--device autoでのGPU自動選択の高度化。 ( #225 )

共有メモリ上にGPUの使用状況を登録して参照することで精度を向上させる。

QSVEncC/QSVEncC.cpp

@@ -1200,7 +1200,12 @@ int run(int argc, TCHAR *argv[]) {
         _tsetlocale(LC_ALL, _T("Japanese"));
     }
 #endif //#if defined(_WIN32) || defined(_WIN64)
-
+    if (Params.ctrl.processMonitorDevUsageReset) {
+        return processMonitorRGYDeviceResetEntry();
+    }
+    if (Params.ctrl.processMonitorDevUsage) {
+        return processMonitorRGYDeviceUsage(Params.ctrl.parentProcessID, (int)Params.device);
+    }
     if (Params.bBenchmark) {
         return run_benchmark(&Params);
     }

QSVPipeline/QSVPipeline.vcxproj

@@ -602,6 +602,7 @@ if exist rgy_rev.h.%PID%.tmp del rgy_rev.h.%PID%.tmp &gt; nul 2&gt;&amp;1</Comma
     <ClCompile Include="rgy_cmd.cpp" />
     <ClCompile Include="rgy_codepage.cpp" />
     <ClCompile Include="rgy_def.cpp" />
+    <ClCompile Include="rgy_device_usage.cpp" />
     <ClCompile Include="rgy_device_vulkan.cpp" />
     <ClCompile Include="rgy_dummy_load.cpp">
       <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>
@@ -1033,6 +1034,7 @@ if exist rgy_rev.h.%PID%.tmp del rgy_rev.h.%PID%.tmp &gt; nul 2&gt;&amp;1</Comma
     <ClInclude Include="rgy_cmd.h" />
     <ClInclude Include="rgy_codepage.h" />
     <ClInclude Include="rgy_def.h" />
+    <ClInclude Include="rgy_device_usage.h" />
     <ClInclude Include="rgy_device_vulkan.h" />
     <ClInclude Include="rgy_dummy_load.h" />
     <ClInclude Include="rgy_env.h" />

QSVPipeline/QSVPipeline.vcxproj.filters

@@ -372,6 +372,9 @@
     <ClCompile Include="rgy_libplacebo.cpp">
       <Filter>ソース ファイル</Filter>
     </ClCompile>
+    <ClCompile Include="rgy_device_usage.cpp">
+      <Filter>ソース ファイル</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="qsv_prm.h">
@@ -716,6 +719,9 @@
     <ClInclude Include="rgy_device_vulkan.h">
       <Filter>ヘッダー ファイル</Filter>
     </ClInclude>
+    <ClInclude Include="rgy_device_usage.h">
+      <Filter>ヘッダー ファイル</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <None Include="rgy_filter.cl">

QSVPipeline/qsv_pipeline.cpp

@@ -3377,6 +3377,18 @@ RGY_ERR CQSVPipeline::deviceAutoSelect(const sInputParams *prm, std::vector<std:
     if (gpuList.size() <= 1) {
         return RGY_ERR_NONE;
     }
+    int maxDeviceUsageCount = 1;
+    std::vector<std::pair<int, int64_t>> deviceUsage;
+    if (gpuList.size() > 1) {
+        RGYDeviceUsage devUsage;
+        deviceUsage = devUsage.getUsage();
+        for (size_t i = 0; i < deviceUsage.size(); i++) {
+            maxDeviceUsageCount = std::max(maxDeviceUsageCount, deviceUsage[i].first);
+            if (deviceUsage[i].first > 0) {
+                PrintMes(RGY_LOG_DEBUG, _T("Device #%d: %d usage.\n"), i, deviceUsage[i].first);
+            }
+        }
+    }
 #if ENABLE_PERF_COUNTER
     PrintMes(RGY_LOG_DEBUG, _T("Auto select device from %d devices.\n"), (int)gpuList.size());
     bool counterIsIntialized = m_pPerfMonitor->isPerfCounterInitialized();
@@ -3403,15 +3415,17 @@ RGY_ERR CQSVPipeline::deviceAutoSelect(const sInputParams *prm, std::vector<std:
             RGYGPUCounterWinEntries(counters).filter_type(L"compute").max()), //vce-opencl
             RGYGPUCounterWinEntries(counters).filter_type(L"3d").max()), //qsv
             RGYGPUCounterWinEntries(counters).filter_type(L"videoprocessing").max());
+        const int deviceUsageCount = (int)gpu->deviceNum() < (int)deviceUsage.size() ? deviceUsage[(int)gpu->deviceNum()].first : 0;
+        double usage_score = 100.0 * (maxDeviceUsageCount - deviceUsageCount) / (double)maxDeviceUsageCount;
         double core_score = 0.0;
         double cc_score = 0.0;
         double ve_score = 100.0 * (1.0 - std::pow(ve_utilization / 100.0, 1.0)) * prm->ctrl.gpuSelect.ve;
         double gpu_score = 100.0 * (1.0 - std::pow(gpu_utilization / 100.0, 1.5)) * prm->ctrl.gpuSelect.gpu;
-        double cl_score = gpu->devInfo() ? 0.0 : -100.0; // openclの初期化に成功したか?
+        double cl_score = gpu->devInfo() ? 0.0 : maxDeviceUsageCount * -100.0; // openclの初期化に成功したか?
 
