search

IBM / aihwkit

IBM Analog Hardware Acceleration Kit
https://aihwkit.readthedocs.io
Apache License 2.0
349 stars 145 forks source link

feat(docs): add half-precision training section in using_simulator docs #678

Open PabloCarmona opened 1 month ago

PabloCarmona commented 1 month ago

Related issues

https://github.com/IBM/aihwkit/issues/623 https://github.com/IBM/aihwkit/issues/677

Description

PabloCarmona commented 1 month ago

@coreylammie @anu-pub Take a look at it and let me know what you think, thanks!

jubueche commented 1 month ago

Compilation is currently not working for the half precision type. To reproduce: Torch version and CUDA: 2.3.0+cu121 AIHWKIT: Current master (6bcc9d6cd9e6210f113a01fa49b16dad7e09761d). Compilation script:

export COMPILER=/gpfs/u/software/npl/gcc/9.3.0/1/
export CXX=$COMPILER/bin/g++
export CC=$COMPILER/bin/gcc
export MKLROOT=$CONDA_PREFIX
export LD_LIBRARY_PATH=$COMPILER/lib64:$COMPILER/lib:${LD_LIBRARY_PATH}
export CUDA_VERSION=12.1
export CUDA_HOME=/usr/local/cuda-${CUDA_VERSION}
export CUDA_TOOLKIT_ROOT_DIR=${CUDA_HOME}
export CUDA_LIB_PATH=${CUDA_HOME}/lib
export CUDA_INCLUDE_DIRS=${CUDA_HOME}/include
export RPU_CUDA_ARCHITECTURS="70;80"
export PATH=${CUDA_HOME}/bin:$COMPILER/bin:${PATH}
export CMAKE_ROOT=/usr/lib64/cmake/
export CMAKE_PREFIX_PATH=$CONDA_PREFIX
cd ~/scratch/public_aihwkit
make build_inplace_cuda flags="-DRPU_CUDA_ARCHITECTURES='70;80' -DRPU_USE_FP16=ON"

Error:

