intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-25 05:24:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix xe: add missing wait on fence before unbind

Browse Source 
Related-To: NEO-7306
Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski
2023-03-17 13:00:44 +00:00
committed by Compute-Runtime-Automation
parent 2d0c61aa33
commit 415d2cb121
9 changed files with 57 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
shared/source/os_interface/linux/drm_neo.cpp
View File

@@ -1454,11 +1454,11 @@ int changeBufferObjectBinding(Drm *drm, OsContext *osContext, uint32_t vmHandleI
vmBind.extensions = castToUint64(extensions.get());
}
if (bind) {
std::unique_lock<std::mutex> lock;
std::unique_lock<std::mutex> lock;
VmBindExtUserFenceT vmBindExtUserFence{};
VmBindExtUserFenceT vmBindExtUserFence{};
if (ioctlHelper->isWaitBeforeBindRequired(bind)) {
if (drm->useVMBindImmediate()) {
lock = drm->lockBindFenceMutex();
@@ -1471,7 +1471,8 @@ int changeBufferObjectBinding(Drm *drm, OsContext *osContext, uint32_t vmHandleI
vmBind.extensions = castToUint64(vmBindExtUserFence);
}
}
}
if (bind) {
ret = ioctlHelper->vmBind(vmBind);
if (ret) {

3
shared/source/os_interface/linux/ioctl_helper.h
View File

@@ -137,6 +137,7 @@ class IoctlHelper {
virtual std::string getFileForMaxGpuFrequencyOfSubDevice(int subDeviceId) const;
virtual std::string getFileForMaxMemoryFrequencyOfSubDevice(int subDeviceId) const;
virtual bool getFabricLatency(uint32_t fabricId, uint32_t &latency, uint32_t &bandwidth) = 0;
virtual bool isWaitBeforeBindRequired(bool bind) const = 0;
uint32_t getFlagsForPrimeHandleToFd() const;
@@ -194,6 +195,7 @@ class IoctlHelperUpstream : public IoctlHelper {
std::string getDrmParamString(DrmParam param) const override;
std::string getIoctlString(DrmIoctl ioctlRequest) const override;
bool getFabricLatency(uint32_t fabricId, uint32_t &latency, uint32_t &bandwidth) override;
bool isWaitBeforeBindRequired(bool bind) const override;
};
template <PRODUCT_FAMILY gfxProduct>
@@ -261,6 +263,7 @@ class IoctlHelperPrelim20 : public IoctlHelper {
std::string getIoctlString(DrmIoctl ioctlRequest) const override;
bool checkIfIoctlReinvokeRequired(int error, DrmIoctl ioctlRequest) const override;
bool getFabricLatency(uint32_t fabricId, uint32_t &latency, uint32_t &bandwidth) override;
bool isWaitBeforeBindRequired(bool bind) const override;
protected:
bool queryHwIpVersion(EngineClassInstance &engineInfo, HardwareIpVersion &ipVersion, int &ret);

5
shared/source/os_interface/linux/ioctl_helper_prelim.cpp
View File

@@ -708,6 +708,11 @@ bool IoctlHelperPrelim20::getFabricLatency(uint32_t fabricId, uint32_t &latency,
bandwidth = info.bandwidth;
return true;
}
bool IoctlHelperPrelim20::isWaitBeforeBindRequired(bool bind) const {
return bind;
}
bool IoctlHelperPrelim20::queryHwIpVersion(EngineClassInstance &engineInfo, HardwareIpVersion &ipVersion, int &ret) {
QueryItem queryItem{};
queryItem.queryId = PRELIM_DRM_I915_QUERY_HW_IP_VERSION;

4
shared/source/os_interface/linux/ioctl_helper_upstream.cpp
View File

@@ -239,4 +239,8 @@ std::string IoctlHelperUpstream::getIoctlString(DrmIoctl ioctlRequest) const {
bool IoctlHelperUpstream::getFabricLatency(uint32_t fabricId, uint32_t &latency, uint32_t &bandwidth) {
return false;
}
bool IoctlHelperUpstream::isWaitBeforeBindRequired(bool bind) const {
return false;
}
} // namespace NEO

21
shared/source/os_interface/linux/xe/ioctl_helper_xe.cpp
View File

@@ -1177,21 +1177,15 @@ int IoctlHelperXe::xeVmBind(const VmBindParams &vmBindParams, bool bindOp) {
}
if (found != -1) {
uint32_t extraBindFlag = 0;
uint64_t waitfence = 0ull;
struct drm_xe_sync sync[1] = {};
sync[0].flags = DRM_XE_SYNC_USER_FENCE | DRM_XE_SYNC_SIGNAL;
extraBindFlag = XE_VM_BIND_FLAG_ASYNC;
auto xeBindExtUserFence = reinterpret_cast<xe_fake_ext_user_fence *>(vmBindParams.extensions);
UNRECOVERABLE_IF(!xeBindExtUserFence);
UNRECOVERABLE_IF(xeBindExtUserFence->tag != VMBIND_FENCE_TAG);
sync[0].addr = xeBindExtUserFence->addr;
sync[0].timeline_value = xeBindExtUserFence->value;
if (bindOp) {
auto xeBindExtUserFence = reinterpret_cast<xe_fake_ext_user_fence *>(vmBindParams.extensions);
UNRECOVERABLE_IF(!xeBindExtUserFence);
UNRECOVERABLE_IF(xeBindExtUserFence->tag != VMBIND_FENCE_TAG);
sync[0].addr = xeBindExtUserFence->addr;
sync[0].timeline_value = xeBindExtUserFence->value;
} else {
sync[0].addr = castToUint64(&waitfence);
sync[0].timeline_value = USER_FENCE_VALUE | vmBindParams.handle;
}
struct drm_xe_vm_bind bind = {};
bind.vm_id = vmBindParams.vmId;
bind.num_binds = 1;
@@ -1209,7 +1203,6 @@ int IoctlHelperXe::xeVmBind(const VmBindParams &vmBindParams, bool bindOp) {
}
if (!bindOp) {
bind.bind.op = XE_VM_BIND_OP_UNMAP;
bind.bind.obj = bindInfo[found].handle;
bind.bind.obj = 0;
if (bindInfo[found].handle & XE_USERPTR_FAKE_FLAG) {
bind.bind.obj_offset = bindInfo[found].userptr;
@@ -1370,6 +1363,10 @@ bool IoctlHelperXe::getFabricLatency(uint32_t fabricId, uint32_t &latency, uint3
return false;
}
bool IoctlHelperXe::isWaitBeforeBindRequired(bool bind) const {
return true;
}
static uint32_t getVectorGetMax(std::vector<uint8_t> *data) {
uint32_t ret = 0;
for (uint32_t i = 0; i < static_cast<uint32_t>(data->size()); i++) {

1
shared/source/os_interface/linux/xe/ioctl_helper_xe.h
View File

@@ -92,6 +92,7 @@ class IoctlHelperXe : public IoctlHelper {
std::string getFileForMaxGpuFrequencyOfSubDevice(int subDeviceId) const override;
std::string getFileForMaxMemoryFrequencyOfSubDevice(int subDeviceId) const override;
bool getFabricLatency(uint32_t fabricId, uint32_t &latency, uint32_t &bandwidth) override;
bool isWaitBeforeBindRequired(bool bind) const override;
std::vector<uint8_t> xeRebuildi915Topology(std::vector<uint8_t> *geomDss, std::vector<uint8_t> *computeDss, std::vector<uint8_t> *euDss);