Fix lbann half issues (elemental#93)

* tick up version to 1.3.3

* fix issues with CPU half

* rework the copy interface

* work on Copy

* add a warmup run to the gemm test

* Various updates to copy dispatch; no more ETI for Copy

* fix an issue with the cuda half type's assignment operators

* make gpu_half_type assignment operators into templates

* Update include/El/blas_like/level1/CopyLocal.hpp

Co-Authored-By: Tim Moon <[email protected]>

* remove some debugging output

* add decls for BaseDistMatrix copy and copyasync

* make gpu_half_type streamable

* be a little more clever about casting to __half

* fix some things

* add overloads of sqrt and pow for half types

* add unary minus for gpu half type

* fix an issue where NVCC decided that rvalue references have value semantics

* add an overload of Log for gpu_half_type

* add exception-throwing bitwise operators for half. This is to appease Aluminum at compile-time and should never be encounted IRL.

* add a bunch of missing library symbols

* add a bunch of transendental functions, etc, for half types

* fix some missing symbols when compiling without half support

* Fixes to instantiate Read/Write with gpu half type

* Add overload for instantiate with half types

* add a few missing symbols

* add a write impl for gpu half matrices

* temp: dispatch gemv through Gemm for __half

* Fix the GEMV as GEMM call

* temporary error-throw for unhandled case

* remove the incy != 1 gemv case

* patch around an issue in the Half library

* expose AbstractMatrix interface to Print()

Co-authored-by: Tim Moon <[email protected]>

CMakeLists.txt

@@ -54,7 +54,7 @@ endif (__GIT_EXECUTABLE)
 # This must be set because version tags
 set(HYDROGEN_VERSION_MAJOR 1)
 set(HYDROGEN_VERSION_MINOR 3)
-set(HYDROGEN_VERSION_PATCH 2)
+set(HYDROGEN_VERSION_PATCH 3)
 set(HYDROGEN_VERSION_MAJOR_MINOR
     "${HYDROGEN_VERSION_MAJOR}.${HYDROGEN_VERSION_MINOR}")
 set(HYDROGEN_VERSION
@@ -539,6 +539,7 @@ if (HYDROGEN_HAVE_CUDA)
   target_link_libraries(Hydrogen_CUDA PUBLIC ${NVTX_LIBRARIES})
   target_link_libraries(Hydrogen_CUDA PUBLIC cuda::toolkit)
 
+  target_link_libraries(Hydrogen_CXX PUBLIC Hydrogen_CUDA)
   list(APPEND HYDROGEN_LIBRARIES Hydrogen_CUDA)
 endif ()
 

cmake/modules/HydrogenETIGeneration.cmake

@@ -0,0 +1,31 @@
+# This function sets the _ARG# variable and recurs through the parameters.
+function (h_gtpe_recur OUTVAR EXPRESSION_TEMPLATE THIS_EPARAM_ID THIS_EPARAM)
+  foreach (_VAL IN LISTS ${THIS_EPARAM})
+    set(_ARG${THIS_EPARAM_ID} "${_VAL}")
+    if (ARGN)
+      math(EXPR _NEXT_ID "${THIS_EPARAM_ID} + 1")
+      h_gtpe_recur(${OUTVAR} "${EXPRESSION_TEMPLATE}" ${_NEXT_ID} ${ARGN})
+    else ()
+      string(CONFIGURE "${EXPRESSION_TEMPLATE}" _THIS_EXPRESSION @ONLY)
+      list(APPEND ${OUTVAR} "${_THIS_EXPRESSION}")
+    endif ()
+  endforeach ()
+  set(${OUTVAR} "${${OUTVAR}}" PARENT_SCOPE)
+endfunction ()
+
+# This function calls the recursive function above to generate each
+# function instance.
+#
+# DO NOT ADD A SEMICOLON TO THE EXPRESSION_TEMPLATE PARAMETER!!!!!!!
+function (h_generate_tensor_product_expression OUTVAR EXPRESSION_TEMPLATE)
+  h_gtpe_recur(_ALL_EXPRESSIONS "${EXPRESSION_TEMPLATE}" 0 ${ARGN})
+  set(${OUTVAR} ${_ALL_EXPRESSIONS} PARENT_SCOPE)
+endfunction ()
+
+# This function adds a semicolon to each function instance in the list
+# and joins them into a string with each function instance on its own
+# line.
+function (h_func_list_to_string OUTVAR INLIST)
+  list(JOIN ${INLIST} ";\n" _TMP)
+  set(${OUTVAR} "${_TMP};" PARENT_SCOPE)
+endfunction ()