save

.devcontainer/Dockerfile

@@ -40,7 +40,6 @@ RUN apt-get update && export DEBIAN_FRONTEND=noninteractive \
     netpbm \
     imagemagick \
     ghostscript \
-    nvidia-cuda-toolkit \
     && apt-get clean && rm -rf /var/lib/apt/lists/* \ 
     && ln -s /usr/bin/clang-$LLVM_VERSION /usr/bin/clang \
     && ln -s /usr/bin/clang++-$LLVM_VERSION /usr/bin/clang++ \
@@ -70,6 +69,17 @@ RUN wget https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cm
     && ln -s /opt/cmake-${CMAKE_VERSION}/bin/* /usr/local/bin \
     && rm /tmp/cmake-install.sh
 
+# Install CUDA
+RUN wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb \
+    && dpkg -i cuda-keyring_1.1-1_all.deb
+
+RUN apt-get update && export DEBIAN_FRONTEND=noninteractive \
+    && apt-get -y install --no-install-recommends \
+    cuda-toolkit-12-6
+
+# add CUDA to the path
+ENV PATH=/usr/local/cuda-12.6/bin:/usr/local/cuda-12.6:$PATH
+
 # Clean up
 RUN apt-get -y remove wget gnupg software-properties-common
 

CMakeLists.txt

@@ -212,3 +212,7 @@ if(SQUINT_BUILD_DOCUMENTATION)
     DESTINATION ${CMAKE_INSTALL_DOCDIR})
 
 endif()
+
+add_executable(main main.cpp)
+target_link_libraries(main PRIVATE SQUINT)
+include_directories(include)

include/squint/tensor/cuda/cuda_context.hpp

@@ -5,32 +5,32 @@
 
 namespace squint::cuda {
 
-class CudaContext {
+class cuda_context {
   public:
-    static auto instance() -> CudaContext & {
-        static CudaContext instance;
+    static auto instance() -> cuda_context & {
+        static cuda_context instance;
         return instance;
     }
 
     [[nodiscard]] auto cublas_handle() const -> cublasHandle_t { return cublas_handle_; }
 
     // Delete copy constructor and assignment operator
-    CudaContext(const CudaContext &) = delete;
-    auto operator=(const CudaContext &) -> CudaContext & = delete;
+    cuda_context(const cuda_context &) = delete;
+    auto operator=(const cuda_context &) -> cuda_context & = delete;
 
     // Delete move constructor and assignment operator
-    CudaContext(CudaContext &&) = delete;
-    auto operator=(CudaContext &&) -> CudaContext & = delete;
+    cuda_context(cuda_context &&) = delete;
+    auto operator=(cuda_context &&) -> cuda_context & = delete;
 
   private:
-    CudaContext() {
+    cuda_context() {
         cublasStatus_t status = cublasCreate(&cublas_handle_);
         if (status != CUBLAS_STATUS_SUCCESS) {
             throw std::runtime_error("Failed to create cuBLAS handle");
         }
     }
 
-    ~CudaContext() { cublasDestroy(cublas_handle_); }
+    ~cuda_context() { cublasDestroy(cublas_handle_); }
 
     cublasHandle_t cublas_handle_{};
 };

include/squint/tensor/cuda/element_wise.hpp

@@ -1,3 +1,4 @@
+#include <cstdint>
 #ifndef SQUINT_TENSOR_CUDA_ELEMENT_WISE_HPP
 
 template <typename T>
@@ -7,17 +8,21 @@ void element_wise_addition(T *output, const T *a, const T *b, const unsigned lon
 
 template <typename T>
 void element_wise_subtraction(T *output, const T *a, const T *b, const unsigned long *dims,
-                           const unsigned long *strides_out, const unsigned long *strides_a,
-                           const unsigned long *strides_b, unsigned long num_dims, unsigned long total_size);
+                              const unsigned long *strides_out, const unsigned long *strides_a,
+                              const unsigned long *strides_b, unsigned long num_dims, unsigned long total_size);
 
 template <typename T>
-void element_wise_equality(T *output, const T *a, const T *b, const unsigned long *dims,
+void element_wise_equality(uint8_t *output, const T *a, const T *b, const unsigned long *dims,
                            const unsigned long *strides_out, const unsigned long *strides_a,
                            const unsigned long *strides_b, unsigned long num_dims, unsigned long total_size);
 
 template <typename T>
-void element_wise_inequality(T *output, const T *a, const T *b, const unsigned long *dims,
-                           const unsigned long *strides_out, const unsigned long *strides_a,
-                           const unsigned long *strides_b, unsigned long num_dims, unsigned long total_size);
+void element_wise_inequality(uint8_t *output, const T *a, const T *b, const unsigned long *dims,
+                             const unsigned long *strides_out, const unsigned long *strides_a,
+                             const unsigned long *strides_b, unsigned long num_dims, unsigned long total_size);
+
+template <typename T>
+void element_wise_negation(T *output, const T *a, const unsigned long *dims, const unsigned long *strides_out,
+                           const unsigned long *strides_a, unsigned long num_dims, unsigned long total_size);
 
 #endif // SQUINT_TENSOR_CUDA_ELEMENT_WISE_HPP

include/squint/tensor/cuda/scalar.hpp

@@ -0,0 +1,7 @@
+#ifndef SQUINT_TENSOR_CUDA_SCALAR_HPP
+
+template <typename T>
+void scalar_multiplication(T scalar, T *output, const T *a, const unsigned long *dims, const unsigned long *strides_out,
+                           const unsigned long *strides_a, unsigned long num_dims, unsigned long total_size);
+
+#endif // SQUINT_TENSOR_CUDA_SCALAR_HPP