[1/80] Building CXX object CMakeFiles/RPU_CPU.dir/src/rpucuda/dense_bit_line_maker.cpp.o
FAILED: CMakeFiles/RPU_CPU.dir/src/rpucuda/dense_bit_line_maker.cpp.o 
/gpfs/u/software/npl/gcc/9.3.0/1/bin/g++ -DRPU_TORCH_CUDA_BUFFERS -DRPU_USE_CUDA -DRPU_USE_FP16 -DRPU_USE_MKL -DRPU_USE_WITH_TORCH -D_GLIBCXX_USE_CXX11_ABI=0 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include/python3.10 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/pybind11/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -isystem /gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include -isystem /usr/local/cuda-12.1/targets/x86_64-linux/include -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/cuda -fopenmp -Wall -Wno-narrowing -Wno-strict-overflow -g -std=gnu++17 -fPIC -MD -MT CMakeFiles/RPU_CPU.dir/src/rpucuda/dense_bit_line_maker.cpp.o -MF CMakeFiles/RPU_CPU.dir/src/rpucuda/dense_bit_line_maker.cpp.o.d -o CMakeFiles/RPU_CPU.dir/src/rpucuda/dense_bit_line_maker.cpp.o -c /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp: In instantiation of ‘int* RPU::DenseBitLineMaker<T>::makeCoincidences(const T*, int, int&, const T*, int, int&, RPU::RNG<T>*, T, T, const RPU::PulsedUpdateMetaParameter<T>&) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:365:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:286:5: error: conversion from ‘int’ to ‘__half’ is ambiguous
  286 |   T A = 0;
      |     ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:287:5: error: conversion from ‘int’ to ‘__half’ is ambiguous
  287 |   T B = 0;
      |     ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp: In instantiation of ‘void RPU::DenseBitLineMaker<T>::generateCoincidences(int*, const int*, int, const int*, int, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:365:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:192:10: error: call of overloaded ‘__half(const int&)’ is ambiguous
  192 |   T bl = (T)BL;
      |          ^~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:195:12: error: call of overloaded ‘__half(const int&)’ is ambiguous
  195 |     T dc = (T)d_counts[i];
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:197:10: error: no match for ‘operator/=’ (operand types are ‘__half’ and ‘__half’)
  197 |       dc /= bl;
      |       ~~~^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:200:55: error: call of overloaded ‘__half(const int&)’ is ambiguous
  200 |         coincidences[idx++] = (int)RPU_ROUNDFUNF(dc * (T)x_counts[j]);
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp: In instantiation of ‘void RPU::DenseBitLineMaker<T>::generateCoincidencesDetI(int*, const T*, int, const T*, int, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:365:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:257:10: error: no match for ‘operator*=’ (operand types are ‘__half’ and ‘const int’)
  257 |       dc *= BL;
      |       ~~~^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp: In instantiation of ‘void RPU::DenseBitLineMaker<T>::generateCountsMean(int*, const T*, int, int, int&, T, RPU::RNG<T>*, int, T, bool, T) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:365:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:165:29: error: operands to ?: have different types ‘float’ and ‘const __half’
  165 |     T v_value = lr < (T)0.0 ? -v[j_v] : v[j_v];
      |                 ~~~~~~~~~~~~^~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:165:29: note:   and each type can be converted to the other
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:168:31: error: no matching function for call to ‘getDiscretizedValue(float, const __half&, const bool&, RPU::RNG<__half>&)’
  168 |     T pp = getDiscretizedValue((T)fabsf((T)v_value) * p, res, sto_round, *rng);
      |            ~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:178:10: note: candidate: ‘template<class T, class RNGClass> T RPU::getDiscretizedValue(T, T, bool, RNGClass&)’
  178 | inline T getDiscretizedValue(T value, T res, bool sto_round, RNGClass &rng) {
      |          ^~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:178:10: note:   template argument deduction/substitution failed:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:168:31: note:   deduced conflicting types for parameter ‘T’ (‘float’ and ‘__half’)
  168 |     T pp = getDiscretizedValue((T)fabsf((T)v_value) * p, res, sto_round, *rng);
      |            ~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:45: error: call of overloaded ‘__half(const int&)’ is ambiguous
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                                             ^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:22: note: in definition of macro ‘MAX’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:22: note: in expansion of macro ‘MIN’
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                      ^~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:45: error: call of overloaded ‘__half(const int&)’ is ambiguous
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                                             ^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:22: note: in definition of macro ‘MAX’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:22: note: in expansion of macro ‘MIN’
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                      ^~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:45: error: call of overloaded ‘__half(const int&)’ is ambiguous
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                                             ^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:35: note: in definition of macro ‘MAX’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:22: note: in expansion of macro ‘MIN’
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                      ^~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:45: error: call of overloaded ‘__half(const int&)’ is ambiguous
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                                             ^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:35: note: in definition of macro ‘MAX’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:176:22: note: in expansion of macro ‘MIN’
  176 |     int ntimes = MAX(MIN((int)RPU_ROUNDFUNF((T)BL * pp), BL), 0);
      |                      ^~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp: In instantiation of ‘void RPU::DenseBitLineMaker<T>::generateDetImplicit(T*, const T*, int, int, int&, T, RPU::RNG<T>*, int, T, bool, T) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:365:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:231:29: error: operands to ?: have different types ‘float’ and ‘const __half’
  231 |     T v_value = lr < (T)0.0 ? -v[j_v] : v[j_v];
      |                 ~~~~~~~~~~~~^~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:231:29: note:   and each type can be converted to the other
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:241:38: error: operands to ?: have different types ‘__half’ and ‘float’
  241 |     pcounts[j] = (v_value >= (T)0.0) ? pp : -pp;
