throw std::runtime_error("Invalid GPU pointer");
+
+    checkCudaErrors(cudaMemcpy(gpu_ptr, cpu_ptr, memory_size, cudaMemcpyHostToDevice));
+}
+
+extern "C" void copy_block_to_cpu(void* cpu_ptr, void* gpu_ptr, unsigned long memory_size) {
+    if (cpu_ptr == nullptr) throw std::runtime_error("Invalid CPU pointer");
+    if (gpu_ptr == nullptr) throw std::runtime_error("Invalid GPU pointer");
+
+    checkCudaErrors(cudaMemcpy(cpu_ptr, gpu_ptr, memory_size, cudaMemcpyDeviceToHost));
+}
+
+// ---------------------------------------
+// --- Memory Functions ------------------
+// ---------------------------------------
+
+
+extern "C" float *move_floats_to_gpu(std::vector<float> &data) {
+    unsigned long memory_size = data.size() * sizeof(float);
+
+    float *gpu_ptr;
+    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, memory_size));
+    checkCudaErrors(cudaMemcpy(gpu_ptr, data.data(), memory_size, cudaMemcpyHostToDevice));
+
+    return gpu_ptr;
+}
+
+extern "C" void free_gpu_float_array(float *gpu_ptr) {
+    if (gpu_ptr == nullptr) throw std::runtime_error("Trying to free nullptr.");
+    checkCudaErrors(cudaFree(gpu_ptr));
+}
+
+extern "C" void *move_void_array_to_gpu(void *data_array, long unsigned memory_size) {
+    if (data_array == nullptr) throw std::runtime_error("No data given.");
+    if (memory_size == 0) throw std::runtime_error("Invalid size.");
+
+    void *gpu_ptr;
+    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, memory_size));
+    checkCudaErrors(cudaMemcpy(gpu_ptr, data_array, memory_size, cudaMemcpyHostToDevice));
+
+    return gpu_ptr;
+}
+
+extern "C" void free_gpu_void_array(void *gpu_ptr) {
+    if (gpu_ptr == nullptr) throw std::runtime_error("Trying to free nullptr.");
+    checkCudaErrors(cudaFree(gpu_ptr));
+}
+
+extern "C" Trajectory *allocate_gpu_trajectory_list(long unsigned num_trj) {
+    Trajectory *gpu_ptr;
+    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, num_trj * sizeof(Trajectory)));
+    return gpu_ptr;
+}
+
+extern "C" void free_gpu_trajectory_list(Trajectory *gpu_ptr) { checkCudaErrors(cudaFree(gpu_ptr)); }
+
+extern "C" void copy_trajectory_list(Trajectory *cpu_ptr, Trajectory *gpu_ptr, long unsigned num_trj,
+                                     bool to_gpu) {
+    if ((cpu_ptr == nullptr) || (gpu_ptr == nullptr)) throw std::runtime_error("Invalid pointer.");
+    long unsigned memory_size = num_trj * sizeof(Trajectory);
+
+    if (to_gpu) {
+        checkCudaErrors(cudaMemcpy(gpu_ptr, cpu_ptr, memory_size, cudaMemcpyHostToDevice));
+    } else {
+        checkCudaErrors(cudaMemcpy(cpu_ptr, gpu_ptr, memory_size, cudaMemcpyDeviceToHost));
+    }
+}
+
+} /* namespace search */
+
+#endif /* KERNELS_MEMORY_CU_ */
diff --git a/src/kbmod/search/kernels/kernel_memory.h b/src/kbmod/search/kernels/kernel_memory.h
@@ -0,0 +1,44 @@
+/*
+ * kernel_memory.h
+ *
+ * Helper functions for transfering KBMOD data to/from GPU.
+ */
+
+#ifndef KERNELS_MEMORY_H_
+#define KERNELS_MEMORY_H_
+
+#include "../trajectory_list.h"
+
+namespace search {
+
+// ---------------------------------------
+// --- Basic Memory Functions ------------
+// ---------------------------------------
+
+extern "C" void* allocate_gpu_block(unsigned long memory_size);
+
+extern "C" void free_gpu_block(void* gpu_ptr);
+
+extern "C" void copy_block_to_gpu(void* cpu_ptr, void* gpu_ptr, unsigned long memory_size);
+
+extern "C" void copy_block_to_cpu(void* cpu_ptr, void* gpu_ptr, unsigned long memory_size);
+
+// ---------------------------------------
+// --- Memory Functions ------------------
+// ---------------------------------------
+
+extern "C" float *move_floats_to_gpu(std::vector<float> &data);
+extern "C" void free_gpu_float_array(float *gpu_ptr);
+
+extern "C" void *move_void_array_to_gpu(void *data_array, long unsigned memory_size);
+extern "C" void free_gpu_void_array(void *gpu_ptr);
+
+extern "C" Trajectory *allocate_gpu_trajectory_list(long unsigned num_trj);
+extern "C" void free_gpu_trajectory_list(Trajectory *gpu_ptr);
+extern "C" void copy_trajectory_list(Trajectory *cpu_ptr, Trajectory *gpu_ptr, long unsigned num_trj,
+                                     bool to_gpu);
+
+
+} /* namespace search */
+
+#endif /* KERNELS_MEMORY_CU_ */
diff --git a/src/kbmod/search/kernels.cu → src/kbmod/search/kernels/kernels.cu b/src/kbmod/search/kernels.cu → src/kbmod/search/kernels/kernels.cu
@@ -16,67 +16,14 @@
 #include <stdio.h>
 #include <float.h>
 
