AI/rocm-systems
2
0
フォーク 0
コード イシュー プルリクエスト Actions パッケージ プロジェクト リリース Wiki アクティビティ

[HIPIFY][#1400] Fix Template Instantiation kernel launch (clang & perl)

ソースを参照 
+ Enclose template instantiation kernel calls into round brackets, leave regular kernel names unchanged (hipify-perl doesn't handle cases with macros).
+ Fix corresponding tests.

PS. hipify-perl couldn't handle correctly the following cases due to macros expansion disability, thus hipify-clang should be used instead:

#define KERNEL_NAME_MACRO axpy<float>
#define KERNEL_CALL_MACRO axpy<float><<<1, 2>>>
#define KERNEL_ARG_LIST_MACRO a, x, y

// CUDA:
KERNEL_NAME_MACRO<<<1, 2>>>(KERNEL_ARG_LIST_MACRO);
KERNEL_CALL_MACRO(KERNEL_ARG_LIST_MACRO);

// hipify-perl:
hipLaunchKernelGGL(KERNEL_NAME_MACRO, dim3(1), dim3(2), 0, 0, KERNEL_ARG_LIST_MACRO);
KERNEL_CALL_MACRO(KERNEL_ARG_LIST_MACRO);

// hipify-clang:
hipLaunchKernelGGL((KERNEL_NAME_MACRO), dim3(1), dim3(2), 0, 0, KERNEL_ARG_LIST_MACRO);
hipLaunchKernelGGL((axpy<float>), dim3(1), dim3(2), 0, 0, KERNEL_ARG_LIST_MACRO);
このコミットが含まれているのは:
Evgeny Mankov
2019-09-10 15:59:06 +03:00
親 29d09eb269
コミット 56ab105e9d
5個のファイルの変更、65行の追加、35行の削除
Patchファイルをダウンロード Diffファイルをダウンロード すべてのファイルを展開 すべてのファイルを折り畳む
tests/hipify-clang/unit_tests/libraries/cuRAND/benchmark_curand_kernel.cpp
+5 -5
ファイルの表示
@@ -121,7 +121,7 @@ struct runner
const size_t states_size = blocks * threads;
// CHECK: CUDA_CALL(hipMalloc((void **)&states, states_size * sizeof(GeneratorState)));
CUDA_CALL(cudaMalloc((void **)&states, states_size * sizeof(GeneratorState)));
// CHECK: hipLaunchKernelGGL(init_kernel, dim3(blocks), dim3(threads), 0, 0, states, seed, offset);
// CHECK: hipLaunchKernelGGL((init_kernel), dim3(blocks), dim3(threads), 0, 0, states, seed, offset);
init_kernel<<<blocks, threads>>>(states, seed, offset);
// CHECK: CUDA_CALL(hipPeekAtLastError());
// CHECK: CUDA_CALL(hipDeviceSynchronize());
@@ -142,7 +142,7 @@ struct runner
const GenerateFunc& generate_func,
const Extra extra)
{
// CHECK: hipLaunchKernelGGL(generate_kernel, dim3(blocks), dim3(threads), 0, 0, states, data, size, generate_func, extra);
// CHECK: hipLaunchKernelGGL((generate_kernel), dim3(blocks), dim3(threads), 0, 0, states, data, size, generate_func, extra);
generate_kernel<<<blocks, threads>>>(states, data, size, generate_func, extra);
}
};
@@ -223,7 +223,7 @@ struct runner<curandStateMtgp32_t>
const GenerateFunc& generate_func,
const Extra extra)
{
// CHECK: hipLaunchKernelGGL(generate_kernel, dim3(std::min((size_t)200, blocks)), dim3(256), 0, 0, states, data, size, generate_func, extra);
// CHECK: hipLaunchKernelGGL((generate_kernel), dim3(std::min((size_t)200, blocks)), dim3(256), 0, 0, states, data, size, generate_func, extra);
generate_kernel<<<std::min((size_t)200, blocks), 256>>>(states, data, size, generate_func, extra);
}
};
@@ -304,7 +304,7 @@ struct runner<curandStateSobol32_t>
CUDA_CALL(cudaMemcpy(directions, h_directions, size, cudaMemcpyHostToDevice));
const size_t blocks_x = next_power2((blocks + dimensions - 1) / dimensions);
// CHECK: hipLaunchKernelGGL(init_kernel, dim3(dim3(blocks_x, dimensions)), dim3(threads), 0, 0, states, directions, offset);
// CHECK: hipLaunchKernelGGL((init_kernel), dim3(dim3(blocks_x, dimensions)), dim3(threads), 0, 0, states, directions, offset);
init_kernel<<<dim3(blocks_x, dimensions), threads>>>(states, directions, offset);
