flatironinstitute · dfm · Oct 30, 2023 · Nov 8, 2021 · Nov 8, 2021 · Nov 9, 2021
diff --git a/.gitignore b/.gitignore
@@ -5,3 +5,4 @@ _skbuild
 dist
 MANIFEST
 __pycache__/
+*.egg-info
diff --git a/.gitmodules b/.gitmodules
@@ -1,3 +1,3 @@
 [submodule "finufft"]
 	path = vendor/finufft
-	url = https://github.com/flatironinstitute/finufft
+	url = https://github.com/lgarrison/finufft
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -2,8 +2,10 @@ cmake_minimum_required(VERSION 3.15)
 project(${SKBUILD_PROJECT_NAME} LANGUAGES C CXX)
 message(STATUS "Using CMake version: " ${CMAKE_VERSION})
 
-set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_LIST_DIR})
+# Add the /cmake directory to the module path so that we can find FFTW
+set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_LIST_DIR}/cmake)
 
+# Handle Python settings passed from scikit-build
 if(SKBUILD)
     set(Python_EXECUTABLE "${PYTHON_EXECUTABLE}")
     set(Python_INCLUDE_DIR "${PYTHON_INCLUDE_DIR}")
@@ -13,23 +15,29 @@ if(SKBUILD)
         OUTPUT_VARIABLE _tmp_dir
         OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ECHO STDOUT)
       list(APPEND CMAKE_PREFIX_PATH "${_tmp_dir}")
+else()
+    find_package(Python COMPONENTS Interpreter Development REQUIRED)
+    execute_process(
+        COMMAND "${Python_EXECUTABLE}" -c "import pybind11; print(pybind11.get_cmake_dir())"
+        OUTPUT_VARIABLE _tmp_dir
+        OUTPUT_STRIP_TRAILING_WHITESPACE COMMAND_ECHO STDOUT)
+      list(APPEND CMAKE_PREFIX_PATH "${_tmp_dir}")
 endif()
 
 set(PYBIND11_NEWPYTHON ON)
 find_package(pybind11 CONFIG REQUIRED)
 find_package(FFTW REQUIRED COMPONENTS FLOAT_LIB DOUBLE_LIB)
 link_libraries(${FFTW_FLOAT_LIB} ${FFTW_DOUBLE_LIB})
 
+# Work out compiler flags
 set(CMAKE_POSITION_INDEPENDENT_CODE ON)
-add_compile_options(-Wall -O3 -funroll-loops)
-
+add_compile_options(-Wall -Wno-unknown-pragmas -O3 -funroll-loops -fdiagnostics-color)
 set(FINUFFT_INCLUDE_DIRS
     ${CMAKE_CURRENT_LIST_DIR}/lib
     ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/include
     ${FFTW_INCLUDE_DIRS})
 
-message(STATUS "FINUFFT include dirs: " "${FINUFFT_INCLUDE_DIRS}")
-
+# Build single and double point versions of the FINUFFT library
 add_library(finufft STATIC
     ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/spreadinterp.cpp
     ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/utils.cpp
@@ -45,10 +53,72 @@ add_library(finufft_32 STATIC
 target_compile_definitions(finufft_32 PUBLIC SINGLE)
 target_include_directories(finufft_32 PRIVATE ${FINUFFT_INCLUDE_DIRS})
 
