Merge pull request #67 from lucasdelimanogueira/tmp

fix sum cuda axis atomicAdd

norch/csrc/cuda.cu

@@ -141,40 +141,27 @@ __global__ void sum_tensor_cuda_kernel(float* data, float* result_data, int size
     }
 }
 
-__global__ void sum_tensor_axis_cuda_kernel(float* data, float* result_data, int size, int axis_size) {
-    __shared__ float partial_sum[THREADS_PER_BLOCK];
+__global__ void sum_tensor_cuda_kernel_axis(float* data, float* result_data, int* strides, int target_axis, int inner_size, int outer_size) {
+    int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
-    int tid = threadIdx.x;
-    int block_offset = blockIdx.x * axis_size;
-    int i = block_offset + tid;
+    if (tid < outer_size) {
+        int outer_index = tid / inner_size;
+        int inner_index = tid % inner_size;
 
-    partial_sum[tid] = 0.0f;
-    while (i < block_offset + axis_size && i < size) {
-        partial_sum[tid] += data[i];
-        i += blockDim.x;
-    }
+        int offset = outer_index * strides[0] + inner_index;
 
-    __syncthreads();
-
-    // Perform block-wise reduction
-    for (int s = blockDim.x / 2; s > 0; s >>= 1) {
-        if (tid < s && i < block_offset + axis_size) {
-            partial_sum[tid] += partial_sum[tid + s];
+        for (int i = 0; i < strides[target_axis]; ++i) {
+            int index = offset + i * strides[target_axis + 1];
+            atomicAdd(&result_data[outer_index * inner_size + inner_index], data[index]);
         }
-        __syncthreads();
-    }
-
-    // Write block sum to global memory
-    if (tid == 0) {
-        result_data[blockIdx.x] = partial_sum[0];
     }
 }
 
 
 __host__ void sum_tensor_cuda(Tensor* tensor, float* result_data, int axis) {
-    cudaMemcpy(result_data, tensor->data, tensor->size * sizeof(float), cudaMemcpyHostToDevice);
 
     if (axis == -1) {
+        cudaMemcpy(result_data, tensor->data, tensor->size * sizeof(float), cudaMemcpyHostToDevice);
         
         int num_blocks = (tensor->size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
 
@@ -197,19 +184,20 @@ __host__ void sum_tensor_cuda(Tensor* tensor, float* result_data, int axis) {
         cudaDeviceSynchronize();
         
     } else {
-        int axis_size = tensor->shape[axis];
 
-        int num_blocks = (tensor->size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+        int target_axis_stride = tensor->strides[axis];
+        int inner_size = tensor->strides[axis + 1];
+        int outer_size = tensor->size / target_axis_stride;
 
-        // First-level reduction
-        sum_tensor_axis_cuda_kernel<<<num_blocks, THREADS_PER_BLOCK>>>(result_data, result_data, tensor->size, axis_size);
+        int* d_strides;
+        cudaMalloc(&d_strides, (tensor->ndim + 1) * sizeof(int));
+        cudaMemcpy(d_strides, tensor->strides, (tensor->ndim + 1) * sizeof(int), cudaMemcpyHostToDevice);
 
-        // If necessary, perform multiple levels of reduction
-        while (num_blocks > 1) {
-            int num_blocks_next = (num_blocks + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
-            sum_tensor_cuda_kernel<<<num_blocks_next, THREADS_PER_BLOCK>>>(result_data, result_data, num_blocks);
-            num_blocks = num_blocks_next;
-        }
+        cudaMemset(result_data, 0, outer_size * sizeof(float));
+
+        int num_threads = outer_size;
+        int num_blocks = (num_threads + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+        sum_tensor_cuda_kernel_axis<<<num_blocks, THREADS_PER_BLOCK>>>(tensor->data, result_data, d_strides, axis, inner_size, outer_size);
 
         cudaError_t error = cudaGetLastError();
         if (error != cudaSuccess) {
@@ -218,6 +206,9 @@ __host__ void sum_tensor_cuda(Tensor* tensor, float* result_data, int axis) {
         }
 
         cudaDeviceSynchronize();
+
+        // Free allocated memory
+        cudaFree(d_strides);
     }
 }
 

norch/csrc/cuda.h

@@ -14,7 +14,7 @@
     __host__ void sub_broadcasted_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_data, int* broadcasted_shape, int broadcasted_size);
 
     __global__ void sum_tensor_cuda_kernel(float* data, float* result_data);
-    __global__ void sum_tensor_axis_cuda_kernel(float* data, float* result_data, int size, int axis_size);
+    __global__ void sum_tensor_cuda_kernel_axis(float* data, float* result_data, int* strides, int target_axis, int inner_size, int outer_size);
     __host__ void sum_tensor_cuda(Tensor* tensor, float* result_data, int axis);
 
     __global__ void sub_tensor_cuda_kernel(float* data1, float* data2, float* result_data, int size);

norch/csrc/tensor.cpp

@@ -210,7 +210,15 @@ extern "C" {
         if (strcmp(tensor->device, "cuda") == 0) {
             
             float* result_data;
-            cudaMalloc((void**)&result_data, tensor->size * sizeof(float));
+            if (axis == -1) {
+                cudaMalloc((void**)&result_data, tensor->size * sizeof(float));
+            } else {
+
+                int target_axis_stride = tensor->strides[axis];
+                int outer_size = tensor->size / target_axis_stride;
+
+                cudaMalloc((void**)&result_data, (outer_size) * sizeof(float));
+            }
             sum_tensor_cuda(tensor, result_data, axis);
             
             if (keepdim) {