nexgenanalytics/simd-math

Library for length agnostic SIMD intrinsic support and the corresponding math operations

simd-math's Issues

compare implementations to check if all have same API and find inconsistencies

we should compare the various classes against one another to check if:

• all have same API
• identify all inconsistencies

This relates to #2

(Potentially) Incorrect size values of simd::mask for Neon and SSE2 ABIs

Both simd_mask<double, simd_abi::sse>::size() and simd_mask<double, simd_abi::neon>::size() return size of 4 (hardcoded), where underlying vector types are __m128d and uint64x2_t.

Suite of unit tests for the Kokkos SIMD class implementation

• organize files to have an include and test subdirs
• add gtest dependency and ensure it works : just clone the gtest library inside and build within the code

Unit tests need to work on:

• cover all public functionality of the SIMD classes and its child classes.
• X86 CPUs
• ARM CPUs
• IBM Power CPUs
• NVIDIA GPUs

templated simd constructor

One of the simd constructor is templated but the template doesn't seem to be used:

template <class Flags>
  SIMD_ALWAYS_INLINE inline simd(double const* ptr, Flags /*flags*/)
    :m_value(_mm_loadu_pd(ptr))
  {}

What is the meaning for "Flags" here ?
Can we simply remove it ?

add readme/instructions for building and running the tests

Possible missing implementations

Some methods seem to be missing in the API:

sqrt, max and min are missing in simd_common.
cbrt is missing for vsx and neon
exp is missing for vsx and neon
log is implemented only for avx, avx512 and sse and only in case of intel_compiler detected
copysign and multiplysign are missing for vsx and neon and also in hip_wavefront, cuda, pack, scalar and vector_size