Validating `vmaf_sycl_import_d3d11_surface` in a local Windows VM¶

This doc is a reproducer for the Windows-only surface-import path added in ADR-0103. CI does not exercise it (the fork's Windows CI is MinGW + no-SYCL, and the D3D11 API requires Intel oneAPI DPC++ on Windows). Validation happens manually in a local Windows VM.

When you need to run this¶

You changed core/src/sycl/d3d11_import.cpp.
You changed core/src/sycl/common.cpp in a way that affects vmaf_sycl_upload_plane (the import path's sink).
You bumped the Intel oneAPI DPC++ toolkit version.

If none of the above, the Linux SYCL + Windows MinGW no-SYCL CI jobs are enough.

Prerequisites on the Windows VM¶

Windows 10/11 (x64) with an Intel GPU. An iGPU (Xe, UHD Graphics) is fine — the D3D11 path doesn't need a discrete Arc.
Intel oneAPI Base Toolkit — installs icx / icpx and the SYCL runtime.
Git for Windows, or a network share mounting the repo from the host.
Meson + Ninja (pip install meson ninja).

Build¶

# From an "Intel oneAPI command prompt" (icpx in PATH + env vars set)
cd C:\path\to\vmaf
meson setup build-sycl libvmaf ^
    -Denable_cuda=false -Denable_sycl=true ^
    --buildtype release
ninja -C build-sycl

If the build fails at the link step with an unknown -fsycl flag, your environment is dropping the Intel oneAPI wrappers — re-open the shell via the oneAPI launcher and retry.

Smoke test (EINVAL paths)¶

The EINVAL paths exercise the argument validation without needing a real decoder. Paste this into test_d3d11_smoke.c on the VM:

#include <stdio.h>
#include <libvmaf/libvmaf_sycl.h>

int main(void) {
    /* Every invalid-argument call must return < 0 without crashing. */
    int bad = vmaf_sycl_import_d3d11_surface(NULL, NULL, NULL, 0, 0, 0, 0, 0);
    printf("null pointers => %d\n", bad);

    /* Add your own positive-path test with a real ID3D11Texture2D once
     * you have a D3D11 device set up — see the "Full reproducer" section
     * below. */
    return bad < 0 ? 0 : 1;
}

Build + run:

icx test_d3d11_smoke.c ^
    -I build-sycl\include /link build-sycl\src\libvmaf.lib
.\test_d3d11_smoke.exe

Expected: prints null pointers => -22 (-EINVAL) and exits 0.

Full reproducer (real D3D11 surface)¶

For the actual import path you need a real decoded surface. The fastest route is to use MediaFoundation's H.264 decoder on a short clip — it outputs IMFSamples backed by ID3D11Texture2D.

/* pseudocode — full code lives in scripts/test-windows-d3d11-import.c */
ID3D11Device *dev; ID3D11DeviceContext *ctx;
D3D11CreateDevice(NULL, D3D_DRIVER_TYPE_HARDWARE, NULL, 0,
                  NULL, 0, D3D11_SDK_VERSION, &dev, NULL, &ctx);

/* Decode one frame, extract its ID3D11Texture2D via
 * IMFDXGIBuffer::GetResource. Call it tex. */

VmafSyclState *sycl_state;
VmafSyclConfig sycl_cfg = { .device_index = 0 };
vmaf_sycl_state_init(&sycl_state, &sycl_cfg);
vmaf_sycl_init_frame_buffers(sycl_state, 1920, 1080, 8);

int rc = vmaf_sycl_import_d3d11_surface(sycl_state, dev, tex,
                                        /* subresource */ 0,
                                        /* is_ref */ 1,
                                        1920, 1080, 8);
printf("import rc = %d\n", rc);  /* expect 0 */

Record the timing: a 1080p8 frame should take ≲ 2 ms end-to-end on a typical iGPU + PCIe Gen3 x8 path. An order-of-magnitude slower result suggests the staging tex isn't in CPU-accessible memory or that the SYCL H2D path is stalling behind a previous submit.

Reporting results¶

Paste into the PR description:

Windows build: SHA + oneAPI toolkit version.
Smoke test: pass / fail + the errno values seen.
Full reproducer: frame count processed, mean per-frame μs, which GPU.

That's enough evidence for review — no screen recording, no log upload.

Validating vmaf_sycl_import_d3d11_surface in a local Windows VM¶