Implement missing host-side math functions (#884)

* add tests for host library, bit 128

* Implement missing host math funcs

* Compiling the bitode library to host code leads to linking errors
some of the OpenCL natives being used in the library code are not
available for the CPU, resulting in linking errors

* OCML does not meant to be compiled in C++ mode
for this, we create a drop-in opencl header defining the missing
datatypes

* Since OCML uses headers in the same directory as the source files,
we must copy the necessary source files into a new directory so that
our replacement headers get picked up instead of the originals

* Implement a new header for mapping math funcs to host-side ocml

* Discover an additional way to trigger 128 bit error

* Discover a discrepancy in host and device func results

* add missing rsqrt host funcs

* fix typo

Author: pvelesko

CMakeLists.txt

@@ -512,9 +512,9 @@ set(HIP_OFFLOAD_COMPILE_OPTIONS_BUILD_
 #             combined compile and link or when the --hip-link is
 #             present at the link step.
 list(APPEND HIP_OFFLOAD_LINK_OPTIONS_INSTALL_
-  "-L${LIB_INSTALL_DIR}" "-lCHIP" "-no-hip-rt")
+  "-L${LIB_INSTALL_DIR}" "-lCHIP" "-no-hip-rt -locml_host_math_funcs")
 list(APPEND HIP_OFFLOAD_LINK_OPTIONS_BUILD_
-  "-L${CMAKE_BINARY_DIR}" "-lCHIP" "-no-hip-rt")
+  "-L${CMAKE_BINARY_DIR}" "-lCHIP" "-no-hip-rt -locml_host_math_funcs")
 
 if(OpenCL_LIBRARY)
   target_link_options(CHIP PUBLIC -Wl,-rpath,${OpenCL_DIR})
@@ -807,4 +807,7 @@ if(LevelZero_LIBRARY)
 endif()
 
 # Include docker targets
-include(cmake/docker.cmake)
+include(cmake/docker.cmake)
+
+add_subdirectory(host_math_funcs)
+target_link_libraries(CHIP PUBLIC ocml_host_math_funcs)