      |                  ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/dense_bit_line_maker.cpp:241:38: note:   and each type can be converted to the other
[2/80] Building CXX object CMakeFiles/RPU_CPU.dir/src/rpucuda/math_util.cpp.o
FAILED: CMakeFiles/RPU_CPU.dir/src/rpucuda/math_util.cpp.o 
/gpfs/u/software/npl/gcc/9.3.0/1/bin/g++ -DRPU_TORCH_CUDA_BUFFERS -DRPU_USE_CUDA -DRPU_USE_FP16 -DRPU_USE_MKL -DRPU_USE_WITH_TORCH -D_GLIBCXX_USE_CXX11_ABI=0 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include/python3.10 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/pybind11/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -isystem /gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include -isystem /usr/local/cuda-12.1/targets/x86_64-linux/include -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/cuda -fopenmp -Wall -Wno-narrowing -Wno-strict-overflow -g -std=gnu++17 -fPIC -MD -MT CMakeFiles/RPU_CPU.dir/src/rpucuda/math_util.cpp.o -MF CMakeFiles/RPU_CPU.dir/src/rpucuda/math_util.cpp.o.d -o CMakeFiles/RPU_CPU.dir/src/rpucuda/math_util.cpp.o -c /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp: In function ‘void RPU::math::scal(int, T, T*, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp:171:12: error: no match for ‘operator*=’ (operand types are ‘half_t’ {aka ‘__half’} and ‘const half_t’ {aka ‘const __half’})
  171 |     X[i_x] *= alpha;
      |     ~~~~~~~^~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp: In function ‘T RPU::math::nrm2(int, const T*, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp:192:15: error: no match for ‘operator+=’ (operand types are ‘half_t’ {aka ‘__half’} and ‘float’)
  192 |     nrm_value += x * x;
      |     ~~~~~~~~~~^~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp: In instantiation of ‘void RPU::math::makeBias(T*, const T*, int, int, bool) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp:396:90:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp:375:24: error: call of overloaded ‘__half(int)’ is ambiguous
  375 |       x_with_bias[i] = (T)1;
      |                        ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
[3/80] Building CXX object CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_dynamic_transfer_device.cpp.o
FAILED: CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_dynamic_transfer_device.cpp.o 
/gpfs/u/software/npl/gcc/9.3.0/1/bin/g++ -DRPU_TORCH_CUDA_BUFFERS -DRPU_USE_CUDA -DRPU_USE_FP16 -DRPU_USE_MKL -DRPU_USE_WITH_TORCH -D_GLIBCXX_USE_CXX11_ABI=0 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include/python3.10 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/pybind11/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -isystem /gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include -isystem /usr/local/cuda-12.1/targets/x86_64-linux/include -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/cuda -fopenmp -Wall -Wno-narrowing -Wno-strict-overflow -g -std=gnu++17 -fPIC -MD -MT CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_dynamic_transfer_device.cpp.o -MF CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_dynamic_transfer_device.cpp.o.d -o CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_dynamic_transfer_device.cpp.o -c /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h: In instantiation of ‘void RPU::DynamicTransferRPUDeviceMetaParameter<T>::initDefaults() [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:67:46: error: ambiguous overload for ‘operator=’ (operand types are ‘__half’ and ‘bool’)
   67 |   void initDefaults() { this->in_chop_random = false; };
      |                         ~~~~~~~~~~~~~~~~~~~~~^~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:218:33: note: candidate: ‘__half& __half::operator=(float)’
  218 |     __CUDA_HOSTDEVICE__ __half &operator=(const float f) { __x = __float2half(f).__x; return *this; }
      |                                 ^~~~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:221:33: note: candidate: ‘__half& __half::operator=(double)’
  221 |     __CUDA_HOSTDEVICE__ __half &operator=(const double f) { __x = __double2half(f).__x; return *this; }
      |                                 ^~~~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half& __half::operator=(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half& __half::operator=(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp: In instantiation of ‘void RPU::DynamicTransferRPUDevice<T>::readAndUpdate(int, int, T, const T*, int, T, int, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:515:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:257:59: error: call of overloaded ‘__half(const unsigned int&)’ is ambiguous
  257 |   const T sample_momentum = (T)MIN(par.tail_weightening / (T)n_samples, (T)1.0);
      |                                                           ^~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:257:59: error: call of overloaded ‘__half(const unsigned int&)’ is ambiguous
  257 |   const T sample_momentum = (T)MIN(par.tail_weightening / (T)n_samples, (T)1.0);
      |                                                           ^~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:268:21: error: call of overloaded ‘__half(const unsigned int&)’ is ambiguous
  268 |     lr_scale /= MIN((T)current_sample, (T)1.0 / from_weight_granularity);
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:268:21: error: call of overloaded ‘__half(const unsigned int&)’ is ambiguous
  268 |     lr_scale /= MIN((T)current_sample, (T)1.0 / from_weight_granularity);
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:272:18: error: call of overloaded ‘__half(const int&)’ is ambiguous
  272 |   T buffer_cap = (T)max_steps * par.buffer_cap;
      |                  ^~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:304:45: error: operands to ?: have different types ‘float’ and ‘__half’
  304 |     dw = (in_chop != this->out_chopper_[j]) ? -dw : dw;
      |          ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:304:45: note:   and each type can be converted to the other
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:305:11: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘float’)
  305 |     omega += dw * lr_scale;
      |     ~~~~~~^~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:311:15: error: no match for ‘operator*=’ (operand types are ‘__half’ and ‘const __half’)
  311 |         omega *= momentum;
      |         ~~~~~~^~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:313:15: error: no match for ‘operator-=’ (operand types are ‘__half’ and ‘float’)
  313 |         omega -= ((T)1.0 - momentum) * n_steps;
      |         ~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘__half’ and ‘float’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:315:23: note: in expansion of macro ‘MAX’
  315 |           omega = MIN(MAX(omega, -buffer_cap), buffer_cap);
      |                       ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:315:23: note: in expansion of macro ‘MAX’
  315 |           omega = MIN(MAX(omega, -buffer_cap), buffer_cap);
