AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

P4 to Git Change 1704517 by gandryey@gera-w8 on 2018/11/08 15:15:03

Browse Source 
SWDEV-165161 - Update OpenCL runtime to control RGP capture based on new RGP protocol.
	- Added logic to control the dispatch capture frame.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.cpp#11 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palgpuopen.hpp#8 edit


[ROCm/clr commit: c80c8964ff]
This commit is contained in:
foreman
2018-11-08 15:20:47 -05:00
parent 009cd3f18d
commit b953979eff
2 changed files with 17 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/rocclr/runtime/device/pal/palgpuopen.cpp
+9 -2
View File
@@ -40,6 +40,7 @@ RgpCaptureMgr::RgpCaptureMgr(Pal::IPlatform* platform, const Device& device)
device_(device),
dev_driver_server_(platform->GetDevDriverServer()),
num_prep_disp_(0),
max_sqtt_disp_(device_.settings().rgpSqttDispCount_),
trace_gpu_mem_limit_(0),
global_disp_count_(1), // Must start from 1 according to RGP spec
user_event_(nullptr),
@@ -252,7 +253,7 @@ void RgpCaptureMgr::PostDispatch(VirtualGPU* gpu)
if (trace_.status_ == TraceStatus::Running) {
amd::ScopedLock traceLock(&trace_mutex_);
trace_.sqtt_disp_count_++;
if (trace_.sqtt_disp_count_ >= device_.settings().rgpSqttDispCount_) {
if (trace_.sqtt_disp_count_ >= max_sqtt_disp_) {
Pal::Result res = EndRGPHardwareTrace(gpu);
if (Pal::Result::ErrorIncompatibleQueue == res) {
// continue until we find the right queue...
@@ -440,7 +441,13 @@ Pal::Result RgpCaptureMgr::PrepareRGPTrace(VirtualGPU* gpu)
const auto traceParameters = rgp_server_->QueryTraceParameters();
num_prep_disp_ = traceParameters.numPreparationFrames;
num_prep_disp_ = traceParameters.captureStartIndex;
uint32_t capture_disp = traceParameters.captureStopIndex - traceParameters.captureStartIndex;
// Validate if the captured dispatches are in the range
if ((capture_disp > 0) && (capture_disp < max_sqtt_disp_)) {
max_sqtt_disp_ = capture_disp;
}
trace_gpu_mem_limit_ = traceParameters.gpuMemoryLimitInMb * 1024 * 1024;
inst_tracing_enabled_ = traceParameters.flags.enableInstructionTokens;
projects/clr/rocclr/runtime/device/pal/palgpuopen.hpp
+8 -5
View File
@@ -384,12 +384,15 @@ private:
DevDriver::RGPProtocol::RGPServer* rgp_server_;
mutable amd::Monitor trace_mutex_;
TraceState trace_;
uint32_t num_prep_disp_;
uint32_t trace_gpu_mem_limit_;
uint32_t global_disp_count_;
RgpSqttMarkerUserEventWithString* user_event_;
bool trace_enabled_; // True if tracing is currently enabled (master flag)
bool inst_tracing_enabled_; // Enable instruction-level SQTT tokens
uint32_t num_prep_disp_;
uint32_t max_sqtt_disp_; // Maximum number of the dispatches allowed in the trace
uint32_t trace_gpu_mem_limit_;
uint32_t global_disp_count_;
bool trace_enabled_; // True if tracing is currently enabled (master flag)
bool inst_tracing_enabled_; // Enable instruction-level SQTT tokens
PAL_DISALLOW_DEFAULT_CTOR(RgpCaptureMgr);
PAL_DISALLOW_COPY_AND_ASSIGN(RgpCaptureMgr);