Trying a bare cuda vector add against pytorch and triton

bench.ipynb

Makefile

CUDA_PATH ?= /usr/local/cuda

.PHONY: clean

vadd.so: vadd.o
	nvcc -shared $^ -o $@ -lcuda

vadd.o: vadd.cu
	nvcc -I $(CUDA_PATH)/include -I$(CUDA_PATH)/samples/common/inc -arch=sm_70 --compiler-options '-fPIC' $^ -c $@

clean:
	rm -f *.o *.so

vadd.cu

// For the CUDA runtime routines (prefixed with "cuda_")
// #include <cuda.h>
#include <cuda_runtime.h>

namespace
{
    __global__ void _vadd(const float *A, const float *B, float *C, int n)
    {
        int i = blockDim.x * blockIdx.x + threadIdx.x;
        if (i < n)
        {
            C[i] = A[i] + B[i];
        }
    }
}

extern "C" void vadd(const float *A, const float *B, float *C, int n, int threads)
{
    const int blocks = (n + threads - 1) / threads;
    _vadd<<<blocks, threads>>>(A, B, C, n);
}