AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-408958 - Use LaunchDelayKernel and modify it to use same kernel based on real time clock for gfx10 and gfx11. (#370)

Browse Source 
Change-Id: Iea8a48e8cbfa1745c7d5535dc5820133a4104087
This commit is contained in:
ROCm CI Service Account
2023-07-20 10:16:56 +05:30
committed by GitHub
parent faa2dd7cfb
commit 04080c2e2e
14 changed files with 79 additions and 209 deletions
Download Patch File Download Diff File Expand all files Collapse all files
catch/unit/stream/hipStreamACb_MultiThread.cc
+3 -29
View File
@@ -25,6 +25,7 @@ multiple Threads.
#include <hip_test_common.hh>
#include <atomic>
#include <utils.hh>
static constexpr size_t N = 4096;
static constexpr int numThreads = 1000;
@@ -43,33 +44,6 @@ static __global__ void device_function(float* C_d, float* A_d, size_t Num) {
for (size_t i = gputhread; i < Num; i += stride) {
C_d[i] = A_d[i] * A_d[i];
}
// Delay thread 1 only in the GPU
if (gputhread == 1) {
uint64_t wait_t = 3200000000, start = clock64(), cur;
do {
cur = clock64() - start;
} while (cur < wait_t);
}
}
static __global__ void device_function_gfx11(float* C_d, float* A_d, size_t Num) {
#if HT_AMD
size_t gputhread = (blockIdx.x * blockDim.x + threadIdx.x);
size_t stride = blockDim.x * gridDim.x;
for (size_t i = gputhread; i < Num; i += stride) {
C_d[i] = A_d[i] * A_d[i];
}
// Delay thread 1 only in the GPU
if (gputhread == 1) {
uint64_t wait_t = 3200000000, start = wall_clock64(), cur;
do {
cur = wall_clock64() - start;
} while (cur < wait_t);
}
#endif
}
static void HIPRT_CB Thread1_Callback(hipStream_t stream, hipError_t status,
@@ -146,10 +120,10 @@ TEST_CASE("Unit_hipStreamAddCallback_MultipleThreads") {
constexpr unsigned threadsPerBlock = 256;
constexpr unsigned blocks = (N + 255)/threadsPerBlock;
auto device_function_used = IsGfx11() ? device_function_gfx11 : device_function;
hipLaunchKernelGGL((device_function_used), dim3(blocks),
hipLaunchKernelGGL((device_function), dim3(blocks),
dim3(threadsPerBlock), 0,
mystream, C_d, A_d, N);
LaunchDelayKernel(std::chrono::milliseconds(2000), mystream);
HIP_CHECK(hipGetLastError());
HIP_CHECK(
hipMemcpyAsync(C1_h, C_d, Nbytes,