      |                       ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘__half’ and ‘float’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:315:23: note: in expansion of macro ‘MAX’
  315 |           omega = MIN(MAX(omega, -buffer_cap), buffer_cap);
      |                       ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:315:23: note: in expansion of macro ‘MAX’
  315 |           omega = MIN(MAX(omega, -buffer_cap), buffer_cap);
      |                       ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp: In instantiation of ‘T RPU::DynamicTransferRPUDevice<T>::getPulseCountLearningRate(T, int, const RPU::PulsedUpdateMetaParameter<T>&) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:515:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:362:12: error: no match for ‘operator*=’ (operand types are ‘__half’ and ‘__half’)
  362 |   count_lr *= count_lr_scale_;
      |   ~~~~~~~~~^~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp: In instantiation of ‘void RPU::DynamicTransferRPUDevice<T>::populate(const RPU::DynamicTransferRPUDeviceMetaParameter<T>&, RPU::RealWorldRNG<T>*) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:515:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:181:3: error: conversion from ‘int’ to ‘const value_type’ {aka ‘const __half’} is ambiguous
  181 |   running_mean_.resize((size_t)this->size_, 0);
      |   ^~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
In file included from /gpfs/u/software/npl/gcc/9.3.0/1/include/c++/9.3.0/vector:67,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:20,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/software/npl/gcc/9.3.0/1/include/c++/9.3.0/bits/stl_vector.h:954:54: note:   initializing argument 2 of ‘void std::vector<_Tp, _Alloc>::resize(std::vector<_Tp, _Alloc>::size_type, const value_type&) [with _Tp = __half; _Alloc = std::allocator<__half>; std::vector<_Tp, _Alloc>::size_type = long unsigned int; std::vector<_Tp, _Alloc>::value_type = __half]’
  954 |       resize(size_type __new_size, const value_type& __x)
      |                                    ~~~~~~~~~~~~~~~~~~^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:182:3: error: conversion from ‘int’ to ‘const value_type’ {aka ‘const __half’} is ambiguous
  182 |   past_mean_.resize((size_t)this->size_, 0);
      |   ^~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
In file included from /gpfs/u/software/npl/gcc/9.3.0/1/include/c++/9.3.0/vector:67,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:20,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/software/npl/gcc/9.3.0/1/include/c++/9.3.0/bits/stl_vector.h:954:54: note:   initializing argument 2 of ‘void std::vector<_Tp, _Alloc>::resize(std::vector<_Tp, _Alloc>::size_type, const value_type&) [with _Tp = __half; _Alloc = std::allocator<__half>; std::vector<_Tp, _Alloc>::size_type = long unsigned int; std::vector<_Tp, _Alloc>::value_type = __half]’
  954 |       resize(size_type __new_size, const value_type& __x)
      |                                    ~~~~~~~~~~~~~~~~~~^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:187:14: error: call of overloaded ‘__half(int)’ is ambiguous
  187 |   T period = (T)par.getInSize() * (T)fabsf(par.transfer_every) * (T)par.getNumInChopSamples();
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:187:66: error: call of overloaded ‘__half(unsigned int)’ is ambiguous
  187 |   T period = (T)par.getInSize() * (T)fabsf(par.transfer_every) * (T)par.getNumInChopSamples();
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘float’ and ‘__half’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:193:11: note: in expansion of macro ‘MAX’
  193 |       MIN(MAX((w_max - weight_granularity) * par.experimental_feedback_target, weight_granularity),
      |           ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:193:11: note: in expansion of macro ‘MAX’
  193 |       MIN(MAX((w_max - weight_granularity) * par.experimental_feedback_target, weight_granularity),
      |           ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘float’ and ‘__half’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:193:11: note: in expansion of macro ‘MAX’
  193 |       MIN(MAX((w_max - weight_granularity) * par.experimental_feedback_target, weight_granularity),
      |           ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:193:11: note: in expansion of macro ‘MAX’
  193 |       MIN(MAX((w_max - weight_granularity) * par.experimental_feedback_target, weight_granularity),
      |           ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:196:30: error: call of overloaded ‘__half(const int&)’ is ambiguous
  196 |   v[FEEDBACK_MOD] = period * (T)par.experimental_feedback_mod;
      |                              ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h: In instantiation of ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter() [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:43:86:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:31:5: error: conversion from ‘bool’ to ‘__half’ is ambiguous
   31 |   T in_chop_random = true;  // random or regular
      |     ^~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h: In instantiation of ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:48:63:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:31:5: error: conversion from ‘bool’ to ‘__half’ is ambiguous
   31 |   T in_chop_random = true;  // random or regular
      |     ^~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h: In instantiation of ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:55:83:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:31:5: error: conversion from ‘bool’ to ‘__half’ is ambiguous
   31 |   T in_chop_random = true;  // random or regular
      |     ^~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘RPU::IOMetaParameter<__half>::IOMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:82:35:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:46:80:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:45:87:   required from ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:43:86:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:86:16: error: call of overloaded ‘__half(int)’ is ambiguous
   86 |   T _out_res = (T)0;
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘constexpr RPU::PulsedUpdateMetaParameter<__half>::PulsedUpdateMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:82:35:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:46:80:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:45:87:   required from ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:43:86:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:267:11: error: call of overloaded ‘__half(int)’ is ambiguous
  267 |   T res = (T)0; // this is taken to be in the range 0..1 as positive and negative phases are done
      |           ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:270:22: error: call of overloaded ‘__half(int)’ is ambiguous
  270 |   T x_res_implicit = (T)0; // in case of implicit pulsing. Assumes range 0..1
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:271:22: error: call of overloaded ‘__half(int)’ is ambiguous
