obs-studio/plugins/obs-qsv11/common_utils_linux.cpp

#include "common_utils.h"

#include <time.h>
#include <cpuid.h>

#include <va/va_drm.h>
#include <va/va_drmcommon.h>
#include <va/va_str.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <fcntl.h>
#include <string.h>
#include <dirent.h>

#include <obs.h>
#include <obs-encoder.h>
#include <obs-nix-platform.h>
#include <graphics/graphics.h>
#include <util/c99defs.h>
#include <util/dstr.h>
#include <util/bmem.h>

// Set during check_adapters to work-around VPL dispatcher not setting a VADisplay
// for the MSDK runtime.
static const char *default_h264_device = nullptr;
static const char *default_hevc_device = nullptr;
static const char *default_av1_device = nullptr;

struct linux_data {
	int fd;
	VADisplay vaDisplay;
};

#define DEVICE_MGR_TYPE MFX_HANDLE_VA_DISPLAY
// This ends up at like 72 for 1440p@120 AV1.
// We may end up hitting this in practice?
constexpr int32_t MAX_ALLOCABLE_SURFACES = 128;

struct surface_info {
	VASurfaceID id;
	int32_t width, height;
	gs_texture_t *tex_y;
	gs_texture_t *tex_uv;
};

mfxStatus simple_alloc(mfxHDL pthis, mfxFrameAllocRequest *request,
		       mfxFrameAllocResponse *response)
{
	if (request->Type & (MFX_MEMTYPE_SYSTEM_MEMORY |
			     MFX_MEMTYPE_VIDEO_MEMORY_PROCESSOR_TARGET))
		return MFX_ERR_UNSUPPORTED;

	response->mids = (mfxMemId *)nullptr;
	response->NumFrameActual = 0;

	mfxSession *session = (mfxSession *)pthis;
	VADisplay display;
	mfxStatus sts =
		MFXVideoCORE_GetHandle(*session, DEVICE_MGR_TYPE, &display);
	MSDK_CHECK_RESULT(sts, MFX_ERR_NONE, sts);

	// https://ffmpeg.org/doxygen/5.1/hwcontext__vaapi_8c_source.html#l00109
	// though earlier comments suggest the driver ignores rt_format so we could choose whatever.
	unsigned int rt_format;
	int32_t pix_format;
	switch (request->Info.FourCC) {
	case MFX_FOURCC_P010:
		rt_format = VA_RT_FORMAT_YUV420_10;
		pix_format = VA_FOURCC_P010;
		break;
	case MFX_FOURCC_NV12:
	default:
		rt_format = VA_RT_FORMAT_YUV420;
		pix_format = VA_FOURCC_NV12;
		break;
	}

	int num_attrs = 2;
	VASurfaceAttrib attrs[2] = {
		{
			.type = VASurfaceAttribMemoryType,
			.flags = VA_SURFACE_ATTRIB_SETTABLE,
			.value =
				{
					.type = VAGenericValueTypeInteger,
					.value =
						{.i = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2},
				},
		},
		{
			.type = VASurfaceAttribPixelFormat,
			.flags = VA_SURFACE_ATTRIB_SETTABLE,
			.value =
				{
					.type = VAGenericValueTypeInteger,
					.value = {.i = (int)pix_format},
				},
		}};

	unsigned int num_surfaces = request->NumFrameSuggested;
	VASurfaceID temp_surfaces[MAX_ALLOCABLE_SURFACES] = {0};
	assert(num_surfaces < MAX_ALLOCABLE_SURFACES);
	VAStatus vasts;
	if ((vasts = vaCreateSurfaces(display, rt_format, request->Info.Width,
				      request->Info.Height, temp_surfaces,
				      num_surfaces, attrs, num_attrs)) !=
	    VA_STATUS_SUCCESS) {
		blog(LOG_ERROR, "failed to create surfaces: %d", vasts);
		return MFX_ERR_MEMORY_ALLOC;
	}

	// Follow the FFmpeg trick and stuff our pointer at the end.
	mfxMemId *mids =
		(mfxMemId *)bmalloc(sizeof(mfxMemId) * num_surfaces + 1);
	struct surface_info *surfaces = (struct surface_info *)bmalloc(
		sizeof(struct surface_info) * num_surfaces);

	mids[num_surfaces] = surfaces; // stuff it
	for (uint64_t i = 0; i < num_surfaces; i++) {
		surfaces[i].id = temp_surfaces[i];
		surfaces[i].width = request->Info.Width;
		surfaces[i].height = request->Info.Height;
		mids[i] = &surfaces[i];

