Fix test hipDynamicShared

This test does not work if block size is greater than wave size
since it relies on lock-step execution of the kernel in the block.
If there are more than waves in the block, the threads in the block
miss synchronization since one wave may finish before another wave.

Due to this bug, the test fails on GFX10 wave32 mode.

This patch fixes that so that it works for block size greater than
wave size.

Change-Id: Ie0097066081df36cb6fe025a71d0ee5a83ec00a2
This commit is contained in:
Yaxun (Sam) Liu
2020-10-08 11:56:14 -04:00
zatwierdzone przez Karthik Jayaprakash
rodzic 86b1600ca7
commit 78269dcbe5
+30 -46
Wyświetl plik
@@ -29,6 +29,22 @@ THE SOFTWARE.
#include "hip/hip_runtime.h"
#include "test_common.h"
template <unsigned batch, typename T>
__device__ void sum(T* sdata, unsigned groupElements, unsigned tid) {
T tmp;
if (groupElements < batch)
return;
// sdata[tid] += sdata[tid - batch/2] does not work when block size is
// greater than wave size because one wave may complete before another
// wave.
if (tid >= batch/2 && tid < groupElements)
tmp = sdata[tid - batch/2];
__syncthreads();
if (tid >= batch/2 && tid < groupElements)
sdata[tid] += tmp;
__syncthreads();
}
template <typename T>
__global__ void testExternSharedKernel(const T* A_d, const T* B_d, T* C_d,
size_t numElements, size_t groupElements) {
@@ -47,51 +63,18 @@ __global__ void testExternSharedKernel(const T* A_d, const T* B_d, T* C_d,
if (tid < groupElements) {
sdata[tid] = static_cast<T>(tid);
}
__syncthreads();
// prefix sum inside dynamic shared memory
if (groupElements >= 512) {
if (tid >= 256) {
sdata[tid] += sdata[tid - 256];
}
__syncthreads();
}
if (groupElements >= 256) {
if (tid >= 128) {
sdata[tid] += sdata[tid - 128];
}
__syncthreads();
}
if (groupElements >= 128) {
if (tid >= 64) {
sdata[tid] += sdata[tid - 64];
}
__syncthreads();
}
if (groupElements >= 64) {
sdata[tid] += sdata[tid - 32];
}
__syncthreads();
if (groupElements >= 32) {
sdata[tid] += sdata[tid - 16];
}
__syncthreads();
if (groupElements >= 16) {
sdata[tid] += sdata[tid - 8];
}
__syncthreads();
if (groupElements >= 8) {
sdata[tid] += sdata[tid - 4];
}
__syncthreads();
if (groupElements >= 4) {
sdata[tid] += sdata[tid - 2];
}
__syncthreads();
if (groupElements >= 2) {
sdata[tid] += sdata[tid - 1];
}
__syncthreads();
sum<512>(sdata, groupElements, tid);
sum<256>(sdata, groupElements, tid);
sum<128>(sdata, groupElements, tid);
sum<64>(sdata, groupElements, tid);
sum<32>(sdata, groupElements, tid);
sum<16>(sdata, groupElements, tid);
sum<8>(sdata, groupElements, tid);
sum<4>(sdata, groupElements, tid);
sum<2>(sdata, groupElements, tid);
C_d[gid] = A_d[gid] + B_d[gid] + sdata[tid % groupElements];
}
@@ -103,9 +86,10 @@ void testExternShared(size_t N, size_t groupElements) {
T *A_h, *B_h, *C_h;
HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N, false);
unsigned blocks = HipTest::setNumBlocks(blocksPerCU, threadsPerBlock, N);
unsigned blocks = N/threadsPerBlock;
assert(N == blocks * threadsPerBlock);
// printf("blocksPerCU: %d\nthreadsPerBlock: %d\nN: %zu\n", blocksPerCU, threadsPerBlock, N);
// printf("blocks: %d\nthreadsPerBlock: %d\nN: %zu\n", blocks, threadsPerBlock, N);
HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
@@ -123,7 +107,7 @@ void testExternShared(size_t N, size_t groupElements) {
// verify
for (size_t i = 0; i < N; ++i) {
size_t tid = (i % groupElements);
size_t tid = (i % min(threadsPerBlock, groupElements));
T sumFromSharedMemory = static_cast<T>(tid * (tid + 1) / 2);
T expected = A_h[i] + B_h[i] + sumFromSharedMemory;
if (C_h[i] != expected) {