chundoong-lab-ta/APWS23/ans/matmul_buffering_ans.cu

#include <cstdio>

#include "matmul.h"

static __global__ void matmul_kernel(float *A, float *B, float *C, int M, int N,
                              int K) {
  int i = blockDim.x * blockIdx.x + threadIdx.x;
  int j = blockDim.y * blockIdx.y + threadIdx.y;
  if (i >= M || j >= N) return;
  float sum = 0.0;
  for (int k = 0; k < K; ++k) sum += A[i * K + k] * B[k * N + j];
  C[i * N + j] = sum;
}

#define BLOCKS 4

static size_t Mbegin[BLOCKS], Mend[BLOCKS];
static cudaStream_t data_stream, calc_stream;
static cudaEvent_t events[BLOCKS];
static float *A_gpu, *B_gpu, *C_gpu;

void matmul_buffering_initialize(size_t M, size_t N, size_t K) {
  for (size_t i = 0; i < BLOCKS; i++) {
    Mbegin[i] = M / BLOCKS * i;
    Mend[i] = M / BLOCKS * (i + 1);
    if (i == BLOCKS - 1) Mend[i] = M;
  }

  cudaStreamCreate(&data_stream);
  cudaStreamCreate(&calc_stream);
  for (int i = 0; i < BLOCKS; i++) { cudaEventCreate(&events[i]); }
  cudaMalloc(&A_gpu, M * K * sizeof(float));
  cudaMalloc(&B_gpu, K * N * sizeof(float));
  cudaMalloc(&C_gpu, M * N * sizeof(float));
}

void matmul_buffering(float *A, float *B, float *C, size_t M, size_t N,
                      size_t K) {
  cudaMemcpyAsync(B_gpu, B, K * N * sizeof(float), cudaMemcpyHostToDevice,
                  data_stream);
  for (int i = 0; i < BLOCKS; i++) {
    cudaMemcpyAsync(&A_gpu[Mbegin[i] * K], &A[Mbegin[i] * K],
                    (Mend[i] - Mbegin[i]) * K * sizeof(float),
                    cudaMemcpyHostToDevice, data_stream);
    cudaEventRecord(events[i], data_stream);
  }

  for (int i = 0; i < BLOCKS; i++) {
    dim3 blockDim(32, 32);
    dim3 gridDim((Mend[i] - Mbegin[i] + 32 - 1) / 32, (N + 32 - 1) / 32);

    cudaStreamWaitEvent(calc_stream, events[i]);
    matmul_kernel<<<gridDim, blockDim, 0, calc_stream>>>(&A_gpu[Mbegin[i] * K], B_gpu,
                                                        &C_gpu[Mbegin[i] * N], (Mend[i] - Mbegin[i]), N, K);
  }

  cudaStreamSynchronize(calc_stream);
  cudaMemcpyAsync(C, C_gpu, M * N * sizeof(float), cudaMemcpyDeviceToHost, data_stream);
  cudaStreamSynchronize(data_stream);
}

void matmul_buffering_finalize(size_t M, size_t N, size_t K) {
  cudaFree(A_gpu);
  cudaFree(B_gpu);
  cudaFree(C_gpu);
  cudaStreamDestroy(data_stream);
  cudaStreamDestroy(calc_stream);
  for (int i = 0; i < BLOCKS; i++) { cudaEventDestroy(events[i]); }
}