The AI CUDA Engineer 👷

import torch
import torch.nn as nn
import torch.nn.functional as F


def module_fn(
    x: torch.Tensor,
    conv1_weight: nn.Parameter,
    conv1_bias: nn.Parameter,
    conv2_weight: nn.Parameter,
    conv2_bias: nn.Parameter,
    fc1_weight: nn.Parameter,
    fc1_bias: nn.Parameter,
    fc2_weight: nn.Parameter,
    fc2_bias: nn.Parameter,
    fc3_weight: nn.Parameter,
    fc3_bias: nn.Parameter,
) -> torch.Tensor:
    """
    Implements a LeNet-5 architecture with ReLU activation.

    Args:
        x (torch.Tensor): The input tensor, shape (batch_size, 1, 32, 32)
        conv1_weight (nn.Parameter): Parameters for first conv layer
        conv1_bias (nn.Parameter): Parameters for first conv layer
        conv2_weight (nn.Parameter): Parameters for second conv layer
        conv2_bias (nn.Parameter): Parameters for second conv layer
        fc1_weight (nn.Parameter): Parameters for first FC layer
        fc1_bias (nn.Parameter): Parameters for first FC layer
        fc2_weight (nn.Parameter): Parameters for second FC layer
        fc3_weight (nn.Parameter): Parameters for third FC layer
        fc3_bias (nn.Parameter): Parameters for third FC layer

    Returns:
        torch.Tensor: The output tensor, shape (batch_size, num_classes)
    """
    # First convolutional layer with ReLU activation and max pooling
    x = F.conv2d(x, conv1_weight, conv1_bias, stride=1)
    x = F.relu(x)
    x = F.max_pool2d(x, kernel_size=2, stride=2)

    # Second convolutional layer with ReLU activation and max pooling
    x = F.conv2d(x, conv2_weight, conv2_bias, stride=1)
    x = F.relu(x)
    x = F.max_pool2d(x, kernel_size=2, stride=2)

    # Flatten the output for the fully connected layers
    x = x.view(-1, 16 * 5 * 5)

    # First fully connected layer with ReLU activation
    x = F.linear(x, fc1_weight, fc1_bias)
    x = F.relu(x)

    # Second fully connected layer with ReLU activation
    x = F.linear(x, fc2_weight, fc2_bias)
    x = F.relu(x)

    # Final fully connected layer
    x = F.linear(x, fc3_weight, fc3_bias)

    return x


class Model(nn.Module):
    def __init__(self, num_classes):
        """
        LeNet-5 architecture implementation in PyTorch.

        :param num_classes: The number of output classes.
        """
        super(Model, self).__init__()

        # Extract parameters from convolutional layers
        conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5, stride=1)
        self.conv1_weight = nn.Parameter(conv1.weight.data.clone())
        self.conv1_bias = nn.Parameter(conv1.bias.data.clone())

        conv2 = nn.Conv2d(in_channels=6, out_channels=16, kernel_size=5, stride=1)
        self.conv2_weight = nn.Parameter(conv2.weight.data.clone())
        self.conv2_bias = nn.Parameter(conv2.bias.data.clone())

        # Extract parameters from fully connected layers
        fc1 = nn.Linear(in_features=16 * 5 * 5, out_features=120)
        self.fc1_weight = nn.Parameter(fc1.weight.data.clone())
        self.fc1_bias = nn.Parameter(fc1.bias.data.clone())

        fc2 = nn.Linear(in_features=120, out_features=84)
        self.fc2_weight = nn.Parameter(fc2.weight.data.clone())
        self.fc2_bias = nn.Parameter(fc2.bias.data.clone())

        fc3 = nn.Linear(in_features=84, out_features=num_classes)
        self.fc3_weight = nn.Parameter(fc3.weight.data.clone())
        self.fc3_bias = nn.Parameter(fc3.bias.data.clone())

    def forward(self, x, fn=module_fn):
        return fn(
            x,
            self.conv1_weight,
            self.conv1_bias,
            self.conv2_weight,
            self.conv2_bias,
            self.fc1_weight,
            self.fc1_bias,
            self.fc2_weight,
            self.fc2_bias,
            self.fc3_weight,
            self.fc3_bias,
        )


# Test code for the LeNet-5 model
batch_size = 1
num_classes = 10


def get_inputs():
    return [torch.randn(batch_size, 1, 32, 32)]


def get_init_inputs():
    return [num_classes]

#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <cublas_v2.h>
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/CUDAUtils.h>

