constant memory optimization

#include <cmath>
#include <iostream>
#include "gpu-new-forward.h"

#define TILE_WIDTH 16
#define BLOCK_SIZE 512

inline void cudaCheckError(cudaError_t error, const char *file, int line) {
    if (error != cudaSuccess) {
        fprintf(stderr, "CUDA error at %s:%d code=%d(%s) \n",
                file, line, static_cast<unsigned int>(error),
                cudaGetErrorString(error));
        exit(EXIT_FAILURE);
    }
}

// Macro to make error checking easier to use
#define CUDA_CHECK(err) cudaCheckError(err, __FILE__, __LINE__)

// Add constant memory for the mask
// Maximum size calculation: typical maximum values might be
// Map_out=16, Channel=4, K=7 -> 16 * 4 * 7 * 7 = 3136 elements
__constant__ float const_mask[4000];  // Conservative size that should handle common cases

__global__ void matrix_unrolling_kernel(const float *input, float *output,
                                        const int Batch, const int Channel,
                                        const int Height, const int Width,
                                        const int K) {

    #define in_4d(i3, i2, i1, i0) input[(i3) * (Channel * Height * Width) + (i2) * (Height * Width) + (i1) * (Width) + i0]
    #define out_3d(i1, i0) output[(i1) * (Batch * W_unroll) + i0]

    // Calculate output dimensions
    const size_t Height_out = Height - K + 1;
    const size_t Width_out = Width - K + 1;
    const size_t W_unroll = Height_out * Width_out;
    const size_t H_unroll = Channel * K * K;
    const size_t W_total_unroll = Batch * W_unroll;

    // Calculate thread indices
    const size_t c = blockIdx.x * blockDim.x + threadIdx.x;
    const size_t hw_pos = blockIdx.y * blockDim.y + threadIdx.y;
    const size_t batch_idx = blockIdx.z * blockDim.z + threadIdx.z;

    // Extract height and width positions
    const size_t h_out = hw_pos / Width_out;
    const size_t w_out = hw_pos % Width_out;

    // Boundary check
    if (c >= Channel || h_out >= Height_out || w_out >= Width_out || batch_idx >= Batch) {
        return;
    }

    // Calculate position in unrolled matrix
    const size_t w_unroll = h_out * Width_out + w_out;
    const size_t w_total_unroll = batch_idx * W_unroll + w_unroll;
    const size_t w_base = c * K * K;

    // Perform unrolling
    for (int p = 0; p < K; p++) {
        for (int q = 0; q < K; q++) {
            int h_unroll = w_base + p * K + q;
            out_3d(h_unroll, w_total_unroll) = in_4d(batch_idx, c, h_out + p, w_out + q);
        }
    }

    #undef in_4d
    #undef out_3d
}

__global__ void matrixMultiplyShared(const float *B, float *C,
                                     int numARows, int numAColumns,
                                     int numBRows, int numBColumns,
                                     int numCRows, int numCColumns)
{
    // Declare shared memory for both matrices
    __shared__ float tileB[TILE_WIDTH][TILE_WIDTH];
    __shared__ float tileA[TILE_WIDTH][TILE_WIDTH];

    int by = blockIdx.y, bx = blockIdx.x, ty = threadIdx.y, tx = threadIdx.x;
    int row = by * TILE_WIDTH + ty, col = bx * TILE_WIDTH + tx;
    float val = 0;

    // Process tiles
    for (int tileId = 0; tileId < (numAColumns - 1) / TILE_WIDTH + 1; tileId++) {
        // Load mask from constant memory into shared memory
        if (row < numARows && tileId * TILE_WIDTH + tx < numAColumns) {
            tileA[ty][tx] = const_mask[row * numAColumns + tileId * TILE_WIDTH + tx];
        } else {
            tileA[ty][tx] = 0;
        }

        // Load tile from matrix B into shared memory
        if (col < numBColumns && tileId * TILE_WIDTH + ty < numBRows) {
            tileB[ty][tx] = B[((size_t) tileId * TILE_WIDTH + ty) * numBColumns + col];
        } else {
            tileB[ty][tx] = 0;
        }

        // Make sure all threads have loaded their data
        __syncthreads();

        // Compute partial dot product using shared memory
        if (row < numCRows && col < numCColumns) {
            for (int i = 0; i < TILE_WIDTH; i++) {
                val += tileA[ty][i] * tileB[i][tx];
            }
        }

        // Synchronize before next iteration
        __syncthreads();
    }

    // Write final result to global memory
    if (row < numCRows && col < numCColumns) {
        C[row * numCColumns + col] = val;
    }
}

__global__ void matrix_permute_kernel(const float *input, float *output, int Map_out,
                                      int Batch, int image_size) {
    int b = blockIdx.y;
    int x = blockIdx.x * BLOCK_SIZE + threadIdx.x;
    if (x < image_size) {
        for (int m = 0; m < Map_out; m++) {
            output[b * Map_out * image_size + m * image_size + x] =
                    input[m * Batch * image_size + b * image_size + x];
        }
    }
}