  271 |   T d_res_implicit = (T)0;
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘RPU::IOMetaParameter<__half>::IOMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:84:54:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:51:56:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:50:64:   required from ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:48:63:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:86:16: error: call of overloaded ‘__half(int)’ is ambiguous
   86 |   T _out_res = (T)0;
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘constexpr RPU::PulsedUpdateMetaParameter<__half>::PulsedUpdateMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:84:54:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:51:56:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:50:64:   required from ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:48:63:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:267:11: error: call of overloaded ‘__half(int)’ is ambiguous
  267 |   T res = (T)0; // this is taken to be in the range 0..1 as positive and negative phases are done
      |           ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:270:22: error: call of overloaded ‘__half(int)’ is ambiguous
  270 |   T x_res_implicit = (T)0; // in case of implicit pulsing. Assumes range 0..1
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:271:22: error: call of overloaded ‘__half(int)’ is ambiguous
  271 |   T d_res_implicit = (T)0;
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_device.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h: In instantiation of ‘constexpr RPU::DriftParameter<__half>::DriftParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:171:25:   required from ‘RPU::SimpleMetaParameter<T>::SimpleMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_simple_device.h:56:30:   required from ‘RPU::PulsedRPUDeviceMetaParameterBase<T>::PulsedRPUDeviceMetaParameterBase() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h:34:33:   required from ‘RPU::VectorRPUDeviceMetaParameter<T>::VectorRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:82:35:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:46:80:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:45:87:   required from ‘RPU::ChoppedTransferRPUDeviceMetaParameter<T>::ChoppedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:43:86:   required from ‘RPU::DynamicTransferRPUDeviceMetaParameter<T>::DynamicTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:103:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:33:16: error: call of overloaded ‘__half(int)’ is ambiguous
   33 |   T wg_ratio = (T)1; // (wmax-wmin)/(gmax-gmin)
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_device.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:34:16: error: call of overloaded ‘__half(int)’ is ambiguous
   34 |   T g_offset = (T)0; // gmin
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_device.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:35:16: error: call of overloaded ‘__half(int)’ is ambiguous
   35 |   T w_offset = (T)0; // w(gmin), i.e. to what value gmin is mapped to in w-space
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:18,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h: In instantiation of ‘T RPU::VectorRPUDeviceMetaParameter<T>::calcWeightGranularity() const [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h:93:5:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h:96:67: error: call of overloaded ‘__half(std::vector<std::unique_ptr<RPU::PulsedRPUDeviceMetaParameterBase<__half>, std::default_delete<RPU::PulsedRPUDeviceMetaParameterBase<__half> > >, std::allocator<std::unique_ptr<RPU::PulsedRPUDeviceMetaParameterBase<__half>, std::default_delete<RPU::PulsedRPUDeviceMetaParameterBase<__half> > > > >::size_type)’ is ambiguous
   96 |       weight_granularity += vec_par[k]->calcWeightGranularity() / (T)vec_par.size();
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:18,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h: In instantiation of ‘T RPU::VectorRPUDeviceMetaParameter<T>::calcNumStates() const [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h:100:5:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_vector_device.h:103:51: error: call of overloaded ‘__half(std::vector<std::unique_ptr<RPU::PulsedRPUDeviceMetaParameterBase<__half>, std::default_delete<RPU::PulsedRPUDeviceMetaParameterBase<__half> > >, std::allocator<std::unique_ptr<RPU::PulsedRPUDeviceMetaParameterBase<__half>, std::default_delete<RPU::PulsedRPUDeviceMetaParameterBase<__half> > > > >::size_type)’ is ambiguous
  103 |       num_states += vec_par[k]->calcNumStates() / (T)vec_par.size();
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_chopped_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_dynamic_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
[4/80] Building CXX object CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu.cpp.o
FAILED: CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu.cpp.o 
/gpfs/u/software/npl/gcc/9.3.0/1/bin/g++ -DRPU_TORCH_CUDA_BUFFERS -DRPU_USE_CUDA -DRPU_USE_FP16 -DRPU_USE_MKL -DRPU_USE_WITH_TORCH -D_GLIBCXX_USE_CXX11_ABI=0 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include/python3.10 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/pybind11/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -isystem /gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include -isystem /usr/local/cuda-12.1/targets/x86_64-linux/include -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/cuda -fopenmp -Wall -Wno-narrowing -Wno-strict-overflow -g -std=gnu++17 -fPIC -MD -MT CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu.cpp.o -MF CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu.cpp.o.d -o CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu.cpp.o -c /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h: In instantiation of ‘constexpr RPU::DriftParameter<__half>::DriftParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:171:25:   required from ‘RPU::SimpleMetaParameter<T>::SimpleMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:217:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:33:16: error: call of overloaded ‘__half(int)’ is ambiguous
   33 |   T wg_ratio = (T)1; // (wmax-wmin)/(gmax-gmin)
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:34:16: error: call of overloaded ‘__half(int)’ is ambiguous
   34 |   T g_offset = (T)0; // gmin
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:35:16: error: call of overloaded ‘__half(int)’ is ambiguous
   35 |   T w_offset = (T)0; // w(gmin), i.e. to what value gmin is mapped to in w-space
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h: In instantiation of ‘void RPU::DriftParameter<T>::printToStream(std::stringstream&) const [with T = __half; std::stringstream = std::basic_stringstream<char>]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:219:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:80:21: error: call of overloaded ‘__half(int)’ is ambiguous
   80 |     if (wg_ratio != (T)1 || g_offset != (T)0 || w_offset != (T)0) {
