sam/sfizz
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Added another subtract SIMD helper

Browse source
This commit is contained in:
paulfd 2019-09-15 22:47:24 +02:00
parent 563e7624d9
commit a900580a66
3 changed files with 44 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
sfizz/SIMDDummy.cpp
View file

@ -126,6 +126,13 @@ void subtract<float, true>(absl::Span<const float> input, absl::Span<float> outp
subtract<float, false>(input, output);
}
template <>
void subtract<float, true>(const float value, absl::Span<float> output) noexcept
{
subtract<float, false>(value, output);
}
template <>
void copy<float, true>(absl::Span<const float> input, absl::Span<float> output) noexcept
{

17
sfizz/SIMDHelpers.h
View file

@ -374,6 +374,21 @@ inline void snippetSubtract(const T*& input, T*& output)
*output++ -= *input++;
}
template <class T>
inline void snippetSubtract(const T value, T*& output)
{
*output++ -= value;
}
template <class T, bool SIMD = SIMDConfig::subtract>
void subtract(const T value, absl::Span<T> output) noexcept
{
auto* out = output.begin();
auto* sentinel = output.end();
while (out < sentinel)
snippetSubtract(value, out);
}
template <class T, bool SIMD = SIMDConfig::subtract>
void subtract(absl::Span<const T> input, absl::Span<T> output) noexcept
{
@ -388,6 +403,8 @@ void subtract(absl::Span<const T> input, absl::Span<T> output) noexcept
template <>
void subtract<float, true>(absl::Span<const float> input, absl::Span<float> output) noexcept;
template <>
void subtract<float, true>(const float value, absl::Span<float> output) noexcept;
template <class T>
void snippetCopy(const T*& input, T*& output)

20
sfizz/SIMDSSE.cpp
View file

@ -564,6 +564,26 @@ void subtract<float, true>(absl::Span<const float> input, absl::Span<float> outp
snippetSubtract<float>(in, out);
}
template <>
void subtract<float, true>(const float value, absl::Span<float> output) noexcept
{
auto* out = output.begin();
auto* sentinel = output.end();
const auto* lastAligned = prevAligned(sentinel);
while (unaligned(out) && out < lastAligned)
snippetSubtract<float>(value, out);
auto mmValue = _mm_set_ps1(value);
while (out < lastAligned) {
_mm_store_ps(out, _mm_sub_ps(_mm_load_ps(out), mmValue));
out += TypeAlignment;
}
while (out < sentinel)
snippetSubtract<float>(value, out);
}
template <>
void copy<float, true>(absl::Span<const float> input, absl::Span<float> output) noexcept
{