-pybind11_add_module(jax_finufft
-    ${CMAKE_CURRENT_LIST_DIR}/lib/jax_finufft.cc
+# Build the XLA bindings to those libraries
+pybind11_add_module(jax_finufft_cpu
+    ${CMAKE_CURRENT_LIST_DIR}/lib/jax_finufft_cpu.cc
     ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/utils_precindep.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/contrib/legendre_rule_fast.c)
-target_link_libraries(jax_finufft PRIVATE finufft finufft_32)
-target_include_directories(jax_finufft PRIVATE ${FINUFFT_INCLUDE_DIRS})
-install(TARGETS jax_finufft DESTINATION .)
+    ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/contrib/legendre_rule_fast.cpp)
+
+target_link_libraries(jax_finufft_cpu PRIVATE finufft finufft_32)
+target_include_directories(jax_finufft_cpu PRIVATE ${FINUFFT_INCLUDE_DIRS})
+install(TARGETS jax_finufft_cpu DESTINATION .)
+
+include(CheckLanguage)
+check_language(CUDA)
+if (CMAKE_CUDA_COMPILER)
+    enable_language(CUDA)
+    set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)
+    if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80;90")
+    endif()
+
+    # Find cufft
+    find_package(CUDAToolkit)
+
+    set(CUFINUFFT_INCLUDE_DIRS
+        ${CMAKE_CURRENT_LIST_DIR}/lib
+        # ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/include/cufinufft/contrib/
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/include/cufinufft/contrib/cuda_samples
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/include
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/contrib
+        ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+
+    set(CUFINUFFT_SOURCES
+        # TODO: 1D not supported via JAX, but needed for compilation
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/1d/spread1d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/1d/interp1d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/1d/cufinufft1d.cu
+
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/2d/spread2d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/2d/interp2d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/2d/cufinufft2d.cu
+
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/3d/spread3d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/3d/interp3d_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/3d/cufinufft3d.cu
+
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/deconvolve_wrapper.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/memtransfer_wrapper.cu
+
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/spreadinterp.cpp
+
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/precision_independent.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/contrib/legendre_rule_fast.cpp
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/utils.cpp
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/cufinufft.cu
+        ${CMAKE_CURRENT_LIST_DIR}/vendor/finufft/src/cuda/common.cu)
+
+    add_library(cufinufft STATIC ${CUFINUFFT_SOURCES})
+    target_include_directories(cufinufft PRIVATE ${CUFINUFFT_INCLUDE_DIRS})
+
+    pybind11_add_module(jax_finufft_gpu
+        ${CMAKE_CURRENT_LIST_DIR}/lib/jax_finufft_gpu.cc
+        ${CMAKE_CURRENT_LIST_DIR}/lib/kernels.cc.cu)
+    target_link_libraries(jax_finufft_gpu PRIVATE cufinufft)
+    target_link_libraries(jax_finufft_gpu PRIVATE ${CUDA_cufft_LIBRARY} ${CUDA_nvToolsExt_LIBRARY})
+    target_include_directories(jax_finufft_gpu PRIVATE ${CUFINUFFT_INCLUDE_DIRS})
+    install(TARGETS jax_finufft_gpu DESTINATION .)
+
+else()
+    message(STATUS "No CUDA compiler found; GPU support will be disabled")
+endif()
diff --git a/FindFFTW.cmake → cmake/FindFFTW.cmake b/FindFFTW.cmake → cmake/FindFFTW.cmake
diff --git a/lib/jax_finufft_common.h b/lib/jax_finufft_common.h
@@ -0,0 +1,21 @@
+#ifndef _JAX_FINUFFT_COMMON_H_
+#define _JAX_FINUFFT_COMMON_H_
+
+// This descriptor is common to both the jax_finufft and jax_finufft_gpu modules
+// We will use the jax_finufft namespace for both
+
+namespace jax_finufft {
+
+template <typename T>
+struct NufftDescriptor {
+  T eps;
+  int iflag;
+  int64_t n_tot;
+  int n_transf;
+  int64_t n_j;
+  int64_t n_k[3];
+};
+
+}
+
+#endif
diff --git a/lib/jax_finufft.cc → lib/jax_finufft_cpu.cc b/lib/jax_finufft.cc → lib/jax_finufft_cpu.cc
@@ -4,6 +4,8 @@
 
 #include "pybind11_kernel_helpers.h"
 
+#include "jax_finufft_cpu.h"
+
 using namespace jax_finufft;
 
 namespace {
@@ -15,7 +17,7 @@ void run_nufft(int type, void *desc_in, T *x, T *y, T *z, std::complex<T> *c, st
   int64_t n_k = 1;
   for (int d = 0; d < ndim; ++d) n_k *= descriptor->n_k[d];
 
-  nufft_opts *opts = new nufft_opts;
+  finufft_opts *opts = new finufft_opts;
   default_opts<T>(opts);
 
   typename plan_type<T>::type plan;
@@ -86,7 +88,7 @@ pybind11::dict Registrations() {
   return dict;
 }
 