// CHECK: CUDA_CALL(hipPeekAtLastError());
// CHECK: CUDA_CALL(hipDeviceSynchronize());
@@ -329,7 +329,7 @@ struct runner<curandStateSobol32_t>
const Extra extra)
{
const size_t blocks_x = next_power2((blocks + dimensions - 1) / dimensions);
// CHECK: hipLaunchKernelGGL(generate_kernel, dim3(dim3(blocks_x, dimensions)), dim3(threads), 0, 0, states, data, size / dimensions, generate_func, extra);
// CHECK: hipLaunchKernelGGL((generate_kernel), dim3(dim3(blocks_x, dimensions)), dim3(threads), 0, 0, states, data, size / dimensions, generate_func, extra);
generate_kernel<<<dim3(blocks_x, dimensions), threads>>>(states, data, size / dimensions, generate_func, extra);
}
};
tests/hipify-clang/unit_tests/samples/axpy.cu
+7 -7
ファイルの表示
@@ -10,7 +10,7 @@
#define KERNEL_CALL_AS_MACRO axpy<float><<<1, kDataLen>>>
#define KERNEL_NAME_MACRO axpy<float>
// CHECK: #define COMPLETE_LAUNCH hipLaunchKernelGGL(axpy, dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y)
// CHECK: #define COMPLETE_LAUNCH hipLaunchKernelGGL((axpy), dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y)
#define COMPLETE_LAUNCH axpy<<<1, kDataLen>>>(a, device_x, device_y)
@@ -48,22 +48,22 @@ int main(int argc, char* argv[]) {
cudaMemcpy(device_x, host_x, kDataLen * sizeof(float), cudaMemcpyHostToDevice);
// Launch the kernel in numerous different strange ways to exercise the prerocessor.
// CHECK: hipLaunchKernelGGL(axpy, dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y);
// CHECK: hipLaunchKernelGGL((axpy), dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y);
axpy<<<1, kDataLen>>>(a, device_x, device_y);
// CHECK: hipLaunchKernelGGL(axpy<float>, dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y);
// CHECK: hipLaunchKernelGGL((axpy<float>), dim3(1), dim3(kDataLen), 0, 0, a, device_x, device_y);
axpy<float><<<1, kDataLen>>>(a, device_x, device_y);
// CHECK: hipLaunchKernelGGL(axpy<float>, dim3(1), dim3(kDataLen), 0, 0, a, TOKEN_PASTE(device, _x), device_y);
// CHECK: hipLaunchKernelGGL((axpy<float>), dim3(1), dim3(kDataLen), 0, 0, a, TOKEN_PASTE(device, _x), device_y);
axpy<float><<<1, kDataLen>>>(a, TOKEN_PASTE(device, _x), device_y);
// CHECK: hipLaunchKernelGGL(axpy<float>, dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL((axpy<float>), dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
axpy<float><<<1, kDataLen>>>(ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL(KERNEL_NAME_MACRO, dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL((KERNEL_NAME_MACRO), dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
KERNEL_NAME_MACRO<<<1, kDataLen>>>(ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL(axpy<float>, dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL((axpy<float>), dim3(1), dim3(kDataLen), 0, 0, ARG_LIST_AS_MACRO);
KERNEL_CALL_AS_MACRO(ARG_LIST_AS_MACRO);
// CHECK: hipLaunchKernelGGL(empty, dim3(1), dim3(kDataLen), 0, 0);
tests/hipify-clang/unit_tests/samples/square.cu
+1 -1
ファイルの表示
@@ -94,7 +94,7 @@ int main(int argc, char *argv[])
const unsigned threadsPerBlock = 256;
printf ("info: launch 'vector_square' kernel\n");
// CHECK: hipLaunchKernelGGL(vector_square, dim3(blocks), dim3(threadsPerBlock), 0, 0, C_d, A_d, N);
// CHECK: hipLaunchKernelGGL((vector_square), dim3(blocks), dim3(threadsPerBlock), 0, 0, C_d, A_d, N);
vector_square <<<blocks, threadsPerBlock>>> (C_d, A_d, N);
printf ("info: copy Device2Host\n");