core/kernels/optimized_matmul.cu

#include <cuda_fp16.h>
#include <cuda_runtime.h>
#include <torch/extension.h>
#include <mma.h>  // Tensor Core WMMA API

using namespace nvcuda;

// Define FP16 type for Tensor Cores
using half_t = __half;

// WMMA tile sizes (fixed for Tensor Cores)
#define WMMA_M 16
#define WMMA_N 16
#define WMMA_K 16
#define BLOCK_SIZE 32

// Optimized GEMM kernel using Tensor Cores
__global__ void optimized_matmul_kernel(
    const half_t* __restrict__ a,    // Matrix A [m, k]
    const half_t* __restrict__ b,    // Matrix B [k, n]
    half_t* __restrict__ c,          // Matrix C [m, n]
    int m, int n, int k)             // Dimensions: A[m,k], B[k,n], C[m,n]
{
    // Shared memory for WMMA tiles
    __shared__ half_t shmem_a[WMMA_M * WMMA_K];
    __shared__ half_t shmem_b[WMMA_K * WMMA_N];

    // WMMA fragments
    wmma::fragment<wmma::matrix_a, WMMA_M, WMMA_N, WMMA_K, half_t, wmma::row_major> a_frag;
    wmma::fragment<wmma::matrix_b, WMMA_M, WMMA_N, WMMA_K, half_t, wmma::col_major> b_frag;
    wmma::fragment<wmma::accumulator, WMMA_M, WMMA_N, WMMA_K, half_t> c_frag;

    // Thread indices
    int tx = threadIdx.x;
    int ty = threadIdx.y;
    int bx = blockIdx.x;
    int by = blockIdx.y;

    // Global tile offsets
    int row = by * WMMA_M + ty;  // Row in C
    int col = bx * WMMA_N + tx;  // Column in C

    // Initialize accumulator
    wmma::fill_fragment(c_frag, __float2half(0.0f));

    // Loop over K dimension in WMMA tiles
    for (int tile_k = 0; tile_k < k; tile_k += WMMA_K) {
        // Load A tile into shared memory (row-major)
        if (row < m && tile_k + tx < k) {
            shmem_a[ty * WMMA_K + tx] = a[row * k + tile_k + tx];
        } else {
            shmem_a[ty * WMMA_K + tx] = __float2half(0.0f);
        }

        // Load B tile into shared memory (col-major)
        if (tile_k + ty < k && col < n) {
            shmem_b[ty * WMMA_N + tx] = b[(tile_k + ty) * n + col];
        } else {
            shmem_b[ty * WMMA_N + tx] = __float2half(0.0f);
        }

        __syncthreads();

        // Load WMMA fragments from shared memory
        wmma::load_matrix_sync(a_frag, shmem_a, WMMA_K);
        wmma::load_matrix_sync(b_frag, shmem_b, WMMA_N);

        // Perform Tensor Core matrix multiply-accumulate
        wmma::mma_sync(c_frag, a_frag, b_frag, c_frag);

        __syncthreads();
    }

    // Store result to global memory
    if (row < m && col < n) {
        wmma::store_matrix_sync(&c[row * n + col], c_frag, n, wmma::mem_row_major);
    }
}

// PyTorch binding
torch::Tensor optimized_matmul(
    torch::Tensor a,  // [m, k]
    torch::Tensor b)  // [k, n]
{
    // Ensure inputs are FP16 and on CUDA
    TORCH_CHECK(a.dtype() == torch::kFloat16, "Matrix A must be FP16");
    TORCH_CHECK(b.dtype() == torch::kFloat16, "Matrix B must be FP16");
    TORCH_CHECK(a.is_cuda(), "Matrix A must be on CUDA");
    TORCH_CHECK(b.is_cuda(), "Matrix B must be on CUDA");
    TORCH_CHECK(a.dim() == 2 && b.dim() == 2, "Inputs must be 2D tensors");
    TORCH_CHECK(a.size(1) == b.size(0), "Inner dimensions must match");

    // Dimensions
    int m = a.size(0);
    int k = a.size(1);
    int n = b.size(1);

    // Output tensor
    auto c = torch::empty({m, n}, 
                          torch::TensorOptions().dtype(torch::kFloat16).device(a.device()));

    // Grid and block dimensions
    dim3 block(BLOCK_SIZE, WMMA_M / WARP_SIZE);  // 32 threads per warp, WMMA_M/32 warps
    dim3 grid((n + WMMA_N - 1) / WMMA_N, (m + WMMA_M - 1) / WMMA_M);

    // Launch kernel
    optimized_matmul_kernel<<<grid, block>>>(
        (half_t*)a.data_ptr(),
        (half_t*)b.data_ptr(),
        (half_t*)c.data_ptr(),
        m, n, k);

    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess) {
        TORCH_CHECK(false, "CUDA error: ", cudaGetErrorString(err));
    }

    return c;
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("optimized_matmul", &optimized_matmul, "Tensor Core-optimized GEMM");
}