		VADRMPRIMESurfaceDescriptor surfDesc = {0};
		if (vaExportSurfaceHandle(display, surfaces[i].id,
					  VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
					  VA_EXPORT_SURFACE_READ_WRITE,
					  &surfDesc) != VA_STATUS_SUCCESS)
			return MFX_ERR_MEMORY_ALLOC;

		obs_enter_graphics();
		// TODO: P010 format support
		assert(surfDesc.num_objects == 1);
		int fds[4] = {0};
		uint32_t strides[4] = {0};
		uint32_t offsets[4] = {0};
		uint64_t modifiers[4] = {0};
		fds[0] =
			surfDesc.objects[surfDesc.layers[0].object_index[0]].fd;
		fds[1] =
			surfDesc.objects[surfDesc.layers[1].object_index[0]].fd;
		strides[0] = surfDesc.layers[0].pitch[0];
		strides[1] = surfDesc.layers[1].pitch[0];
		offsets[0] = surfDesc.layers[0].offset[0];
		offsets[1] = surfDesc.layers[1].offset[0];
		modifiers[0] =
			surfDesc.objects[surfDesc.layers[0].object_index[0]]
				.drm_format_modifier;
		modifiers[1] =
			surfDesc.objects[surfDesc.layers[1].object_index[0]]
				.drm_format_modifier;

		surfaces[i].tex_y = gs_texture_create_from_dmabuf(
			surfDesc.width, surfDesc.height,
			surfDesc.layers[0].drm_format, GS_R8, 1, fds, strides,
			offsets, modifiers);
		surfaces[i].tex_uv = gs_texture_create_from_dmabuf(
			surfDesc.width / 2, surfDesc.height,
			surfDesc.layers[1].drm_format, GS_R8G8, 1, fds + 1,
			strides + 1, offsets + 1, modifiers + 1);
		obs_leave_graphics();

		close(surfDesc.objects[surfDesc.layers[0].object_index[0]].fd);
		if (!surfaces[i].tex_y || !surfaces[i].tex_uv) {
			return MFX_ERR_MEMORY_ALLOC;
		}
	}

	response->mids = (mfxMemId *)mids;
	response->NumFrameActual = num_surfaces;
	return MFX_ERR_NONE;
}

mfxStatus simple_lock(mfxHDL pthis, mfxMemId mid, mfxFrameData *ptr)
{
	UNUSED_PARAMETER(pthis);
	UNUSED_PARAMETER(mid);
	UNUSED_PARAMETER(ptr);
	return MFX_ERR_UNSUPPORTED;
}

mfxStatus simple_unlock(mfxHDL pthis, mfxMemId mid, mfxFrameData *ptr)
{
	UNUSED_PARAMETER(pthis);
	UNUSED_PARAMETER(mid);
	UNUSED_PARAMETER(ptr);
	return MFX_ERR_UNSUPPORTED;
}

mfxStatus simple_gethdl(mfxHDL pthis, mfxMemId mid, mfxHDL *handle)
{
	UNUSED_PARAMETER(pthis);
	if (NULL == handle)
		return MFX_ERR_INVALID_HANDLE;

	// Seemingly undocumented, but Pair format defined by
	// oneVPL-intel-gpu-intel-onevpl-23.1.0/_studio/mfx_lib/encode_hw/av1/linux/base/av1ehw_base_va_packer_lin.cpp
	// https://github.com/intel/vpl-gpu-rt/blob/4170dd9fa1ea319dda81b6189616ecc9b178a321/_studio/shared/src/libmfx_core_vaapi.cpp#L1464
	mfxHDLPair *pPair = (mfxHDLPair *)handle;

