Adding bindless compilation toggles

Change-Id: I35e37e9319a660fe0e4588f7abdc821557948b18

Author: jchodor

runtime/os_interface/debug_variables_base.inl

@@ -105,6 +105,8 @@ DECLARE_DEBUG_VARIABLE(bool, AddClGlSharing, false, "Add cl-gl extension")
 DECLARE_DEBUG_VARIABLE(bool, EnableFormatQuery, false, "Enable sharing format querying")
 DECLARE_DEBUG_VARIABLE(bool, EnableFreeMemory, false, "Enable freeMemory in memory manager")
 DECLARE_DEBUG_VARIABLE(bool, ForceSamplerLowFilteringPrecision, false, "Force Low Filtering Precision Sampler mode")
+DECLARE_DEBUG_VARIABLE(bool, UseBindlessBuffers, false, "Force compiler to use bindless buffer addressing instead of stateful one")
+DECLARE_DEBUG_VARIABLE(bool, UseBindlessImages, false, "Force compiler to use bindless image addressing instead of stateful one")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableIntelVme, -1, "-1: default, 0: disabled, 1: Enables cl_intel_motion_estimation extension")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableIntelAdvancedVme, -1, "-1: default, 0: disabled, 1: Enables cl_intel_advanced_motion_estimation extension")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableBlitterOperationsSupport, -1, "-1: default, 0: disable, 1: enable")

runtime/program/program.cpp

@@ -79,6 +79,15 @@ Program::Program(ExecutionEnvironment &executionEnvironment, Context *context, b
             DebugManager.flags.DisableStatelessToStatefulOptimization.get()) {
             internalOptions += "-cl-intel-greater-than-4GB-buffer-required ";
         }
+
+        if (DebugManager.flags.UseBindlessBuffers.get()) {
+            internalOptions += "-cl-intel-use-bindless-buffers ";
+        }
+
+        if (DebugManager.flags.UseBindlessImages.get()) {
+            internalOptions += "-cl-intel-use-bindless-images ";
+        }
+
         kernelDebugEnabled = pDevice->isSourceLevelDebuggerActive();
 
         auto enableStatelessToStatefullWithOffset = pDevice->getHardwareCapabilities().isStatelesToStatefullWithOffsetSupported;

unit_tests/program/program_tests.cpp

@@ -1763,6 +1763,42 @@ TEST_F(ProgramTests, ProgramCtorSetsProperInternalOptionsWhenStatelessToStateful
     }
 }
 
+TEST_F(ProgramTests, WhenCreatingProgramThenBindlessIsEnabledOnlyIfDebugFlagIsEnabled) {
+    using namespace testing;
+    DebugManagerStateRestore restorer;
+
+    {
+        EXPECT_FALSE(DebugManager.flags.UseBindlessBuffers.get());
+        EXPECT_FALSE(DebugManager.flags.UseBindlessImages.get());
+        MockProgram programNoBindless(*pDevice->getExecutionEnvironment(), pContext, false);
+        EXPECT_THAT(programNoBindless.getInternalOptions(), Not(HasSubstr(std::string("-cl-intel-use-bindless-buffers"))));
+        EXPECT_THAT(programNoBindless.getInternalOptions(), Not(HasSubstr(std::string("-cl-intel-use-bindless-images"))));
+    }
+
+    {
+        DebugManager.flags.UseBindlessBuffers.set(true);
+        MockProgram programNoBindless(*pDevice->getExecutionEnvironment(), pContext, false);
+        EXPECT_THAT(programNoBindless.getInternalOptions(), HasSubstr(std::string("-cl-intel-use-bindless-buffers ")));
+        EXPECT_THAT(programNoBindless.getInternalOptions(), Not(HasSubstr(std::string("-cl-intel-use-bindless-images"))));
+    }
+
+    {
+        DebugManager.flags.UseBindlessBuffers.set(false);
+        DebugManager.flags.UseBindlessImages.set(true);
+        MockProgram programNoBindless(*pDevice->getExecutionEnvironment(), pContext, false);
+        EXPECT_THAT(programNoBindless.getInternalOptions(), Not(HasSubstr(std::string("-cl-intel-use-bindless-buffers"))));
+        EXPECT_THAT(programNoBindless.getInternalOptions(), HasSubstr(std::string("-cl-intel-use-bindless-images ")));
+    }
+
+    {
+        DebugManager.flags.UseBindlessBuffers.set(true);
+        DebugManager.flags.UseBindlessImages.set(true);
+        MockProgram programNoBindless(*pDevice->getExecutionEnvironment(), pContext, false);
+        EXPECT_THAT(programNoBindless.getInternalOptions(), HasSubstr(std::string("-cl-intel-use-bindless-buffers ")));
+        EXPECT_THAT(programNoBindless.getInternalOptions(), HasSubstr(std::string("-cl-intel-use-bindless-images ")));
+    }
+}
+
 TEST_F(ProgramTests, givenDeviceThatSupportsSharedSystemMemoryAllocationWhenProgramIsCompiledThenItForcesStatelessCompilation) {
     pDevice->deviceInfo.sharedSystemMemCapabilities = CL_UNIFIED_SHARED_MEMORY_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_ATOMIC_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ACCESS_INTEL | CL_UNIFIED_SHARED_MEMORY_CONCURRENT_ATOMIC_ACCESS_INTEL;
     MockProgram program(*pDevice->getExecutionEnvironment(), pContext, false);

unit_tests/test_files/igdrcl.config

@@ -123,4 +123,6 @@ AllocateSharedAllocationsWithCpuAndGpuStorage = -1
 EnableSharedSystemUsmSupport = -1
 ForcePerDssBackedBufferProgramming = 0
 ForceSamplerLowFilteringPrecision = 0
+UseBindlessBuffers = 0
+UseBindlessImages = 0
 PrintProgramBinaryProcessingTime = 0