fix GPU + only generate files instantiations

daphne-eu · Sep 26, 2024 · a629d8d · a629d8d
1 parent 16f933c
commit a629d8d
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 28 deletions.
diff --git a/src/runtime/local/kernels/CMakeLists.txt b/src/runtime/local/kernels/CMakeLists.txt
@@ -24,7 +24,7 @@ if(USE_CUDA AND CMAKE_CUDA_COMPILER)
     add_custom_command(
             OUTPUT ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels.cpp ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json
             COMMAND python3 ARGS genKernelInst.py kernels.json
-                    ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels.cpp ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json CUDA
+                    ${PROJECT_BINARY_DIR}/src/runtime/local/kernels/CUDAkernels ${PROJECT_SOURCE_DIR}/lib/CUDAcatalog.json CUDA
             MAIN_DEPENDENCY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/kernels.json
             DEPENDS ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/genKernelInst.py
             WORKING_DIRECTORY ${PROJECT_SOURCE_DIR}/src/runtime/local/kernels/

diff --git a/src/runtime/local/kernels/genKernelInst.py b/src/runtime/local/kernels/genKernelInst.py
@@ -280,10 +280,10 @@ def getDefaultHeaders() -> str:
         "#include <runtime/local/context/DaphneContext.h>\n" \
         "#include <runtime/local/instrumentation/KernelInstrumentation.h>\n"
 
-def codegenKernelInfos(kernelInfos, catalog_entries) -> Tuple[str,
-                                                               List[str]]:
+def codegenKernelInfos(kernelInfos, catalog_entries) -> Tuple[bool, str]:
     ops_inst_str: str = ""
     header_str: str = getDefaultHeaders()
+    didGenerateCode = False
 
     for kernelInfo in kernelInfos:
         kernelTemplateInfo = kernelInfo["kernelTemplate"]
@@ -312,6 +312,7 @@ def codegenKernelInfos(kernelInfos, catalog_entries) -> Tuple[str,
                             generateKernelInstantiation(kernelTemplateInfo, instantiation,
                                                         api.get("opCodes", None), outBuf, catalog_entries, API)
                         ops_inst_str += outBuf.getvalue()
+                        didGenerateCode = True
         else:
             if API == "CPP":
                 # Comment reporting the kernel name.
@@ -327,9 +328,10 @@ def codegenKernelInfos(kernelInfos, catalog_entries) -> Tuple[str,
                 for instantiation in kernelInfo["instantiations"]:
                     generateKernelInstantiation(kernelTemplateInfo, instantiation, opCodes, outBuf, catalog_entries, API)
                 ops_inst_str += outBuf.getvalue()
+                didGenerateCode = True
 
     file = header_str + '\nextern \"C\" {\n' + ops_inst_str + "}\n"
-    return file
+    return didGenerateCode, file
 
 
 if __name__ == "__main__":
@@ -355,30 +357,10 @@ def codegenKernelInfos(kernelInfos, catalog_entries) -> Tuple[str,
     with open(inSpecPath, "r") as inFile:
         kernelInfos = json.load(inFile)
 
-        def getTImplCount(kernelInfo):
-            if "api" in kernelInfo:
-                sum = 0
-                for api in kernelInfo["api"]:
-                    if api == API:
-                        sum += len(api['instantiations'])
-                return sum
-            else:
-                return len(kernelInfo['instantiations'])
-        kernelInfos.sort(key=getTImplCount, reverse=True)
-
         for kernelInfo in kernelInfos:
-            file = codegenKernelInfos([kernelInfo], catalog_entries)
-            cppCodegen.append(file)
-        #  topN = 75
-        #  for i in range(0, topN):
-        #      file = codegenKernelInfos([kernelInfos[i]],
-        #                                          catalog_entries)
-        #      cppCodegen.append(file)
-        #
-        #  # Generate a single .cpp file for all remaining kernels
-        #  file = codegenKernelInfos(kernelInfos[topN:],
-        #                                      catalog_entries)
-        #  cppCodegen.append(file)
+            didGenerateCode, file = codegenKernelInfos([kernelInfo], catalog_entries)
+            if didGenerateCode:
+                cppCodegen.append(file)
 
     cppFiles = []
     # Write the generated C++ code to separate files.
@@ -393,4 +375,4 @@ def getTImplCount(kernelInfo):
         outCatalog.write(json.dumps(catalog_entries, indent=2))
 
     #  Lists all generated *.cpp files.
-    print(cppFiles)
+    #  print(cppFiles)