-#include "common.h"
+#include "../common.h"
 #include "cuda_errors.h"
-#include "psi_phi_array_ds.h"
-#include "trajectory_list.h"
+#include "../psi_phi_array_ds.h"
+#include "../trajectory_list.h"
 
-namespace search {
-
-// ---------------------------------------
-// --- Memory Functions ------------------
-// ---------------------------------------
-
-extern "C" float *move_floats_to_gpu(std::vector<float> &data) {
-    unsigned long memory_size = data.size() * sizeof(float);
-
-    float *gpu_ptr;
-    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, memory_size));
-    checkCudaErrors(cudaMemcpy(gpu_ptr, data.data(), memory_size, cudaMemcpyHostToDevice));
-
-    return gpu_ptr;
-}
-
-extern "C" void free_gpu_float_array(float *gpu_ptr) {
-    if (gpu_ptr == nullptr) throw std::runtime_error("Trying to free nullptr.");
-    checkCudaErrors(cudaFree(gpu_ptr));
-}
-
-extern "C" void *move_void_array_to_gpu(void *data_array, long unsigned memory_size) {
-    if (data_array == nullptr) throw std::runtime_error("No data given.");
-    if (memory_size == 0) throw std::runtime_error("Invalid size.");
-
-    void *gpu_ptr;
-    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, memory_size));
-    checkCudaErrors(cudaMemcpy(gpu_ptr, data_array, memory_size, cudaMemcpyHostToDevice));
+#include "kernel_memory.h"
 
-    return gpu_ptr;
-}
-
-extern "C" void free_gpu_void_array(void *gpu_ptr) {
-    if (gpu_ptr == nullptr) throw std::runtime_error("Trying to free nullptr.");
-    checkCudaErrors(cudaFree(gpu_ptr));
-}
-
-extern "C" Trajectory *allocate_gpu_trajectory_list(long unsigned num_trj) {
-    Trajectory *gpu_ptr;
-    checkCudaErrors(cudaMalloc((void **)&gpu_ptr, num_trj * sizeof(Trajectory)));
-    return gpu_ptr;
-}
-
-extern "C" void free_gpu_trajectory_list(Trajectory *gpu_ptr) { checkCudaErrors(cudaFree(gpu_ptr)); }
-
-extern "C" void copy_trajectory_list(Trajectory *cpu_ptr, Trajectory *gpu_ptr, long unsigned num_trj,
-                                     bool to_gpu) {
-    if ((cpu_ptr == nullptr) || (gpu_ptr == nullptr)) throw std::runtime_error("Invalid pointer.");
-    long unsigned memory_size = num_trj * sizeof(Trajectory);
-
-    if (to_gpu) {
-        checkCudaErrors(cudaMemcpy(gpu_ptr, cpu_ptr, memory_size, cudaMemcpyHostToDevice));
-    } else {
-        checkCudaErrors(cudaMemcpy(cpu_ptr, gpu_ptr, memory_size, cudaMemcpyDeviceToHost));
-    }
-}
+namespace search {
 
 // ---------------------------------------
 // --- Data Access Functions -------------

diff --git a/src/kbmod/search/psi_phi_array.cpp b/src/kbmod/search/psi_phi_array.cpp
@@ -2,16 +2,13 @@
 #include "psi_phi_array_utils.h"
 #include "pydocs/psi_phi_array_docs.h"
 
-namespace search {
-
 // Declaration of CUDA functions that will be linked in.
 #ifdef HAVE_CUDA
-extern "C" float* move_floats_to_gpu(std::vector<float>& data);
-extern "C" void free_gpu_float_array(float* gpu_ptr);
-extern "C" void* move_void_array_to_gpu(void* data_array, long unsigned memory_size);
-extern "C" void free_gpu_void_array(void* gpu_ptr);
+#include "kernels/kernel_memory.h"
 #endif
 
+namespace search {
+
 // -------------------------------------------------------
 // --- Implementation of core data structure functions ---
 // -------------------------------------------------------

diff --git a/src/kbmod/search/trajectory_list.cpp b/src/kbmod/search/trajectory_list.cpp
@@ -4,17 +4,13 @@
 #include <parallel/algorithm>
 #include <algorithm>
 
-namespace search {
-
 // Declaration of CUDA functions that will be linked in.
 #ifdef HAVE_CUDA
-extern "C" Trajectory *allocate_gpu_trajectory_list(long unsigned num_trj);
+#include "kernels/kernel_memory.h"
+#endif
 
-extern "C" void free_gpu_trajectory_list(Trajectory *gpu_ptr);
 
-extern "C" void copy_trajectory_list(Trajectory *cpu_ptr, Trajectory *gpu_ptr, long unsigned num_trj,
-                                     bool to_gpu);
-#endif
+namespace search {
 
 // -------------------------------------------------------
 // --- Implementation of core data structure functions ---