	// first must be a pointer to a VASurfaceID and will be dereferenced by
	// the driver.
	pPair->first = &((struct surface_info *)mid)->id;
	pPair->second = 0;

	return MFX_ERR_NONE;
}

mfxStatus simple_free(mfxHDL pthis, mfxFrameAllocResponse *response)
{
	if (response->mids == nullptr || response->NumFrameActual == 0)
		return MFX_ERR_NONE;

	mfxSession *session = (mfxSession *)pthis;
	VADisplay display;
	mfxStatus sts =
		MFXVideoCORE_GetHandle(*session, DEVICE_MGR_TYPE, &display);
	MSDK_CHECK_RESULT(sts, MFX_ERR_NONE, sts);

	struct surface_info *surfs =
		(struct surface_info *)response->mids[response->NumFrameActual];
	VASurfaceID temp_surfaces[MAX_ALLOCABLE_SURFACES] = {0};
	obs_enter_graphics();
	for (int i = 0; i < response->NumFrameActual; i++) {
		temp_surfaces[i] = *(VASurfaceID *)response->mids[i];
		gs_texture_destroy(surfs[i].tex_y);
		gs_texture_destroy(surfs[i].tex_uv);
	}
	obs_leave_graphics();

	bfree(surfs);
	bfree(response->mids);
	if (vaDestroySurfaces(display, temp_surfaces,
			      response->NumFrameActual) != VA_STATUS_SUCCESS)
		return MFX_ERR_MEMORY_ALLOC;

	return MFX_ERR_NONE;
}

mfxStatus simple_copytex(mfxHDL pthis, mfxMemId mid, void *tex, mfxU64 lock_key,
			 mfxU64 *next_key)
{
	UNUSED_PARAMETER(lock_key);
	UNUSED_PARAMETER(next_key);

	profile_start("copy_tex");

	mfxSession *session = (mfxSession *)pthis;
	VADisplay display;
	mfxStatus sts =
		MFXVideoCORE_GetHandle(*session, DEVICE_MGR_TYPE, &display);
	MSDK_CHECK_RESULT(sts, MFX_ERR_NONE, sts);

	struct encoder_texture *ptex = (struct encoder_texture *)tex;
	struct surface_info *surf = (struct surface_info *)mid;

	obs_enter_graphics();
	gs_copy_texture(surf->tex_y, ptex->tex[0]);
	gs_copy_texture(surf->tex_uv, ptex->tex[1]);
	obs_leave_graphics();

	profile_end("copy_tex");
	return MFX_ERR_NONE;
}

struct get_drm_device_params {
	const char **out_path;
	uint32_t idx;
};

bool get_drm_device(void *param, const char *node, uint32_t idx)
{
	struct get_drm_device_params *p = (struct get_drm_device_params *)param;
	if (idx == p->idx) {
		*p->out_path = node;
		return false;
	}
	return true;
}

// Initialize Intel VPL Session, device/display and memory manager
mfxStatus Initialize(mfxVersion ver, mfxSession *pSession,
		     mfxFrameAllocator *pmfxAllocator, mfxHDL *deviceHandle,
		     bool bCreateSharedHandles, enum qsv_codec codec,
		     void **data)
{
	UNUSED_PARAMETER(ver);
	UNUSED_PARAMETER(deviceHandle);
	UNUSED_PARAMETER(bCreateSharedHandles);
	mfxStatus sts = MFX_ERR_NONE;
	mfxVariant impl;

	// Initialize Intel VPL Session
	mfxLoader loader = MFXLoad();
	mfxConfig cfg = MFXCreateConfig(loader);

	impl.Type = MFX_VARIANT_TYPE_U32;
	impl.Data.U32 = MFX_IMPL_TYPE_HARDWARE;
	MFXSetConfigFilterProperty(
		cfg, (const mfxU8 *)"mfxImplDescription.Impl", impl);

	impl.Type = MFX_VARIANT_TYPE_U32;
	impl.Data.U32 = INTEL_VENDOR_ID;
	MFXSetConfigFilterProperty(
		cfg, (const mfxU8 *)"mfxImplDescription.VendorID", impl);

	impl.Type = MFX_VARIANT_TYPE_U32;
	impl.Data.U32 = MFX_ACCEL_MODE_VIA_VAAPI_DRM_RENDER_NODE;
	MFXSetConfigFilterProperty(
		cfg, (const mfxU8 *)"mfxImplDescription.AccelerationMode",
		impl);