-        gpuscore[gpu->deviceNum()] = cc_score + ve_score + gpu_score + core_score + cl_score;
-        PrintMes(RGY_LOG_DEBUG, _T("GPU #%d (%s) score: %.1f: VE %.1f, GPU %.1f, CC %.1f, Core %.1f, CL %.1f.\n"), gpu->deviceNum(), gpu->name().c_str(),
-            gpuscore[gpu->deviceNum()], ve_score, gpu_score, cc_score, core_score, cl_score);
+        gpuscore[gpu->deviceNum()] = usage_score + cc_score + ve_score + gpu_score + core_score + cl_score;
+        PrintMes(RGY_LOG_DEBUG, _T("GPU #%d (%s) score: %.1f: Use: %.1f, VE %.1f, GPU %.1f, CC %.1f, Core %.1f, CL %.1f.\n"), gpu->deviceNum(), gpu->name().c_str(),
+            gpuscore[gpu->deviceNum()], usage_score, ve_score, gpu_score, cc_score, core_score, cl_score);
     }
     std::sort(gpuList.begin(), gpuList.end(), [&](const std::unique_ptr<QSVDevice> &a, const std::unique_ptr<QSVDevice> &b) {
         if (gpuscore.at(a->deviceNum()) != gpuscore.at(b->deviceNum())) {
@@ -3430,6 +3444,7 @@ RGY_ERR CQSVPipeline::deviceAutoSelect(const sInputParams *prm, std::vector<std:
 
 RGY_ERR CQSVPipeline::InitSession(const sInputParams *inputParam, std::vector<std::unique_ptr<QSVDevice>>& deviceList) {
     auto err = RGY_ERR_NONE;
+    const int deviceCount = (int)deviceList.size();
     if (deviceList.size() == 0) {
         PrintMes(RGY_LOG_DEBUG, _T("No device found for QSV encoding!\n"));
         return RGY_ERR_DEVICE_NOT_FOUND;
@@ -3445,6 +3460,10 @@ RGY_ERR CQSVPipeline::InitSession(const sInputParams *inputParam, std::vector<st
         m_device = std::move(deviceList.front());
         PrintMes(RGY_LOG_DEBUG, _T("InitSession: selected device #%d: %s.\n"), (int)m_device->deviceNum(), m_device->name().c_str());
     }
+    if (deviceCount > 1) {
+        RGYDeviceUsage devUsage;
+        devUsage.startProcessMonitor((int)m_device->deviceNum());
+    }
 
     //使用できる最大のversionをチェック
     m_device->mfxSession().QueryVersion(&m_mfxVer);

QSVPipeline/qsv_pipeline.h

@@ -55,6 +55,7 @@
 #include "rgy_output.h"
 #include "rgy_opencl.h"
 #include "rgy_dummy_load.h"
+#include "rgy_device_usage.h"
 #include "qsv_vpp_mfx.h"
 #include "qsv_mfx_dec.h"
 #include "qsv_pipeline_ctrl.h"

QSVPipeline/rgy_cmd.cpp

@@ -6988,6 +6988,14 @@ int parse_one_ctrl_option(const TCHAR *option_name, const TCHAR *strInput[], int
         ctrl->enableVulkan = true;
         return 0;
     }
+    if (IS_OPTION("process-monitor-dev-usage")) {
+        ctrl->processMonitorDevUsage = true;
+        return 0;
+    }
+    if (IS_OPTION("process-monitor-dev-usage-reset")) {
+        ctrl->processMonitorDevUsageReset = true;
+        return 0;
+    }
     return -10;
 }
 
@@ -8357,7 +8365,9 @@ tstring gen_cmd(const RGYParamControl *param, const RGYParamControl *defaultPrm,
         }
     }
     OPT_NUM(_T("--perf-monitor-interval"), perfMonitorInterval);
-    OPT_NUM(_T("--parent-pid"), parentProcessID);
+    if (param->parentProcessID != defaultPrm->parentProcessID) {
+        cmd << strsprintf(_T(" --parent-pid %x"), param->parentProcessID);
+    }
     if (param->gpuSelect != defaultPrm->gpuSelect) {
         std::basic_stringstream<TCHAR> tmp;
         tmp.str(tstring());
@@ -8373,6 +8383,8 @@ tstring gen_cmd(const RGYParamControl *param, const RGYParamControl *defaultPrm,
     OPT_BOOL(_T("--enable-opencl"), _T("--disable-opencl"), enableOpenCL);
 #endif
     OPT_BOOL(_T("--enable-vulkan"), _T("--disable-vulkan"), enableVulkan);
+    OPT_BOOL(_T("--process-monitor-dev-usage"), _T(""), processMonitorDevUsage);
+    OPT_BOOL(_T("--process-monitor-dev-usage-reset"), _T(""), processMonitorDevUsageReset);
     return cmd.str();
 }