      |                     ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:80:41: error: call of overloaded ‘__half(int)’ is ambiguous
   80 |     if (wg_ratio != (T)1 || g_offset != (T)0 || w_offset != (T)0) {
      |                                         ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:17,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/weight_drifter.h:80:61: error: call of overloaded ‘__half(int)’ is ambiguous
   80 |     if (wg_ratio != (T)1 || g_offset != (T)0 || w_offset != (T)0) {
      |                                                             ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::applyWeightUpdate(T*) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1447:10: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘__half’)
 1447 |     w[i] += dw;
      |     ~~~~~^~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::decayWeights(T, bool) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1557:38: error: operands to ?: have different types ‘float’ and ‘__half’
 1557 |   T decay_rate = (lifetime > (T)1.0) ? ((T)1.0 / lifetime) : (T)0.0;
      |                  ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1557:38: note:   and each type can be converted to the other
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1569:14: error: no match for ‘operator*=’ (operand types are ‘__half’ and ‘__half’)
 1569 |         w[i] *= (i % this->x_size_ == last_col) ? (T)1.0 : decay_scale;
      |         ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::clipWeights(T) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘__half’ and ‘float’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1594:18: note: in expansion of macro ‘MAX’
 1594 |       w[i] = MIN(MAX(w[i], -clip), clip);
      |                  ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:22: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                      ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1594:18: note: in expansion of macro ‘MAX’
 1594 |       w[i] = MIN(MAX(w[i], -clip), clip);
      |                  ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘__half’ and ‘float’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1594:18: note: in expansion of macro ‘MAX’
 1594 |       w[i] = MIN(MAX(w[i], -clip), clip);
      |                  ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                                ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:29:35: note: in definition of macro ‘MIN’
   29 | #define MIN(a, b) (((a) < (b)) ? (a) : (b))
      |                                   ^
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1594:18: note: in expansion of macro ‘MAX’
 1594 |       w[i] = MIN(MAX(w[i], -clip), clip);
      |                  ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::diffuseWeights() [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1615:12: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘float’)
 1615 |       w[i] += diffusion * rng_->sampleGauss();
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::diffuseWeightsPink() [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1686:27: error: call of overloaded ‘__half(int)’ is ambiguous
 1686 |     T last_noise_value = ((T)__builtin_popcountll(fls));
      |                          ~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1718:22: error: call of overloaded ‘__half(int)’ is ambiguous
 1718 |     T noise_value = ((T)__builtin_popcountll(fls));
      |                     ~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1720:10: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘float’)
 1720 |     w[j] += amp * (noise_value - last_noise_value);
      |     ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::copyIndexedInput(T*, const T*, int, const int*, int, int, int, bool, int, const int*) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:893:33: error: call of overloaded ‘__half(int&)’ is ambiguous
  893 |                               ? (T)j_shifted
      |                                 ^~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:905:33: error: call of overloaded ‘__half(int&)’ is ambiguous
  905 |                               ? (T)j_shifted
      |                                 ^~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:925:33: error: call of overloaded ‘__half(int&)’ is ambiguous
  925 |                               ? (T)j_shifted
      |                                 ^~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:934:33: error: call of overloaded ‘__half(int&)’ is ambiguous
  934 |                               ? (T)j_shifted
      |                                 ^~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp: In instantiation of ‘void RPU::RPUSimple<T>::copyIndexedOutput(T*, const T*, int, const int*, int, int, int, bool, int, const int*) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1765:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:978:69: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘const __half’)
  978 |           out_tensor[(j_shifted - 2) + i_dim3 * output_matrix_size] += src_tensor[idx];
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:989:68: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘const __half’)
  989 |           out_tensor[(j_shifted - 2) + i / M * output_matrix_size] += src_tensor[idx];
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1007:69: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘const __half’)
 1007 |           out_tensor[(j_shifted - 2) + i_dim3 * output_matrix_size] += src_tensor[idx];
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu.cpp:1014:70: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘const __half’)
 1014 |           out_tensor[(j_shifted - 2) + idx / M * output_matrix_size] += src_tensor[idx];
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~
[6/80] Building CXX object CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_buffered_transfer_device.cpp.o
FAILED: CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_buffered_transfer_device.cpp.o 
/gpfs/u/software/npl/gcc/9.3.0/1/bin/g++ -DRPU_TORCH_CUDA_BUFFERS -DRPU_USE_CUDA -DRPU_USE_FP16 -DRPU_USE_MKL -DRPU_USE_WITH_TORCH -D_GLIBCXX_USE_CXX11_ABI=0 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include/python3.10 -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/pybind11/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include -I/gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/lib/python3.10/site-packages/torch/include/torch/csrc/api/include -isystem /gpfs/u/home/ANFM/ANFMbchl/scratch/miniconda3/envs/torch-nightly/include -isystem /usr/local/cuda-12.1/targets/x86_64-linux/include -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda -isystem /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/cuda -fopenmp -Wall -Wno-narrowing -Wno-strict-overflow -g -std=gnu++17 -fPIC -MD -MT CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_buffered_transfer_device.cpp.o -MF CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_buffered_transfer_device.cpp.o.d -o CMakeFiles/RPU_CPU.dir/src/rpucuda/rpu_buffered_transfer_device.cpp.o -c /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp: In instantiation of ‘T RPU::BufferedTransferRPUDeviceMetaParameter<T>::getTransferLR(int, int, T) const [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:70:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:60:14: error: no match for ‘operator*=’ (operand types are ‘__half’ and ‘float’)
