AI/rocm-systems
2
0 Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

Replace hipLaunchKernel -> hipLaunchKernelGGL

Telusuri Sumber 
Change-Id: I4d99009e1199811d417becf1e1b934ec4d4e30be
This commit is contained in:
Maneesh Gupta committed 2018-10-17 14:32:25 +05:30
1 orang tua 3bf8e9ad8e
melakukan e1fe095471
78 files changed
+246 -264
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

hipamd/tests/src/runtimeApi/memory/hipArray.cpp
+3 -3
Melihat File
@@ -67,7 +67,7 @@ void memcpy2Dtest(size_t numW, size_t numH, bool usePinnedHost) {
HIPCHECK(hipMemcpy2D(A_d, pitch_A, A_h, width, width, numH, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy2D(B_d, pitch_B, B_h, width, width, numH, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0, A_d, B_d, C_d,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0, A_d, B_d, C_d,
(pitch_C / sizeof(T)) * numH);
HIPCHECK(hipMemcpy2D(C_h, width, C_d, pitch_C, width, numH, hipMemcpyDeviceToHost));
@@ -117,7 +117,7 @@ void memcpyArraytest(size_t numW, size_t numH, bool usePinnedHost, bool usePitch
HIPCHECK(hipMemcpyToArray(A_d, 0, 0, (void*)A_h, width, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpyToArray(B_d, 0, 0, (void*)B_h, width, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
(T*)A_d->data, (T*)B_d->data, (T*)C_d->data, numW);
HIPCHECK(hipMemcpy(C_h, C_d->data, width, hipMemcpyDeviceToHost));
@@ -156,7 +156,7 @@ void memcpyArraytest(size_t numW, size_t numH, bool usePinnedHost, bool usePitch
hipMemcpyHostToDevice));
}
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
(T*)A_d->data, (T*)B_d->data, (T*)C_d->data, numW * numH);
HIPCHECK(hipMemcpy2D((void*)C_h, width, (void*)C_d->data, width, width, numH,
hipamd/tests/src/runtimeApi/memory/hipHostGetFlags.cpp
+2 -2
Melihat File
@@ -32,7 +32,7 @@ THE SOFTWARE.
#define LEN 1024 * 1024
#define SIZE LEN * sizeof(float)
__global__ void Add(hipLaunchParm lp, float* Ad, float* Bd, float* Cd) {
__global__ void Add(float* Ad, float* Bd, float* Cd) {
int tx = threadIdx.x + blockIdx.x * blockDim.x;
Cd[tx] = Ad[tx] + Bd[tx];
}
@@ -74,7 +74,7 @@ int main() {
dim3 dimGrid(LEN / 512, 1, 1);
dim3 dimBlock(512, 1, 1);
hipLaunchKernel(HIP_KERNEL_NAME(Add), dimGrid, dimBlock, 0, 0, Ad, Bd, Cd);
hipLaunchKernelGGL(HIP_KERNEL_NAME(Add), dimGrid, dimBlock, 0, 0, Ad, Bd, Cd);
HIPCHECK(
hipMemcpy(C, Cd, SIZE, hipMemcpyDeviceToHost)); // Note this really HostToHost not
hipamd/tests/src/runtimeApi/memory/hipHostRegister.cpp
+2 -2
Melihat File
@@ -28,7 +28,7 @@ THE SOFTWARE.
#include "test_common.h"
#include <malloc.h>
__global__ void Inc(hipLaunchParm lp, float* Ad) {
__global__ void Inc(float* Ad) {
int tx = threadIdx.x + blockIdx.x * blockDim.x;
Ad[tx] = Ad[tx] + float(1);
}
@@ -99,7 +99,7 @@ int main(int argc, char* argv[]) {
// Reference the registered device pointer Ad from inside the kernel:
for (int i = 0; i < num_devices; i++) {
HIPCHECK(hipSetDevice(i));
hipLaunchKernel(Inc, dim3(N / 512), dim3(512), 0, 0, Ad[i]);
hipLaunchKernelGGL(Inc, dim3(N / 512), dim3(512), 0, 0, Ad[i]);
HIPCHECK(hipDeviceSynchronize());
}
hipamd/tests/src/runtimeApi/memory/hipMemcpy.cpp
+1 -1
Melihat File
@@ -230,7 +230,7 @@ void memcpytest2(DeviceMemory<T>* dmem, HostMemory<T>* hmem, size_t numElements,
useMemkindDefault ? hipMemcpyDefault : hipMemcpyHostToDevice));
}
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const T*>(dmem->A_d()), static_cast<const T*>(dmem->B_d()),
dmem->C_d(), numElements);
hipamd/tests/src/runtimeApi/memory/hipMemcpyDtoD.cpp
+2 -2
Melihat File
@@ -51,7 +51,7 @@ int main() {
HIPCHECK(hipSetDevice(0));
HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(A_d), static_cast<const int*>(B_d), C_d, N);