	const char *device_path = NULL;
	int fd = -1;
	if (pmfxAllocator) {
		obs_video_info ovi;
		obs_get_video_info(&ovi);
		struct get_drm_device_params params = {&device_path,
						       (uint32_t)ovi.adapter};
		obs_enter_graphics();
		gs_enum_adapters(get_drm_device, &params);
		obs_leave_graphics();
	} else {
		if (codec == QSV_CODEC_AVC && default_h264_device)
			device_path = default_h264_device;
		else if (codec == QSV_CODEC_HEVC && default_hevc_device)
			device_path = default_hevc_device;
		else if (codec == QSV_CODEC_AV1 && default_av1_device)
			device_path = default_av1_device;
	}
	fd = open(device_path, O_RDWR);
	if (fd < 0) {
		blog(LOG_ERROR, "Failed to open device '%s'", device_path);
		return MFX_ERR_DEVICE_FAILED;
	}

	mfxHDL vaDisplay = vaGetDisplayDRM(fd);
	if (!vaDisplay) {
		return MFX_ERR_DEVICE_FAILED;
	}

	sts = MFXCreateSession(loader, 0, pSession);
	if (MFX_ERR_NONE > sts) {
		blog(LOG_ERROR, "Failed to initialize MFX");
		MSDK_PRINT_RET_MSG(sts);
		close(fd);
		return sts;
	}

	// VPL expects the VADisplay to be initialized.
	int major;
	int minor;
	if (vaInitialize(vaDisplay, &major, &minor) != VA_STATUS_SUCCESS) {
		blog(LOG_ERROR, "Failed to initialize VA-API");
		vaTerminate(vaDisplay);
		close(fd);
		return MFX_ERR_DEVICE_FAILED;
	}

	sts = MFXVideoCORE_SetHandle(*pSession, DEVICE_MGR_TYPE, vaDisplay);
	MSDK_CHECK_RESULT(sts, MFX_ERR_NONE, sts);

	if (pmfxAllocator) {
		// Allow us to access the session during allocation.
		pmfxAllocator->pthis = pSession;
		pmfxAllocator->Alloc = simple_alloc;
		pmfxAllocator->Free = simple_free;
		pmfxAllocator->Lock = simple_lock;
		pmfxAllocator->Unlock = simple_unlock;
		pmfxAllocator->GetHDL = simple_gethdl;

		sts = MFXVideoCORE_SetFrameAllocator(*pSession, pmfxAllocator);
		MSDK_CHECK_RESULT(sts, MFX_ERR_NONE, sts);
	}

	struct linux_data *d =
		(struct linux_data *)bmalloc(sizeof(struct linux_data));
	d->fd = fd;
	d->vaDisplay = (VADisplay)vaDisplay;
	*data = d;

	return sts;
}

void Release() {}

// Release per session resources.
void ReleaseSessionData(void *data)
{
	struct linux_data *d = (struct linux_data *)data;
	if (d) {
		vaTerminate(d->vaDisplay);
		close(d->fd);
		bfree(d);
	}
}

void mfxGetTime(mfxTime *timestamp)
{
	clock_gettime(CLOCK_MONOTONIC, timestamp);
}

double TimeDiffMsec(mfxTime tfinish, mfxTime tstart)
{
	UNUSED_PARAMETER(tfinish);
	UNUSED_PARAMETER(tstart);
	//unused so far
	return 0.0;
}

extern "C" void util_cpuid(int cpuinfo[4], int level)
{
	__get_cpuid(level, (unsigned int *)&cpuinfo[0],
		    (unsigned int *)&cpuinfo[1], (unsigned int *)&cpuinfo[2],
		    (unsigned int *)&cpuinfo[3]);
}

struct vaapi_device {
	int fd;
	VADisplay display;
	const char *driver;
};

static void vaapi_open(const char *device_path, struct vaapi_device *device)
{
	int fd = open(device_path, O_RDWR);
	if (fd < 0) {
		return;
	}

	VADisplay display = vaGetDisplayDRM(fd);
	if (!display) {
		close(fd);
		return;
	}

