Cleanup and refactor

This commit is contained in:
2024-04-11 22:52:41 +02:00
parent 4b9d123e94
commit 18522c2dea
9 changed files with 81 additions and 72 deletions

View File

@@ -47,21 +47,3 @@ TEST(ActivationFunctionsTest, SigmoidSanityCheck) {
cudaDeviceReset();
}
// void print_vec(float* d_vec, int length) {
// std::vector<float> h_vec(length);
// CUDA_CHECK(cudaMemcpy(
// h_vec.data(), d_vec, sizeof(float) * length, cudaMemcpyDeviceToHost
// ));
// float sum = 0.0f;
// for (int i = 0; i < length; ++i) {
// std::cout << h_vec[i] << ", ";
// sum += h_vec[i];
// }
// std::cout << std::endl;
// }

View File

@@ -61,7 +61,7 @@ TEST(MatMulTest, MatVecMulTest) {
for (int j = 0; j < w; j++) {
sum += matrix[i * w + j] * vector[j];
}
EXPECT_NEAR(sum, output_gpu[i], 1e-5);
EXPECT_NEAR(sum, output_gpu[i], 1e-5f);
}
cudaFree(d_matrix);
@@ -151,7 +151,7 @@ TEST(MatMulTest, VecExpTest) {
EXPECT_EQ(cudaStatus, cudaSuccess);
for (int i = 0; i < 6; i++) {
EXPECT_NEAR(expected[i], output[i], 1e7);
EXPECT_NEAR(expected[i], output[i], 1e7f);
}
cudaFree(d_input);
@@ -193,7 +193,6 @@ TEST(MatMulTest, SumReduceTest) {
int remaining = gridSize;
while (remaining > 1) {
std::cout << remaining << std::endl;
int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::sum_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_sum, d_sum, remaining);
remaining = blocks_needed;