Visible to Intel only — GUID: GUID-E9F4ED64-CB7F-4A0B-9D62-1AC0CDCC2B88
Execution Model Overview
Thread Mapping and GPU Occupancy
Kernels
Using Libraries for GPU Offload
Host/Device Memory, Buffer and USM
Host/Device Coordination
Using Multiple Heterogeneous Devices
Compilation
OpenMP Offloading Tuning Guide
Multi-GPU, Multi-Stack and Multi-C-Slice Architecture and Programming
Level Zero
Performance Profiling and Analysis
Configuring GPU Device
Sub-Groups and SIMD Vectorization
Removing Conditional Checks
Registerization and Avoiding Register Spills
Small Register Mode vs. Large Register Mode
Shared Local Memory
Pointer Aliasing and the Restrict Directive
Synchronization among Threads in a Kernel
Considerations for Selecting Work-Group Size
Reduction
Kernel Launch
Executing Multiple Kernels on the Device at the Same Time
Submitting Kernels to Multiple Queues
Avoiding Redundant Queue Constructions
Programming Intel® XMX Using SYCL Joint Matrix Extension
Doing I/O in the Kernel
Visible to Intel only — GUID: GUID-E9F4ED64-CB7F-4A0B-9D62-1AC0CDCC2B88
Avoid Declaring Buffers in a Loop
When kernels are repeatedly launched inside a for-loop, you can prevent repeated allocation and freeing of a buffer by declaring the buffer outside the loop. Declaring a buffer inside the loop introduces repeated host-to-device and device-to-host memory copies.
In the following example, the kernel is repeatedly launched inside a for-loop. The buffer C is used as a temporary array, where it is used to hold values in an iteration, and the values assigned in one iteration are not used in any other iteration. Since the buffer C is declared inside the for-loop, it is allocated and freed in every loop iteration. In addition to the allocation and freeing of the buffer, the memory associated with the buffer is redundantly transferred from host to device and device to host in each iteration.
//==============================================================
// Copyright © 2022 Intel Corporation
//
// SPDX-License-Identifier: MIT
// =============================================================
#include <CL/sycl.hpp>
#include <stdio.h>
constexpr int N = 25;
constexpr int STEPS = 100000;
int main() {
int AData[N];
int BData[N];
int CData[N];
sycl::queue Q;
// Create 2 buffers, each holding N integers
sycl::buffer<int> ABuf(&AData[0], N);
sycl::buffer<int> BBuf(&BData[0], N);
Q.submit([&](auto &h) {
// Create device accessors.
// The property no_init lets the runtime know that the
// previous contents of the buffer can be discarded.
sycl::accessor aA(ABuf, h, sycl::write_only, sycl::no_init);
sycl::accessor aB(BBuf, h, sycl::write_only, sycl::no_init);
h.parallel_for(N, [=](auto i) {
aA[i] = 10;
aB[i] = 20;
});
});
for (int j = 0; j < STEPS; j++) {
sycl::buffer<int> CBuf(&CData[0], N);
Q.submit([&](auto &h) {
// Create device accessors.
sycl::accessor aA(ABuf, h);
sycl::accessor aB(BBuf, h);
sycl::accessor aC(CBuf, h);
h.parallel_for(N, [=](auto i) {
aC[i] = (aA[i] < aB[i]) ? -1 : 1;
aA[i] += aC[i];
aB[i] -= aC[i];
});
});
} // end for
// Create host accessors.
const sycl::host_accessor haA(ABuf);
const sycl::host_accessor haB(BBuf);
printf("%d %d\n", haA[N / 2], haB[N / 2]);
return 0;
}
A better approach would be to declare the buffer C before the for-loop, so that it is allocated and freed only once, resulting in improved performance by avoiding the redundant data transfers between host and device. The following kernel shows this change.
//==============================================================
// Copyright © 2022 Intel Corporation
//
// SPDX-License-Identifier: MIT
// =============================================================
#include <CL/sycl.hpp>
#include <stdio.h>
constexpr int N = 25;
constexpr int STEPS = 100000;
int main() {
int AData[N];
int BData[N];
int CData[N];
sycl::queue Q;
// Create 3 buffers, each holding N integers
sycl::buffer<int> ABuf(&AData[0], N);
sycl::buffer<int> BBuf(&BData[0], N);
sycl::buffer<int> CBuf(&CData[0], N);
Q.submit([&](auto &h) {
// Create device accessors.
// The property no_init lets the runtime know that the
// previous contents of the buffer can be discarded.
sycl::accessor aA(ABuf, h, sycl::write_only, sycl::no_init);
sycl::accessor aB(BBuf, h, sycl::write_only, sycl::no_init);
h.parallel_for(N, [=](auto i) {
aA[i] = 10;
aB[i] = 20;
});
});
for (int j = 0; j < STEPS; j++) {
Q.submit([&](auto &h) {
// Create device accessors.
sycl::accessor aA(ABuf, h);
sycl::accessor aB(BBuf, h);
sycl::accessor aC(CBuf, h);
h.parallel_for(N, [=](auto i) {
aC[i] = (aA[i] < aB[i]) ? -1 : 1;
aA[i] += aC[i];
aB[i] -= aC[i];
});
});
} // end for
// Create host accessors.
const sycl::host_accessor haA(ABuf);
const sycl::host_accessor haB(BBuf);
printf("%d %d\n", haA[N / 2], haB[N / 2]);
return 0;
}