	// VA-API is noisy by default.
	vaSetInfoCallback(display, nullptr, nullptr);
	vaSetErrorCallback(display, nullptr, nullptr);

	int major;
	int minor;
	if (vaInitialize(display, &major, &minor) != VA_STATUS_SUCCESS) {
		vaTerminate(display);
		close(fd);
		return;
	}

	const char *driver = vaQueryVendorString(display);
	if (strstr(driver, "Intel i965 driver") != nullptr) {
		blog(LOG_WARNING,
		     "Legacy intel-vaapi-driver detected, incompatible with QSV");
		vaTerminate(display);
		close(fd);
		return;
	}

	device->fd = fd;
	device->display = display;
	device->driver = driver;
}

static void vaapi_close(struct vaapi_device *device)
{
	vaTerminate(device->display);
	close(device->fd);
}

static uint32_t vaapi_check_support(VADisplay display, VAProfile profile,
				    VAEntrypoint entrypoint)
{
	VAConfigAttrib attrib[1];
	attrib->type = VAConfigAttribRateControl;

	VAStatus va_status =
		vaGetConfigAttributes(display, profile, entrypoint, attrib, 1);

	uint32_t rc = 0;
	switch (va_status) {
	case VA_STATUS_SUCCESS:
		rc = attrib->value;
		break;
	case VA_STATUS_ERROR_UNSUPPORTED_PROFILE:
	case VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT:
	default:
		break;
	}

	return (rc & VA_RC_CBR || rc & VA_RC_CQP || rc & VA_RC_VBR);
}

static bool vaapi_supports_h264(VADisplay display)
{
	bool ret = false;
	ret |= vaapi_check_support(display, VAProfileH264ConstrainedBaseline,
				   VAEntrypointEncSlice);
	ret |= vaapi_check_support(display, VAProfileH264Main,
				   VAEntrypointEncSlice);
	ret |= vaapi_check_support(display, VAProfileH264High,
				   VAEntrypointEncSlice);

	if (!ret) {
		ret |= vaapi_check_support(display,
					   VAProfileH264ConstrainedBaseline,
					   VAEntrypointEncSliceLP);
		ret |= vaapi_check_support(display, VAProfileH264Main,
					   VAEntrypointEncSliceLP);
		ret |= vaapi_check_support(display, VAProfileH264High,
					   VAEntrypointEncSliceLP);
	}

	return ret;
}

static bool vaapi_supports_av1(VADisplay display)
{
	bool ret = false;
	// Are there any devices with non-LowPower entrypoints?
	ret |= vaapi_check_support(display, VAProfileAV1Profile0,
				   VAEntrypointEncSlice);
	ret |= vaapi_check_support(display, VAProfileAV1Profile0,
				   VAEntrypointEncSliceLP);
	return ret;
}

static bool vaapi_supports_hevc(VADisplay display)
{
	bool ret = false;
	ret |= vaapi_check_support(display, VAProfileHEVCMain,
				   VAEntrypointEncSlice);
	ret |= vaapi_check_support(display, VAProfileHEVCMain,
				   VAEntrypointEncSliceLP);
	return ret;
}

bool check_adapter(void *param, const char *node, uint32_t idx)
{
	struct vaapi_device device = {0};
	struct adapter_info *adapters = (struct adapter_info *)param;

	vaapi_open(node, &device);
	if (!device.display)
		return true;

	struct adapter_info *adapter = &adapters[idx];
	adapter->is_intel = strstr(device.driver, "Intel") != nullptr;
	// This is currently only used for LowPower coding which is busted on VA-API anyway.
	adapter->is_dgpu = false;
	adapter->supports_av1 = vaapi_supports_av1(device.display);
	adapter->supports_hevc = vaapi_supports_hevc(device.display);

	if (adapter->is_intel && default_h264_device == nullptr)
		default_h264_device = strdup(node);

	if (adapter->is_intel && adapter->supports_av1 &&
	    default_av1_device == nullptr)
		default_av1_device = strdup(node);

	if (adapter->is_intel && adapter->supports_hevc &&
	    default_hevc_device == nullptr)
		default_hevc_device = strdup(node);

	vaapi_close(&device);
	return true;
}

void check_adapters(struct adapter_info *adapters, size_t *adapter_count)
{
	obs_enter_graphics();
	uint32_t gs_count = gs_get_adapter_count();
	if (*adapter_count < gs_count) {
		blog(LOG_WARNING, "Too many video adapters: %ld < %d",
		     *adapter_count, gs_count);
		obs_leave_graphics();
		return;
	}
	gs_enum_adapters(check_adapter, adapters);
	obs_leave_graphics();
}