Skip to content

Commit

Permalink
Fixed bmm.cu, updated block size.
Browse files Browse the repository at this point in the history
  • Loading branch information
@linjames0
linjames0 authored Sep 24, 2023
1 parent 6517fae commit 7654bc4
Showing 1 changed file with 105 additions and 90 deletions.
195 changes: 105 additions & 90 deletions bmm.cu
View file
Original file line number Diff line number Diff line change
@@ -1,104 +1,119 @@
#include <stdio.h>
#include <stdlib.h>

__global__ void bmm(float *d_A, float *d_B, float *d_C, int batch_size int M, int N, int P) {
__global__ void bmm(float *d_A, float *d_B, float *d_C, int batch_size, int M, int N, int P) {
int row = blockIdx.y * blockDim.y + threadIdx.y;
int col = blockIdx.x * blockDim.x + threadIdx.x;
int batch = blockIdx.z * blockDim.z + threadIdx.z;

if(batch < batch_size && row < M && col < P) {
float sum = 0.0f;

// compute the dot product for each row of A and col of B
for(int i = 0; i < N; ++i) {
sum += d_A[batch * M * N + row * N + i] * d_B[batch * N * P + i * P + col];
}
d_C[batch * M * P + row * P + col] = sum;
}
if (batch < batch_size && row < M && col < P) {
float sum = 0.0f;

// compute the dot product for each row of A and col of B
for (int i = 0; i < N; ++i) {
sum += d_A[batch * M * N + row * N + i] * d_B[batch * N * P + i * P + col];
}
printf("%f\n", sum);
d_C[batch * M * P + row * P + col] = sum;
printf("%f\n", d_C[batch * M * P + row * P + col]);
}
}

int main() {
// variable initialization
int M = 2;
int N = 3;
int P = 5;
int M = 2;
int N = 3;
int P = 5;
int batch_size = 4;

float *h_A, *h_B, *h_C;
float *d_A, *d_B, *d_C;

// memory allocation
h_A = (float *)malloc(M * N * sizeof(float));
h_B = (float *)malloc(N * P * sizeof(float));
h_C = (float *)malloc(M * P * sizeof(float));

cudaMalloc((void**)&d_A, M * N * sizeof(float));
cudaMalloc((void**)&d_B, N * P * sizeof(float));
cudaMalloc((void**)&d_C, M * P * sizeof(float));

// initial data
for(int i = 0; i < M; ++i) {
for(int j = 0; j < N; ++j) {
h_A[i * N + j] = (float) (rand() % 10 + 1);
}
}

for(int i = 0; i < N; ++i) {
for(int j = 0; j < P; ++j) {
h_B[i * P + j] = (float) (rand() % 10 + 1);
}
}

// copy CPU data to GPU memory blocks
cudaMemcpy(d_A, h_A, M * N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_B, h_B, N * P * sizeof(float), cudaMemcpyHostToDevice);

// set grid and block dimensions
dim3 blockDim(16, 16);
dim3 gridDim((P + blockDim.x - 1)/blockDim.x, (M + blockDim.y - 1)/blockDim.y);

// run matmul
matMul<<<gridDim, blockDim>>>(d_A, d_B, d_C, M, N, P);

// transfer data from device to host
cudaMemcpy(h_C, d_C, M * P * sizeof(float), cudaMemcpyDeviceToHost);

// print statements
printf("Matrix A:\n--------\n");
for(int i = 0; i < M; ++i) {
for(int j = 0; j < N; ++j) {
printf("%f ", h_A[i * N + j]);
}
printf("\n");
}

printf("--------\n");
printf("Matrix B:\n--------\n");
for(int i = 0; i < N; ++i) {
for(int j = 0; j < P; ++j) {
printf("%f ", h_B[i * P + j]);
}
printf("\n");
}

printf("--------\n");
printf("Matrix C:\n--------\n");
for(int i = 0; i < M; ++i) {
for(int j = 0; j < P; ++j) {
printf("%f ", h_C[i * P + j]);
}
printf("\n");
}
printf("--------\n");

// clean up device memory
cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);

free(h_A);
free(h_B);
free(h_C);

return 0;
float *h_A, *h_B, *h_C;
float *d_A, *d_B, *d_C;

// memory allocation
h_A = (float *)malloc(batch_size * M * N * sizeof(float));
h_B = (float *)malloc(batch_size * N * P * sizeof(float));
h_C = (float *)malloc(batch_size * M * P * sizeof(float));

cudaMalloc((void**)&d_A, batch_size * M * N * sizeof(float));
cudaMalloc((void**)&d_B, batch_size * N * P * sizeof(float));
cudaMalloc((void**)&d_C, batch_size * M * P * sizeof(float));

// initial data
for (int batch = 0; batch < batch_size; ++batch) {
for (int i = 0; i < M; ++i) {
for (int j = 0; j < N; ++j) {
h_A[batch * M * N + i * N + j] = (float) (rand() % 10 + 1);
}
}

for (int i = 0; i < N; ++i) {
for (int j = 0; j < P; ++j) {
h_B[batch * N * P + i * P + j] = (float) (rand() % 10 + 1);
}
}
}

// copy CPU data to GPU memory blocks
cudaMemcpy(d_A, h_A, batch_size * M * N * sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_B, h_B, batch_size * N * P * sizeof(float), cudaMemcpyHostToDevice);

// set grid and block dimensions
dim3 blockDim(8, 8, batch_size);
dim3 gridDim((P + blockDim.x - 1)/blockDim.x, (M + blockDim.y - 1)/blockDim.y, (batch_size + blockDim.z - 1)/blockDim.z);

// run batch matmul
bmm<<<gridDim, blockDim>>>(d_A, d_B, d_C, batch_size, M, N, P);

// sync
cudaDeviceSynchronize();

// check for errors
cudaError_t err = cudaGetLastError();
if (err != cudaSuccess) {
printf("CUDA Error: %s\n", cudaGetErrorString(err));
}

// transfer data from device to host
cudaMemcpy(h_C, d_C, batch_size * M * P * sizeof(float), cudaMemcpyDeviceToHost);

// print statements
for (int batch = 0; batch < batch_size; ++batch) {
printf("Matrix A:\n--------\n");
for (int i = 0; i < M; ++i) {
for (int j = 0; j < N; ++j) {
printf("%f ", h_A[batch * M * N + i * N + j]);
}
printf("\n");
}

printf("--------\n");
printf("Matrix B:\n--------\n");
for (int i = 0; i < N; ++i) {
for (int j = 0; j < P; ++j) {
printf("%f ", h_B[batch * N * P + i * P + j]);
}
printf("\n");
}

printf("--------\n");
printf("Matrix C:\n--------\n");
for (int i = 0; i < M; ++i) {
for (int j = 0; j < P; ++j) {
printf("%f ", h_C[batch * M * P + i * P + j]);
}
printf("\n");
}
printf("--------\n");
}

// clean up device memory
cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);

free(h_A);
free(h_B);
free(h_C);

return 0;
}

0 comments on commit 7654bc4

Please sign in to comment.