HIPCHECK(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
@@ -62,7 +62,7 @@ int main() {
HIPCHECK(hipMemcpyDtoD((hipDeviceptr_t)X_d, (hipDeviceptr_t)A_d, Nbytes));
HIPCHECK(hipMemcpyDtoD((hipDeviceptr_t)Y_d, (hipDeviceptr_t)B_d, Nbytes));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(X_d), static_cast<const int*>(Y_d), Z_d, N);
HIPCHECK(hipMemcpyDtoH(C_h, (hipDeviceptr_t)Z_d, Nbytes));
HIPCHECK(hipDeviceSynchronize());
hipamd/tests/src/runtimeApi/memory/hipMemcpyDtoDAsync.cpp
+2 -2
Melihat File
@@ -52,7 +52,7 @@ int main() {
HIPCHECK(hipSetDevice(0));
HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(A_d), static_cast<const int*>(B_d), C_d, N);
HIPCHECK(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
@@ -63,7 +63,7 @@ int main() {
HIPCHECK(hipMemcpyDtoDAsync((hipDeviceptr_t)X_d, (hipDeviceptr_t)A_d, Nbytes, s));
HIPCHECK(hipMemcpyDtoDAsync((hipDeviceptr_t)Y_d, (hipDeviceptr_t)B_d, Nbytes, s));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(X_d), static_cast<const int*>(Y_d), Z_d, N);
HIPCHECK(hipMemcpyDtoHAsync(C_h, (hipDeviceptr_t)Z_d, Nbytes, s));
HIPCHECK(hipStreamSynchronize(s));
hipamd/tests/src/runtimeApi/memory/hipMemcpyPeer.cpp
+2 -2
Melihat File
@@ -50,7 +50,7 @@ int main() {
HIPCHECK(hipSetDevice(0));
HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(A_d), static_cast<const int*>(B_d), C_d, N);
HIPCHECK(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
@@ -62,7 +62,7 @@ int main() {
Nbytes); // this call is eqv to hipMemcpy(hipMemcpyD2D) which goes via stg bufs.
hipMemcpyPeer(Y_d, 1, B_d, 0, Nbytes);
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(X_d), static_cast<const int*>(Y_d), Z_d, N);
HIPCHECK(hipMemcpy(C_h, Z_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
hipamd/tests/src/runtimeApi/memory/hipMemcpyPeerAsync.cpp
+2 -2
Melihat File
@@ -54,7 +54,7 @@ int main() {
HIPCHECK(hipSetDevice(0));
HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(A_d), static_cast<const int*>(B_d), C_d, N);
HIPCHECK(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
@@ -65,7 +65,7 @@ int main() {
HIPCHECK(hipMemcpyPeerAsync(X_d, 1, A_d, 0, Nbytes, s));
HIPCHECK(hipMemcpyPeerAsync(Y_d, 1, B_d, 0, Nbytes, s));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(X_d), static_cast<const int*>(Y_d), Z_d, N);
HIPCHECK(hipMemcpy(C_h, Z_d, Nbytes, hipMemcpyDeviceToHost));
HIPCHECK(hipDeviceSynchronize());
hipamd/tests/src/runtimeApi/memory/hipMemcpy_simple.cpp
+1 -1
Melihat File
@@ -61,7 +61,7 @@ void simpleTest1() {
HIPCHECK(memcopy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(memcopy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
hipLaunchKernel(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
static_cast<const int*>(A_d), static_cast<const int*>(B_d), C_d, N);
HIPCHECK(memcopy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
hipamd/tests/src/runtimeApi/memory/hipMemoryAllocateCoherent.cpp
+2 -2
Melihat File
@@ -29,7 +29,7 @@ THE SOFTWARE.
#include <cstdio>
#include "hip/hip_runtime.h"
__global__ void Kernel(hipLaunchParm lp, volatile float* hostRes) {
__global__ void Kernel(volatile float* hostRes) {
int tid = threadIdx.x + blockIdx.x * blockDim.x;
hostRes[tid] = tid + 1;
__threadfence_system();
@@ -45,7 +45,7 @@ int main() {
hipHostMalloc((void**)&hostRes, blocks * sizeof(float), hipHostMallocMapped);
hostRes[0] = 0;
hostRes[1] = 0;
hipLaunchKernel(HIP_KERNEL_NAME(Kernel), dim3(1), dim3(blocks), 0, 0, hostRes);
hipLaunchKernelGGL(HIP_KERNEL_NAME(Kernel), dim3(1), dim3(blocks), 0, 0, hostRes);
int eleCounter = 0;
while (eleCounter < blocks) {
// blocks until the value changes