Skip to content

Commit 32a7916

Browse files
devshgraphicsprogrammingdevshgraphicsprogramming
committed
Make sure to use steady clock, high resolution was responsible for time going backwards and deadlocks on GLsync waits.
GLES backend has a bug where the tests timeout!? Vulkan has a bug because new Buffer API is not ready yet.
1 parent 0f65fba commit 32a7916

File tree

11 files changed

+23
-16
lines changed

11 files changed

+23
-16
lines changed

‎examples_tests/48.ArithmeticUnitTest/main.cpp

Lines changed: 5 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -201,6 +201,7 @@ bool validateResults(ILogicalDevice* device, const uint32_t* inputData, const ui
201201
auto dataFromBuffer = reinterpret_cast<uint32_t*>(mem->getMappedPointer());
202202
const uint32_t subgroupSize = (*dataFromBuffer++);
203203

204+
// TODO: parallel for
204205
// now check if the data obtained has valid values
205206
uint32_t* tmp = new uint32_t[workgroupSize];
206207
uint32_t* ballotInput = new uint32_t[workgroupSize];
@@ -281,7 +282,7 @@ bool runTest(
281282
passed = validateResults<Arithmetic,max_op>(device, inputData, workgroupSize, workgroupCount, buffers[6].get(),logger)&&passed;
282283
if(is_workgroup_test)
283284
{
284-
passed = validateResults<Arithmetic,ballot>(device, inputData, workgroupSize, workgroupCount, buffers[7].get(),logger) && passed;
285+
passed = validateResults<Arithmetic,ballot>(device, inputData, workgroupSize, workgroupCount, buffers[7].get(),logger)&&passed;
285286
}
286287

287288
return passed;
@@ -300,7 +301,7 @@ class ArythmeticUnitTestApp : public NonGraphicalApplicationBase
300301
void onAppInitialized_impl() override
301302
{
302303
CommonAPI::InitOutput initOutput;
303-
CommonAPI::InitWithNoExt(initOutput, nbl::video::EAT_OPENGL, "Subgroup Arithmetic Test");
304+
CommonAPI::InitWithNoExt(initOutput, video::EAT_OPENGL, "Subgroup Arithmetic Test");
304305
gl = std::move(initOutput.apiConnection);
305306
gpuPhysicalDevice = std::move(initOutput.physicalDevice);
306307
logicalDevice = std::move(initOutput.logicalDevice);
@@ -335,7 +336,7 @@ class ArythmeticUnitTestApp : public NonGraphicalApplicationBase
335336
params.queueFamilyIndexCount = 0;
336337
params.queueFamilyIndices = nullptr;
337338
params.sharingMode = ESM_CONCURRENT;
338-
params.usage = IGPUBuffer::EUF_STORAGE_BUFFER_BIT;
339+
params.usage = core::bitflag(IGPUBuffer::EUF_STORAGE_BUFFER_BIT)|IGPUBuffer::EUF_TRANSFER_SRC_BIT;
339340
IDriverMemoryBacked::SDriverMemoryRequirements reqs;
340341
reqs.vulkanReqs.memoryTypeBits = ~0u;
341342
reqs.vulkanReqs.alignment = 256u;
@@ -424,7 +425,7 @@ class ArythmeticUnitTestApp : public NonGraphicalApplicationBase
424425
core::smart_refctd_ptr<IGPUCommandBuffer> cmdbuf;
425426
logicalDevice->createCommandBuffers(cmdPool.get(), IGPUCommandBuffer::EL_PRIMARY, 1u, &cmdbuf);
426427
computeQueue->startCapture();
427-
for (uint32_t workgroupSize = 1u; workgroupSize <= 1024u; workgroupSize++)
428+
for (uint32_t workgroupSize=45u; workgroupSize<=1024u; workgroupSize++)
428429
{
429430
core::smart_refctd_ptr<IGPUComputePipeline> pipelines[kTestTypeCount];
430431
for (uint32_t i = 0u; i < kTestTypeCount; i++)