-PYBIND11_MODULE(jax_finufft, m) {
+PYBIND11_MODULE(jax_finufft_cpu, m) {
   m.def("registrations", &Registrations);
   m.def("build_descriptorf", &build_descriptor<float>);
   m.def("build_descriptor", &build_descriptor<double>);

diff --git a/lib/jax_finufft.h → lib/jax_finufft_cpu.h b/lib/jax_finufft.h → lib/jax_finufft_cpu.h
@@ -7,16 +7,6 @@
 
 namespace jax_finufft {
 
-template <typename T>
-struct NufftDescriptor {
-  T eps;
-  int iflag;
-  int64_t n_tot;
-  int n_transf;
-  int64_t n_j;
-  int64_t n_k[3];
-};
-
 template <typename T>
 struct plan_type;
 
@@ -31,31 +21,31 @@ struct plan_type<float> {
 };
 
 template <typename T>
-void default_opts(nufft_opts* opts);
+void default_opts(finufft_opts* opts);
 
 template <>
-void default_opts<float>(nufft_opts* opts) {
+void default_opts<float>(finufft_opts* opts) {
   finufftf_default_opts(opts);
 }
 
 template <>
-void default_opts<double>(nufft_opts* opts) {
+void default_opts<double>(finufft_opts* opts) {
   finufft_default_opts(opts);
 }
 
 template <typename T>
 int makeplan(int type, int dim, int64_t* nmodes, int iflag, int ntr, T eps,
-             typename plan_type<T>::type* plan, nufft_opts* opts);
+             typename plan_type<T>::type* plan, finufft_opts* opts);
 
 template <>
 int makeplan<float>(int type, int dim, int64_t* nmodes, int iflag, int ntr, float eps,
-                    typename plan_type<float>::type* plan, nufft_opts* opts) {
+                    typename plan_type<float>::type* plan, finufft_opts* opts) {
   return finufftf_makeplan(type, dim, nmodes, iflag, ntr, eps, plan, opts);
 }
 
 template <>
 int makeplan<double>(int type, int dim, int64_t* nmodes, int iflag, int ntr, double eps,
-                     typename plan_type<double>::type* plan, nufft_opts* opts) {
+                     typename plan_type<double>::type* plan, finufft_opts* opts) {
   return finufft_makeplan(type, dim, nmodes, iflag, ntr, eps, plan, opts);
 }
 

diff --git a/lib/jax_finufft_gpu.cc b/lib/jax_finufft_gpu.cc
@@ -0,0 +1,37 @@
+// This file defines the Python interface to the XLA custom call implemented on the CPU.
+// It is exposed as a standard pybind11 module defining "capsule" objects containing our
+// method. For simplicity, we export a separate capsule for each supported dtype.
+
+#include "pybind11_kernel_helpers.h"
+#include "kernels.h"
+
+using namespace jax_finufft;
+
+namespace {
+
+pybind11::dict Registrations() {
+  pybind11::dict dict;
+
+  // TODO: do we prefer to keep these names the same as the CPU version or prefix them with "cu"?
+  // dict["nufft1d1f"] = encapsulate_function(nufft1d1f);
+  // dict["nufft1d2f"] = encapsulate_function(nufft1d2f);
+  dict["nufft2d1f"] = encapsulate_function(nufft2d1f);
+  dict["nufft2d2f"] = encapsulate_function(nufft2d2f);
+  dict["nufft3d1f"] = encapsulate_function(nufft3d1f);
+  dict["nufft3d2f"] = encapsulate_function(nufft3d2f);
+
+  // dict["nufft1d1"] = encapsulate_function(nufft1d1);
+  // dict["nufft1d2"] = encapsulate_function(nufft1d2);
+  dict["nufft2d1"] = encapsulate_function(nufft2d1);
+  dict["nufft2d2"] = encapsulate_function(nufft2d2);
+  dict["nufft3d1"] = encapsulate_function(nufft3d1);
+  dict["nufft3d2"] = encapsulate_function(nufft3d2);
+
+  return dict;
+}
+
+PYBIND11_MODULE(jax_finufft_gpu, m) {
+  m.def("registrations", &Registrations);
+}
+
+}  // namespace