In the tlp_v2_upsert_kernel_with_io interface of the upsert.cuh file, when the cooperative thread group writes values in a pipelined operation, why is __pipeline_wait_prior(0) used here? Does using __pipeline_wait_prior(0) also block prefetch operations, preventing the pipeline effect from being achieved?
constexpr uint32_t GROUP_BUFs =
GROUP_SIZE * 2 * STRIDE_S * sizeof(S) / sizeof(VecV);
constexpr uint32_t GROUP_BUF = GROUP_BUFs / 2;
auto sm_values_buffer =
reinterpret_cast<VecV*>(&(sm_bucket_scores[0][0])) + groupID * GROUP_BUFs;
auto occupy_result_next = g.shfl(occupy_result, 0);
if ((occupy_result_next != OccupyResult::ILLEGAL) &&
(occupy_result_next != OccupyResult::REFUSED)) {
VecV* dst = sm_values_buffer;
auto kv_idx_next = g.shfl(kv_idx, 0);
const VecV* src = values + kv_idx_next * dim;
CopyValue::ldg_sts(rank, dst, src, dim);
}
__pipeline_commit();
for (int i = 0; i < GROUP_SIZE; i++) {
if (i + 1 < GROUP_SIZE) {
auto occupy_result_next = g.shfl(occupy_result, i + 1);
if ((occupy_result_next != OccupyResult::ILLEGAL) &&
(occupy_result_next != OccupyResult::REFUSED)) {
VecV* dst = sm_values_buffer + diff_buf(i) * GROUP_BUF;
auto kv_idx_next = g.shfl(kv_idx, i + 1);
const VecV* src = values + kv_idx_next * dim;
CopyValue::ldg_sts(rank, dst, src, dim);
}
}
__pipeline_commit();
auto occupy_result_cur = g.shfl(occupy_result, i);
if ((occupy_result_cur != OccupyResult::ILLEGAL) &&
(occupy_result_cur != OccupyResult::REFUSED)) {
VecV* src = sm_values_buffer + same_buf(i) * GROUP_BUF;
// --------- question about this code block
__pipeline_wait_prior(0); // this line
// ----------
VecV* dst = g.shfl(bucket_value_ptr, i);
__pipeline_wait_prior(1);
CopyValue::lds_stg(rank, dst, src, dim);
}
}Thanks !
In the
tlp_v2_upsert_kernel_with_iointerface of theupsert.cuhfile, when the cooperative thread group writes values in a pipelined operation, why is__pipeline_wait_prior(0)used here? Does using __pipeline_wait_prior(0) also block prefetch operations, preventing the pipeline effect from being achieved?Thanks !