‎examples_tests/Readme.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -79,7 +79,7 @@ N = No support
7979
| 45.BRDFEvalTest | S | S | S | S | S | S | N | S | |
8080
| 46.SamplingValidation | S | S | S | S | S | S | N | S | |
8181
| 47.DerivMapTest | Y | Y | B | S | N | S | N | N | |
82-
| 48.ArithmeticUnitTest | Y | Y | Y | S | N | S | N | S | |
82+
| 48.ArithmeticUnitTest | Y | B | B | S | S | S | N | S | |
8383
| 49.ComputeFFT | S | N | S | S | N | S | N | N | |
8484
| 50.NewAPITest | W | W | W | W | W | W | W | W | |
8585
| 51.RadixSort | W | N | W | W | N | W | N | W | |

‎include/nbl/core/EventDeferredHandler.h

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -105,7 +105,7 @@ class DeferredEventHandlerST
105105
auto prev = mEvents.before_begin();
106106
for (auto it = mEvents.begin(); it!=mEvents.end(); )
107107
{
108-
if (it->m_event.wait_until(std::chrono::high_resolution_clock::now()+std::chrono::microseconds(250ull)))
108+
if (it->m_event.wait_until(std::chrono::steady_clock::now()+std::chrono::microseconds(250ull)))
109109
{
110110
it->m_function();
111111
it = mEvents.erase_after(prev);

‎include/nbl/video/IGPUFence.h

Lines changed: 3 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -65,10 +65,11 @@ class GPUEventWrapper : public core::Uncopyable
6565
return *this;
6666
}
6767

68-
template<class Clock=std::chrono::high_resolution_clock, class Duration=typename Clock::duration>
68+
template<class Clock=std::chrono::steady_clock, class Duration=typename Clock::duration>
6969
inline static std::chrono::time_point<Clock,Duration> default_wait()
7070
{
71-
return std::chrono::high_resolution_clock::now()+std::chrono::nanoseconds(50000ull); // 50 us
71+
//return typename Clock::now()+std::chrono::nanoseconds(50000ull); // 50 us
72+
return Clock::now()+std::chrono::nanoseconds(50000ull); // 50 us
7273
}
7374

7475
IGPUFence::E_STATUS waitFenceWrapper(IGPUFence* fence, uint64_t timeout);

‎include/nbl/video/alloc/SubAllocatedDataBuffer.h

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -227,7 +227,7 @@ class SubAllocatedDataBuffer : protected core::impl::FriendOfHeterogenousMemoryA
227227
return multi_alloc(GPUEventWrapper::default_wait(),count,std::forward<Args>(args)...);
228228
}
229229
//!
230-
template<class Clock=typename std::chrono::high_resolution_clock, typename... Args>
230+
template<class Clock=typename std::chrono::steady_clock, typename... Args>
231231
inline size_type multi_alloc(const std::chrono::time_point<Clock>& maxWaitPoint, const Args&... args) noexcept
232232
{
233233
#ifdef _NBL_DEBUG

‎include/nbl/video/utilities/CPropertyPoolHandler.h

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -187,7 +187,7 @@ class CPropertyPoolHandler final : public core::IReferenceCounted, public core::
187187
StreamingTransientDataBufferMT<>* const upBuff, IGPUCommandBuffer* const cmdbuf, IGPUFence* const fence, IGPUQueue* const queue,
188188
const asset::SBufferBinding<video::IGPUBuffer>& scratch, UpStreamingRequest*& requests, const uint32_t requestCount,
189189
uint32_t& waitSemaphoreCount, IGPUSemaphore* const*& semaphoresToWaitBeforeOverwrite, const asset::E_PIPELINE_STAGE_FLAGS*& stagesToWaitForPerSemaphore,
190-
system::logger_opt_ptr logger, const std::chrono::high_resolution_clock::time_point& maxWaitPoint=std::chrono::high_resolution_clock::now()+std::chrono::microseconds(500u)
190+
system::logger_opt_ptr logger, const std::chrono::steady_clock::time_point& maxWaitPoint=std::chrono::steady_clock::now()+std::chrono::microseconds(500u)
191191
);
192192