// ReLU kernel with balanced workload distribution
__global__ void relu_kernel(float* input, int size) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = gridDim.x * blockDim.x;
    for (int i = idx; i < size; i += stride) {
        input[i] = fmaxf(0.0f, input[i]);
    }
}

// Max pooling kernel with balanced workload distribution
__global__ void max_pool2d_kernel(
    const float* __restrict__ input, float* __restrict__ output,
    int batch_size, int channels, int height, int width,
    int pool_height, int pool_width, int stride
) {
    int out_h = (height - pool_height) / stride + 1;
    int out_w = (width - pool_width) / stride + 1;
    int total_elements = batch_size * channels * out_h * out_w;
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride_threads = blockDim.x * gridDim.x;

    for (; idx < total_elements; idx += stride_threads) {
        int b = idx / (channels * out_h * out_w);
        int c = (idx / (out_h * out_w)) % channels;
        int h = (idx / out_w) % out_h;
        int w = idx % out_w;

        int in_h_start = h * stride;
        int in_w_start = w * stride;

        float max_val = -FLT_MAX;
        for (int i = in_h_start; i < in_h_start + pool_height; ++i) {
            for (int j = in_w_start; j < in_w_start + pool_width; ++j) {
                float val = input[((b * channels + c) * height + i) * width + j];
                max_val = fmaxf(max_val, val);
            }
        }
        output[idx] = max_val;
    }
}

// Linear layer kernel with balanced workload distribution
__global__ void linear_kernel(
    const float* __restrict__ input,
    const float* __restrict__ weight,
    const float* __restrict__ bias,
    float* __restrict__ output,
    int in_features, int out_features
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;

    for (; idx < out_features; idx += stride) {
        float val = bias[idx];
        for (int i = 0; i < in_features; ++i) {
            val += input[i] * weight[idx * in_features + i];
        }
        output[idx] = val;
    }
}

// Forward function for the LeNet-5 architecture
torch::Tensor forward(
    torch::Tensor x,
    torch::Tensor conv1_weight, torch::Tensor conv1_bias,
    torch::Tensor conv2_weight, torch::Tensor conv2_bias,
    torch::Tensor fc1_weight, torch::Tensor fc1_bias,
    torch::Tensor fc2_weight, torch::Tensor fc2_bias,
    torch::Tensor fc3_weight, torch::Tensor fc3_bias
) {
    // Ensure inputs are on CUDA
    x = x.to(torch::kCUDA);
    conv1_weight = conv1_weight.to(torch::kCUDA);
    conv1_bias = conv1_bias.to(torch::kCUDA);
    conv2_weight = conv2_weight.to(torch::kCUDA);
    conv2_bias = conv2_bias.to(torch::kCUDA);
    fc1_weight = fc1_weight.to(torch::kCUDA);
    fc1_bias = fc1_bias.to(torch::kCUDA);
    fc2_weight = fc2_weight.to(torch::kCUDA);
    fc2_bias = fc2_bias.to(torch::kCUDA);
    fc3_weight = fc3_weight.to(torch::kCUDA);
    fc3_bias = fc3_bias.to(torch::kCUDA);

    const int block_size = 256;
    const int num_blocks = (x.numel() + block_size - 1) / block_size;

    // First convolutional layer
    auto conv1 = torch::conv2d(x, conv1_weight, conv1_bias, {1, 1});
    relu_kernel<<<num_blocks, block_size>>>(conv1.data_ptr<float>(), conv1.numel());
    auto pool1 = torch::max_pool2d(conv1, {2, 2}, {2, 2});

    // Second convolutional layer
    auto conv2 = torch::conv2d(pool1, conv2_weight, conv2_bias, {1, 1});
    relu_kernel<<<num_blocks, block_size>>>(conv2.data_ptr<float>(), conv2.numel());
    auto pool2 = torch::max_pool2d(conv2, {2, 2}, {2, 2});

    // Flatten the output
    auto flat = pool2.view({pool2.size(0), -1});

    // First fully connected layer
    auto fc1 = torch::linear(flat, fc1_weight, fc1_bias);
    relu_kernel<<<num_blocks, block_size>>>(fc1.data_ptr<float>(), fc1.numel());

    // Second fully connected layer
    auto fc2 = torch::linear(fc1, fc2_weight, fc2_bias);
    relu_kernel<<<num_blocks, block_size>>>(fc2.data_ptr<float>(), fc2.numel());

    // Final fully connected layer
    auto fc3 = torch::linear(fc2, fc3_weight, fc3_bias);

    return fc3;
}

// PyBind11 module definition
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("forward", &forward, "LeNet-5 forward pass");
}