host_math_funcs/CMakeLists.txt

@@ -0,0 +1,136 @@
+message(STATUS "Building host math functions")
+set(ROCm-Device-Libs_SRC_DIR ${CMAKE_SOURCE_DIR}/bitcode/ROCm-Device-Libs)
+set(OCML_SRC_DIR ${ROCm-Device-Libs_SRC_DIR}/ocml/src)
+set(OCML_INCLUDE_DIR ${ROCm-Device-Libs_SRC_DIR}/ocml/include)
+set(OCML_LIB_DIR ${ROCm-Device-Libs_SRC_DIR}/ocml/lib)
+
+# Add include directories for OCML
+include_directories(
+  ${OCML_INCLUDE_DIR}
+  ${ROCm-Device-Libs_SRC_DIR}/irif/inc
+  ${ROCm-Device-Libs_SRC_DIR}/oclc/inc
+  ${ROCm-Device-Libs_SRC_DIR}/ocml/inc
+  ${CMAKE_CURRENT_SOURCE_DIR}
+)
+
+
+# Create host_math_funcs directory in build
+file(MAKE_DIRECTORY ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# Copy ncdfF.cl to the build directory
+
+# cospi			OpenCL in chipStar (extern)
+# cospif			OpenCL in chipStar (calling cospi)
+file(COPY ${OCML_SRC_DIR}/trigredF.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/sincospiredF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/tanpiredF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/trigpiredF.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/trigpiredF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/cospiF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+file(COPY ${OCML_SRC_DIR}/trigpiredD.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/sincospiredD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/tanpiredD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/trigpiredD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/cospiD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+file(COPY ${OCML_SRC_DIR}/erfcD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/erfcF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+file(COPY ${OCML_SRC_DIR}/expD_base.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/expF_base.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/logD_base.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/logF_base.h DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/logD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/logF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+
+file(COPY ${OCML_SRC_DIR}/expD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/expF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# erfcinv			OCML in chipStar (__ocml_erfcinv_f64)
+file(COPY ${OCML_SRC_DIR}/erfcinvD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # erfcinvf		OCML in chipStar (__ocml_erfcinv_f32)
+file(COPY ${OCML_SRC_DIR}/erfcinvF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # erfcx			OCML in chipStar (__ocml_erfcx_f64)
+file(COPY ${OCML_SRC_DIR}/erfcxD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # erfcxf			OCML in chipStar (__ocml_erfcx_f32)
+file(COPY ${OCML_SRC_DIR}/erfcxF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # erfinv			OCML in chipStar (__ocml_erfinv_f64)
+file(COPY ${OCML_SRC_DIR}/erfinvD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # erfinvf			OCML in chipStar (__ocml_erfinv_f32)
+file(COPY ${OCML_SRC_DIR}/erfinvF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # normcdf			OCML in chipStar (__ocml_ncdf_f64)
+file(COPY ${OCML_SRC_DIR}/ncdfD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # normcdff		OCML in chipStar (__ocml_ncdf_f32)
+file(COPY ${OCML_SRC_DIR}/ncdfF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # normcdfinv		OCML in chipStar (__ocml_ncdfinv_f64)
+file(COPY ${OCML_SRC_DIR}/ncdfinvD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # normcdfinvf		OCML in chipStar (__ocml_ncdfinv_f32)
+file(COPY ${OCML_SRC_DIR}/ncdfinvF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # rcbrt			OCML in chipStar (__ocml_rcbrt_f64)
+file(COPY ${OCML_SRC_DIR}/rcbrtD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # rcbrtf			OCML in chipStar (__ocml_rcbrt_f32)
+file(COPY ${OCML_SRC_DIR}/rcbrtF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # sincospi		OCML in chipStar (__ocml_sincospi_f64)
+file(COPY ${OCML_SRC_DIR}/sincospiD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # sincospif		OCML in chipStar (__ocml_sincospi_f32)
+file(COPY ${OCML_SRC_DIR}/sincospiF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # sinpi			OpenCL in chipStar (extern)
+file(COPY ${OCML_SRC_DIR}/sinpiD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# # sinpif			OpenCL in chipStar (call sinpi)
+file(COPY ${OCML_SRC_DIR}/sinpiF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# rsqrt			OCML in chipStar (__ocml_rsqrt_f64)
+file(COPY ${OCML_SRC_DIR}/rsqrtD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# rsqrtf			OCML in chipStar (__ocml_rsqrt_f32)
+file(COPY ${OCML_SRC_DIR}/rsqrtF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+# llmax			calling max() in chipStar
+# llmin			calling min() in chipStar
+# ullmax			calling max() in chipStar
+# ullmin			calling min() in chipStar
+# umax			calling max() in chipStar
+# umin			calling min() in chipStar
+
+# signbit			OCML in chipStar (__ocml_signbit_f64)
+file(COPY ${OCML_SRC_DIR}/signbitD.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+file(COPY ${OCML_SRC_DIR}/signbitF.cl DESTINATION ${CMAKE_BINARY_DIR}/host_math_funcs)
+
+
+# Gather all .cl sources from the OCML source directory
+file(GLOB OCML_CL_SOURCES ${CMAKE_BINARY_DIR}/host_math_funcs/*.cl)
+
+# Print the list of gathered .cl sources for debugging
+message(STATUS "OCML CL Sources: ${OCML_CL_SOURCES}")
+
+# Set the language for the target
+set_source_files_properties(${OCML_CL_SOURCES} PROPERTIES LANGUAGE C)
+
+# Add the library target
+add_library(ocml_host_math_funcs STATIC ${OCML_CL_SOURCES})
+set_target_properties(ocml_host_math_funcs PROPERTIES
+    ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}"
+)
+# Set the linker language explicitly
+set_target_properties(ocml_host_math_funcs PROPERTIES LINKER_LANGUAGE C)
+
+# Add ocml_host_math_funcs to the export set
+install(TARGETS ocml_host_math_funcs EXPORT hip-targets)
+install(TARGETS ocml_host_math_funcs EXPORT CHIPTargets)