   60 |     lr_gamma *= this->gamma_vec[from_device_idx] / this->gamma_vec[to_device_idx];
      |     ~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp: In instantiation of ‘void RPU::BufferedTransferRPUDevice<T>::readAndUpdate(int, int, T, const T*, int, T, int, int) [with T = __half]’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:325:16:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:177:17: error: call of overloaded ‘__half(const int&)’ is ambiguous
  177 |   T max_steps = (T)this->transfer_pwu_->getUpPar().desired_BL;
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:199:13: error: no match for ‘operator+=’ (operand types are ‘__half’ and ‘float’)
  199 |       omega += v_out[j] * lr_abs;
      |       ~~~~~~^~~~~~~~~~~~~~~~~~~~
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: error: operands to ?: have different types ‘__half’ and ‘float’
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                   ~~~~~~~~~~~~~^~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:201:19: note: in expansion of macro ‘MAX’
  201 |       T n_steps = MAX(MIN((T)truncf(omega / buffer_granularity), max_steps), -max_steps);
      |                   ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:32:32: note:   and each type can be converted to the other
   32 | #define MAX(a, b) (((a) > (b)) ? (a) : (b))
      |                   ~~~~~~~~~~~~~^~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:201:19: note: in expansion of macro ‘MAX’
  201 |       T n_steps = MAX(MIN((T)truncf(omega / buffer_granularity), max_steps), -max_steps);
      |                   ^~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:204:41: error: operands to ?: have different types ‘float’ and ‘__half’
  204 |         fp_w[i_w] = (n_steps != (T)0.0) ? omega * par.momentum : omega;
      |                     ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:204:41: note:   and each type can be converted to the other
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:207:33: error: operands to ?: have different types ‘float’ and ‘__half’
  207 |             (n_steps != (T)0.0) ? omega - sub_momentum * n_steps * buffer_granularity : omega;
      |             ~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:207:33: note:   and each type can be converted to the other
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘RPU::IOMetaParameter<__half>::IOMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:82:35:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:46:80:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:70:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:86:16: error: call of overloaded ‘__half(int)’ is ambiguous
   86 |   T _out_res = (T)0;
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘constexpr RPU::PulsedUpdateMetaParameter<__half>::PulsedUpdateMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:82:35:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:46:80:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter() [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:70:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:267:11: error: call of overloaded ‘__half(int)’ is ambiguous
  267 |   T res = (T)0; // this is taken to be in the range 0..1 as positive and negative phases are done
      |           ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:270:22: error: call of overloaded ‘__half(int)’ is ambiguous
  270 |   T x_res_implicit = (T)0; // in case of implicit pulsing. Assumes range 0..1
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:271:22: error: call of overloaded ‘__half(int)’ is ambiguous
  271 |   T d_res_implicit = (T)0;
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘RPU::IOMetaParameter<__half>::IOMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:84:54:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:51:56:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:70:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:86:16: error: call of overloaded ‘__half(int)’ is ambiguous
   86 |   T _out_res = (T)0;
      |                ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h: In instantiation of ‘constexpr RPU::PulsedUpdateMetaParameter<__half>::PulsedUpdateMetaParameter()’:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:84:54:   required from ‘RPU::TransferRPUDeviceMetaParameter<T>::TransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:51:56:   required from ‘RPU::BufferedTransferRPUDeviceMetaParameter<T>::BufferedTransferRPUDeviceMetaParameter(const RPU::PulsedRPUDeviceMetaParameterBase<T>&, int) [with T = __half]’
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:70:17:   required from here
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:267:11: error: call of overloaded ‘__half(int)’ is ambiguous
  267 |   T res = (T)0; // this is taken to be in the range 0..1 as positive and negative phases are done
      |           ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:270:22: error: call of overloaded ‘__half(int)’ is ambiguous
  270 |   T x_res_implicit = (T)0; // in case of implicit pulsing. Assumes range 0..1
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:215:25: note: candidate: ‘__half::__half(double)’
  215 |     __CUDA_HOSTDEVICE__ __half(const double f) { __x = __double2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:214:25: note: candidate: ‘__half::__half(float)’
  214 |     __CUDA_HOSTDEVICE__ __half(const float f) { __x = __float2half(f).__x;  }
      |                         ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(const __half&)’
  192 | struct __CUDA_ALIGN__(2) __half {
      |                          ^~~~~~
/usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.hpp:192:26: note: candidate: ‘constexpr __half::__half(__half&&)’
In file included from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
/gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_pulsed_meta_parameter.h:271:22: error: call of overloaded ‘__half(int)’ is ambiguous
  271 |   T d_res_implicit = (T)0;
      |                      ^~~~
In file included from /usr/local/cuda-12.1/targets/x86_64-linux/include/cuda_fp16.h:4019,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/utility_functions.h:35,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/math_util.h:16,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rng.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_forward_backward_pass.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.h:15,
                 from /gpfs/u/scratch/ANFM/ANFMbchl/public_aihwkit/src/rpucuda/rpu_buffered_transfer_device.cpp:13:
jubueche commented 1 month ago

