batched matmul 2D and 3D tensors

norch/csrc/cpu.cpp

@@ -45,6 +45,26 @@ void matmul_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data) {
     }
 }
 
+
+void batched_matmul_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data) {
+
+    int tensor2_offset = tensor2->shape[1] * tensor2->shape[2];
+    int result_data_offset = tensor1->shape[0] * tensor2->shape[2];
+
+    for (int batch = 0; batch < tensor2->shape[0]; batch++) {
+    
+        for (int i = 0; i < tensor1->shape[0]; i++) {
+            for (int j = 0; j < tensor2->shape[2]; j++) {
+                float sum = 0.0;
+                for (int k = 0; k < tensor1->shape[1]; k++) {
+                    sum += tensor1->data[i * tensor1->shape[1] + k] * tensor2->data[batch*tensor2_offset + (k * tensor2->shape[2] + j)];
+                }
+                result_data[(batch * result_data_offset) + (i * tensor2->shape[2] + j)] = sum;
+            }
+        }
+    }
+} 
+
 void pow_tensor_cpu(Tensor* tensor, float power, float* result_data) {
     
     for (int i = 0; i < tensor->size; i++) {

norch/csrc/cpu.h

@@ -8,6 +8,7 @@ void sum_tensor_cpu(Tensor* tensor1, float* result_data);
 void sub_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data);
 void elementwise_mul_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data);
 void matmul_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data);
+void batched_matmul_tensor_cpu(Tensor* tensor1, Tensor* tensor2, float* result_data);
 void pow_tensor_cpu(Tensor* tensor, float power, float* result_data);
 void scalar_mul_tensor_cpu(Tensor* tensor, float scalar, float* result_data);
 void ones_like_tensor_cpu(Tensor* tensor, float* result_data);

norch/csrc/cuda.cu

@@ -5,7 +5,6 @@
 
 #define THREADS_PER_BLOCK 128
 #define TILE_SIZE 32
-#define SHMEM_SIZE THREADS_PER_BLOCK * sizeof(float)
 
 __host__ void cpu_to_cuda(Tensor* tensor) {
     
@@ -61,7 +60,7 @@ __host__ void add_tensor_cuda(Tensor* tensor1, Tensor* tensor2, float* result_da
 
 
 __global__ void sum_tensor_cuda_kernel(float* data, float* result_data, int size) {
-    __shared__ float partial_sum[SHMEM_SIZE];
+    __shared__ float partial_sum[THREADS_PER_BLOCK * sizeof(float)];
 
     int tid = threadIdx.x;
     int i = blockIdx.x * blockDim.x + threadIdx.x;

norch/csrc/tensor.cpp

@@ -327,9 +327,10 @@ extern "C" {
     }
 
     Tensor* matmul_tensor(Tensor* tensor1, Tensor* tensor2) {
+        //MxN @ NxP = MxP
         // Check if tensors have compatible shapes for matrix multiplication
         if (tensor1->shape[1] != tensor2->shape[0]) {
-            fprintf(stderr, "Incompatible shapes for matrix multiplication\n");
+            fprintf(stderr, "Incompatible shapes for matrix multiplication %dx%d and %dx%d\n", tensor1->shape[0], tensor1->shape[1], tensor2->shape[0], tensor2->shape[1]);
             exit(1);
         }
 
@@ -387,6 +388,69 @@ extern "C" {
         }
     }
 
+    Tensor* batched_matmul_tensor(Tensor* tensor1, Tensor* tensor2) {
+        //MxN @ BATCHxNxP = BATCHxMxP
+        // Check if tensors have compatible shapes for matrix multiplication
+        if (tensor1->shape[1] != tensor2->shape[1]) {
+            fprintf(stderr, "Incompatible shapes for matrix multiplication %dx%d and %dx%d\n", tensor1->shape[0], tensor1->shape[1], tensor2->shape[0], tensor2->shape[1]);
+            exit(1);
+        }
+
+        if (strcmp(tensor1->device, tensor2->device) != 0) {
+            fprintf(stderr, "Tensors must be on the same device: %s and %s\n", tensor1->device, tensor2->device);
+            exit(1);
+        }
+
+        char* device = (char*)malloc(strlen(tensor1->device) + 1);
+        if (device != NULL) {
+            strcpy(device, tensor1->device);
+        } else {
+            fprintf(stderr, "Memory allocation failed\n");
+            exit(-1);
+        }
+
+        int ndim = 3;
+        int* shape = (int*)malloc(ndim * sizeof(int));
+        if (shape == NULL) {
+            fprintf(stderr, "Memory allocation failed\n");
+            exit(1);
+        }
+        
+        shape[0] = tensor2->shape[0];;
+        shape[1] = tensor1->shape[0];
+        shape[2] = tensor2->shape[2];  
+
+        int size = 1;
+        for (int i = 0; i < ndim; i++) {
+            size *= shape[i];
+        }
+
+        float* result_data = (float*)malloc(size * sizeof(float));
+        if (result_data == NULL) {
+            fprintf(stderr, "Memory allocation failed\n");
+            exit(1);
+        }
+
+        if (strcmp(tensor1->device, "cuda") == 0) {
+
+            float* result_data;
+            cudaMalloc((void **)&result_data, size * sizeof(float));
+            matmul_tensor_cuda(tensor1, tensor2, result_data);
+            return create_tensor(result_data, shape, ndim, device);
+        } 
+        else {
+            float* result_data = (float*)malloc(size * sizeof(float));
+            if (result_data == NULL) {
+                fprintf(stderr, "Memory allocation failed\n");
+                exit(1);
+            }
+            batched_matmul_tensor_cpu(tensor1, tensor2, result_data);
+            return create_tensor(result_data, shape, ndim, device);
+        }
+
+    }
+
+
     Tensor* pow_tensor(Tensor* tensor, float power) {
         char* device = (char*)malloc(strlen(tensor->device) + 1);
         if (device != NULL) {