193193
// utility to help you fill out the tail move scatter request after the free, properly, returns if you actually need to transfer anything

‎include/nbl/video/utilities/IUtilities.h

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -317,7 +317,7 @@ namespace nbl::video
317317
const uint32_t subSize = core::min(alllocationSize,size);
318318
// cannot use `multi_place` because of the extra padding size we could have added
319319
uint32_t localOffset = video::StreamingTransientDataBufferMT<>::invalid_address;
320-
m_defaultUploadBuffer.get()->multi_alloc(std::chrono::high_resolution_clock::now()+std::chrono::microseconds(500u),1u,&localOffset,&alllocationSize,&alignment);
320+
m_defaultUploadBuffer.get()->multi_alloc(std::chrono::steady_clock::now()+std::chrono::microseconds(500u),1u,&localOffset,&alllocationSize,&alignment);
321321
// copy only the unpadded part
322322
if (localOffset != video::StreamingTransientDataBufferMT<>::invalid_address)
323323
{

‎src/nbl/video/COpenGLSync.cpp

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -32,7 +32,7 @@ uint64_t COpenGLSync::prewait() const
3232
if (sync)
3333
return 0ull;
3434

35-
using clock_t = std::chrono::high_resolution_clock;
35+
using clock_t = std::chrono::steady_clock;
3636
auto start = clock_t::now();
3737
while (!sync)
3838
{
@@ -53,7 +53,7 @@ COpenGLSync::E_STATUS COpenGLSync::waitCPU(IOpenGL_FunctionTable* _gl, uint64_t
5353
return (cachedRetval = ES_TIMEOUT_EXPIRED);
5454
timeout -= spintime;
5555

56-
GLenum status = _gl->glSync.pglClientWaitSync(sync, lockedTable?GL_SYNC_FLUSH_COMMANDS_BIT:0, timeout); // GL_SYNC_FLUSH_COMMANDS_BIT to flags?
56+
GLenum status = _gl->glSync.pglClientWaitSync(sync, lockedTable ? GL_SYNC_FLUSH_COMMANDS_BIT:0, timeout);
5757
switch (status)
5858
{
5959
case GL_ALREADY_SIGNALED:

‎src/nbl/video/IOpenGL_LogicalDevice.h

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -852,7 +852,12 @@ class IOpenGL_LogicalDevice : public ILogicalDevice, protected impl::IOpenGL_Log
852852
if(notFirstRun && now>=timeoutPoint)
853853
return IGPUFence::ES_TIMEOUT;
854854
else if (_waitAll) // all fences have to get signalled anyway so no use round robining
855-
timeout = std::chrono::duration_cast<std::chrono::nanoseconds>(timeoutPoint-now).count();
855+
{
856+
if (timeoutPoint>now)
857+
timeout = std::chrono::duration_cast<std::chrono::nanoseconds>(timeoutPoint-now).count();
858+
else
859+
timeout = 0ull;
860+
}
856861
else if (i==0u) // if we're only looking for one to succeed then poll with increasing timeouts until deadline
857862
timeout <<= 1u;
858863
}

‎src/nbl/video/utilities/CPropertyPoolHandler.cpp

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -174,7 +174,7 @@ uint32_t CPropertyPoolHandler::transferProperties(
174174
StreamingTransientDataBufferMT<>* const upBuff, IGPUCommandBuffer* const cmdbuf, IGPUFence* const fence, IGPUQueue* const queue,
175175
const asset::SBufferBinding<video::IGPUBuffer>& scratch, UpStreamingRequest* &requests, const uint32_t requestCount,
176176
uint32_t& waitSemaphoreCount, IGPUSemaphore* const*& semaphoresToWaitBeforeOverwrite, const asset::E_PIPELINE_STAGE_FLAGS*& stagesToWaitForPerSemaphore,
177-
system::logger_opt_ptr logger, const std::chrono::high_resolution_clock::time_point& maxWaitPoint
177+
system::logger_opt_ptr logger, const std::chrono::steady_clock::time_point& maxWaitPoint
178178
)
179179
{
180180
if (!requestCount)

0 commit comments

Comments
 (0)