__host__ void GPUInterface::conv_forward_gpu_prolog(const float *host_output, const float *host_input, const float *host_mask, float **device_output_ptr, float **device_input_ptr, float **device_mask_ptr, const int Batch, const int Map_out, const int Channel, const int Height, const int Width, const int K)
{
    // Calculate sizes
    const int Height_out = Height - K + 1;
    const int Width_out = Width - K + 1;

    const int input_size = Batch * Channel * Height * Width * sizeof(float);
    const int mask_size = Map_out * Channel * K * K * sizeof(float);
    const int output_size = Batch * Map_out * Height_out * Width_out * sizeof(float);

    // Verify mask size
    if (Map_out * Channel * K * K > 4000) {
        std::cerr << "Error: Mask size exceeds constant memory allocation" << std::endl;
        exit(-1);
    }

    // Copy mask to constant memory instead of global memory
    cudaMemcpyToSymbol(const_mask, host_mask, mask_size);

    // Allocate memory for input and output
    cudaMalloc((void**)device_input_ptr, input_size);
    cudaMalloc((void**)device_output_ptr, output_size);

    // Copy input to device
    cudaMemcpy(*device_input_ptr, host_input, input_size, cudaMemcpyHostToDevice);

    // Set mask pointer to NULL since we're using constant memory
    *device_mask_ptr = NULL;
}

__host__ void GPUInterface::conv_forward_gpu(float *device_output, const float *device_input, const float *device_mask, const int Batch, const int Map_out, const int Channel, const int Height, const int Width, const int K)
{
    const int Height_out = Height - K + 1;
    const int Width_out = Width - K + 1;
    const int Height_unrolled = Channel * K * K;
    const int Width_unrolled = Batch * Height_out * Width_out;

    float *unrolled_matrix;
    float *matmul_output;
    cudaMalloc((void**)&unrolled_matrix, (size_t) Batch * Channel * K * K * Height_out * Width_out * sizeof(float));
    cudaMalloc((void**)&matmul_output, (Batch * Map_out * Height_out * Width_out) * sizeof(float));

    dim3 blockDim(4, 256, 1);
    dim3 gridDim(
        (Channel + blockDim.x - 1) / blockDim.x,
        (Height_out * Width_out + blockDim.y - 1) / blockDim.y,
        (Batch + blockDim.z - 1) / blockDim.z
    );

    matrix_unrolling_kernel<<<gridDim, blockDim>>>(
        device_input, unrolled_matrix,
        Batch, Channel, Height, Width, K
    );

    dim3 dimGrid((Width_unrolled - 1)/TILE_WIDTH + 1, (Map_out - 1)/TILE_WIDTH + 1, 1);
    dim3 dimBlock(TILE_WIDTH, TILE_WIDTH, 1);

    matrixMultiplyShared<<<dimGrid, dimBlock>>>(
        unrolled_matrix, matmul_output,
        Map_out, Height_unrolled,
        Height_unrolled, Width_unrolled,
        Map_out, Width_unrolled
    );

    const int out_image_size = Height_out * Width_out;
    dim3 permute_kernel_grid_dim((out_image_size - 1) / BLOCK_SIZE + 1, Batch, 1);
    matrix_permute_kernel<<<permute_kernel_grid_dim, BLOCK_SIZE>>>(
        matmul_output, device_output,
        Map_out, Batch, out_image_size
    );

    cudaFree(matmul_output);
    cudaFree(unrolled_matrix);
}

__host__ void GPUInterface::conv_forward_gpu_epilog(float *host_output, float *device_output, float *device_input, float *device_mask, const int Batch, const int Map_out, const int Channel, const int Height, const int Width, const int K)
{
    const int Height_out = Height - K + 1;
    const int Width_out = Width - K + 1;
    const int output_size = Batch * Map_out * Height_out * Width_out * sizeof(float);

    cudaMemcpy(host_output, device_output, output_size, cudaMemcpyDeviceToHost);

    cudaFree(device_output);
    cudaFree(device_input);
}

__host__ void GPUInterface::get_device_properties()
{
    int deviceCount;
    cudaGetDeviceCount(&deviceCount);

    for(int dev = 0; dev < deviceCount; dev++)
    {
        cudaDeviceProp deviceProp;
        cudaGetDeviceProperties(&deviceProp, dev);

        std::cout<<"Device "<<dev<<" name: "<<deviceProp.name<<std::endl;
        std::cout<<"Computational capabilities: "<<deviceProp.major<<"."<<deviceProp.minor<<std::endl;
        std::cout<<"Max Global memory size: "<<deviceProp.totalGlobalMem<<std::endl;
        std::cout<<"Max Constant memory size: "<<deviceProp.totalConstMem<<std::endl;
        std::cout<<"Max Shared memory size per block: "<<deviceProp.sharedMemPerBlock<<std::endl;
        std::cout<<"Max threads per block: "<<deviceProp.maxThreadsPerBlock<<std::endl;
        std::cout<<"Max block dimensions: "<<deviceProp.maxThreadsDim[0]<<" x, "<<deviceProp.maxThreadsDim[1]<<" y, "<<deviceProp.maxThreadsDim[2]<<" z"<<std::endl;
        std::cout<<"Max grid dimensions: "<<deviceProp.maxGridSize[0]<<" x, "<<deviceProp.maxGridSize[1]<<" y, "<<deviceProp.maxGridSize[2]<<" z"<<std::endl;
        std::cout<<"Warp Size: "<<deviceProp.warpSize<<std::endl;
    }
}