add an environment variable control for the suggested local work size

Author: bashbaug

layers/10_cmdbufemu/README.md

@@ -34,6 +34,7 @@ The following environment variables can modify the behavior of the command buffe
 |----------------------|----------|-----------------|
 | `CMDBUFEMU_EnhancedErrorChecking` | Enables additional error checking when commands are added to a command buffer using a command buffer "test queue".  By default, the additional error checking is disabled. | `export CMDBUFEMU_EnhancedErrorChecking=1`<br/><br/>`set CMDBUFEMU_EnhancedErrorChecking=1` |
 | `CMDBUFEMU_KernelForProfiling` | Enables use of an empty kernel for event profiling instead of event profiling on a command-queue barrier.  By default, to minimize overhead, the empty kernel is not used. | `export CMDBUFEMU_KernelForProfiling=1`<br/><br/>`set CMDBUFEMU_KernelForProfiling=1` |
+| `CMDBUFEMU_SuggestedLocalWorkSize` | Enables use of the suggested local work-group size extension to eliminate `NULL` local work-group sizes.  Only valid when an implementation supports the local work-group size extension and the command is not mutable.  By default, use of the suggested local work-group size is enabled. | `export CMDBUFEMU_SuggestedLocalWorkSize=0`<br/><br/>`set CMDBUFEMU_SuggestedLocalWorkSize=0` |
 
 ## Known Limitations
 

layers/10_cmdbufemu/emulate.cpp

@@ -2023,7 +2023,7 @@ std::unique_ptr<NDRangeKernel> NDRangeKernel::create(
             local_work_size,
             local_work_size + work_dim);
     }
-    else if( isMutable == false )
+    else if( g_SuggestedLocalWorkSize && isMutable == false )
     {
         command->local_work_size.resize(work_dim);
         cl_int checkError = cmdbuf->clGetKernelSuggestedLocalWorkSize(

layers/10_cmdbufemu/emulate.h

@@ -11,6 +11,7 @@
 
 extern bool g_EnhancedErrorChecking;
 extern bool g_KernelForProfiling;
+extern bool g_SuggestedLocalWorkSize;
 
 extern const struct _cl_icd_dispatch* g_pNextDispatch;
 

layers/10_cmdbufemu/main.cpp

@@ -35,10 +35,16 @@
 bool g_EnhancedErrorChecking = false;
 
 // Using kernels for profiling can fix issues with some implementations
-// that do not properly support event profiling on barrkers.
+// that do not properly support event profiling on barriers.
 
 bool g_KernelForProfiling = false;
 
+// Using the suggested local work-group size can reduce overhead by determining
+// the values for a NULL local work-group size when the command buffer is
+// created rather than when it is executed.
+
+bool g_SuggestedLocalWorkSize = true;
+
 const struct _cl_icd_dispatch* g_pNextDispatch = NULL;
 
 static cl_int CL_API_CALL
@@ -231,7 +237,7 @@ static void _init_dispatch()
 }
 
 CL_API_ENTRY cl_int CL_API_CALL clGetLayerInfo(
-    cl_layer_info  param_name,
+    cl_layer_info param_name,
     size_t param_value_size,
     void* param_value,
     size_t* param_value_size_ret)
@@ -251,10 +257,17 @@ CL_API_ENTRY cl_int CL_API_CALL clGetLayerInfo(
 #if defined(CL_LAYER_NAME)
     case CL_LAYER_NAME:
         {
+            char str[256];
+            snprintf(str, 256, "Emulation Layer for "
+                CL_KHR_COMMAND_BUFFER_EXTENSION_NAME
+                " (EEC: %s, KFP: %s, SLWS: %s)",
+                g_EnhancedErrorChecking ? "Y" : "N",
+                g_KernelForProfiling ? "Y" : "N",
+                g_SuggestedLocalWorkSize ? "Y" : "N");
             auto ptr = (char*)param_value;
             return writeStringToMemory(
                 param_value_size,
-                "Emulation Layer for " CL_KHR_COMMAND_BUFFER_EXTENSION_NAME,
+                str,
                 param_value_size_ret,
                 ptr);
         }
@@ -290,6 +303,7 @@ CL_API_ENTRY cl_int CL_API_CALL clInitLayerWithProperties(
 
     getControl("CMDBUFEMU_EnhancedErrorChecking", g_EnhancedErrorChecking);
     getControl("CMDBUFEMU_KernelForProfiling", g_KernelForProfiling);
+    getControl("CMDBUFEMU_SuggestedLocalWorkSize", g_SuggestedLocalWorkSize);
 
     g_pNextDispatch = target_dispatch;