norch/tensor.py

@@ -291,22 +291,38 @@ class Tensor:
         return self  
     
     def __matmul__(self, other):
-        if self.ndim != 2 or other.ndim != 2:
-            raise ValueError("Matrix multiplication requires 2D tensors")
+        if other.ndim == 3:
+            #batched 3D matmul
 
-        if self.shape[1] != other.shape[0]:
-            raise ValueError("Incompatible shapes for matrix multiplication")
+            Tensor._C.batched_matmul_tensor.argtypes = [ctypes.POINTER(CTensor), ctypes.POINTER(CTensor)]
+            Tensor._C.batched_matmul_tensor.restype = ctypes.POINTER(CTensor)
+            
+            result_tensor_ptr = Tensor._C.batched_matmul_tensor(self.tensor, other.tensor)
 
-        Tensor._C.matmul_tensor.argtypes = [ctypes.POINTER(CTensor), ctypes.POINTER(CTensor)]
-        Tensor._C.matmul_tensor.restype = ctypes.POINTER(CTensor)
+            result_data = Tensor()
+            result_data.tensor = result_tensor_ptr
+            result_data.shape = [other.shape[0], self.shape[0], other.shape[2]]
+            result_data.ndim = 3
+            result_data.device = self.device
+            
+        else:
+            #2D matmul
+            if self.ndim != 2 or other.ndim != 2:
+                raise ValueError("Matrix multiplication requires 2D tensors")
 
-        result_tensor_ptr = Tensor._C.matmul_tensor(self.tensor, other.tensor)
+            if self.shape[1] != other.shape[0]:
+                raise ValueError("Incompatible shapes for matrix multiplication")
 
-        result_data = Tensor()
-        result_data.tensor = result_tensor_ptr
-        result_data.shape = [self.shape[0], other.shape[1]]
-        result_data.ndim = 2
-        result_data.device = self.device
+            Tensor._C.matmul_tensor.argtypes = [ctypes.POINTER(CTensor), ctypes.POINTER(CTensor)]
+            Tensor._C.matmul_tensor.restype = ctypes.POINTER(CTensor)
+
+            result_tensor_ptr = Tensor._C.matmul_tensor(self.tensor, other.tensor)
+
+            result_data = Tensor()
+            result_data.tensor = result_tensor_ptr
+            result_data.shape = [self.shape[0], other.shape[1]]
+            result_data.ndim = 2
+            result_data.device = self.device
 
         result_data.requires_grad = self.requires_grad or other.requires_grad
         if result_data.requires_grad:

test.py

@@ -20,8 +20,25 @@ if __name__ == "__main__":
     import random
     import numpy as np
 
+    a = norch.Tensor([
+        [[1.234, 2.123], [3.635, 4.456], [5.678, 6.789]],
+        [[7.890, 8.901], [9.012, 1.234], [2.345, 3.456]],
+        [[4.567, 5.678], [6.789, 7.890], [8.901, 9.012]],
+        [[1.234, 2.345], [3.456, 4.567], [5.678, 6.789]],
+        [[7.890, 8.901], [9.012, 1.234], [2.345, 3.456]]
+    ])
+
+    b = norch.Tensor([
+        [1.234, 2.123, 1.5],
+        [5.678, 6.789, 1.293],
+        [3.635, 4.456, 1.0202],
+        [7.890, 8.901, 1.91],
+    ])
+
+    result = b @ a
+    print(result)
     
-    #a = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")
+    #a = norch.Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")
     #b = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")
     #c = Tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3]])#.to("cuda")
 
@@ -37,7 +54,6 @@ if __name__ == "__main__":
     print(a.grad)"""
     
     """#print(a)
-    """
     N = 10
     a = norch.Tensor([[1 for _ in range(N)] for _ in range(N)])
     #b = norch.Tensor([[random.uniform(0, 1) for _ in range(N)] for _ in range(N)])
@@ -56,7 +72,7 @@ if __name__ == "__main__":
     print("\n\n")
 
     """
-    
+    """
     a = [[random.uniform(0, 1) for _ in range(N)] for _ in range(N)]
     b = [[random.uniform(0, 1) for _ in range(N)] for _ in range(N)]
     ini = time.time()