@maljoras I think this is something that you could maybe easily fix. Or you could tell me how I could resolve it.

maljoras commented 1 month ago

Yes, one has to enable the FP16 auto-conversion during the compilation if I remember correctly. You have to make sure that __CUDA_NO_HALF_CONVERSIONS__ is NOT defined. See https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__INTRINSIC__HALF.html#group__CUDA__MATH__INTRINSIC__HALF

maljoras commented 1 month ago

You could try to set additionally -DRPU_PARAM_FP16=ON. Maybe I set a compiler directive wrongly somewhere. I remember that with this flag it was working fine.

maljoras commented 1 month ago

You might also want to use bfloat as default FP16 type (with -DRPU_BFLOAT_AS_FP16=ON)

maljoras commented 1 month ago

Actually, could also be related to the rather old compiler you are using? One way to go about it might to change each conversion (T) var_or_number to static_cast<T>(var_or_number) . But as far as I remember it compiled fine last time I checked. However, CPU version of fp16 had some problems still, that's why I mentioned it to be experimental.

jubueche commented 1 month ago

I suspect the compiler version to be the problem. Do you know what compiler version worked? On the system I using to check this it is not working. This is my GCC compiler version:

Not searching for unused variables given on the command line.
-- The C compiler identification is GNU 9.3.0
-- The CXX compiler identification is GNU 9.3.0

Regarding the other things you mentioned: I tried -DRPU_BFLOAT_AS_FP16=ON and -DRPU_PARAM_FP16=ON and this and didn't help. __CUDA_NO_HALF_CONVERSIONS__ also did not change anything.

kaoutar55 commented 1 week ago

@jubueche have you tried a different compiler version? what is the status on this testing?