Skip to content

Commit

Permalink
Add a slightly more efficient cubic resampler
Browse files Browse the repository at this point in the history 
Not as beneficial as I'd have hoped, but there does seem to be a slight
improvement.
  • Loading branch information
@kcat
kcat committed Feb 17, 2024
1 parent e39956c commit ed75f54
Show file tree
Hide file tree
Showing 4 changed files with 348 additions and 44 deletions.
8 changes: 8 additions & 0 deletions alc/alu.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -203,6 +203,14 @@ inline ResamplerFunc SelectResampler(Resampler resampler, uint increment)
if((CPUCapFlags&CPU_CAP_NEON))
return Resample_<CubicTag,NEONTag>;
#endif
#ifdef HAVE_SSE4_1
if((CPUCapFlags&CPU_CAP_SSE4_1))
return Resample_<CubicTag,SSE4Tag>;
#endif
#ifdef HAVE_SSE2
if((CPUCapFlags&CPU_CAP_SSE2))
return Resample_<CubicTag,SSE2Tag>;
#endif
#ifdef HAVE_SSE
if((CPUCapFlags&CPU_CAP_SSE))
return Resample_<CubicTag,SSETag>;
Expand Down
136 changes: 104 additions & 32 deletions core/mixer/mixer_neon.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,8 @@ struct FastBSincTag;
#pragma GCC target("fpu=neon")
#endif

using uint = unsigned int;

namespace {

constexpr uint BSincPhaseDiffBits{MixerFracBits - BSincPhaseBits};
Expand All @@ -39,6 +41,19 @@ constexpr uint CubicPhaseDiffBits{MixerFracBits - CubicPhaseBits};
constexpr uint CubicPhaseDiffOne{1 << CubicPhaseDiffBits};
constexpr uint CubicPhaseDiffMask{CubicPhaseDiffOne - 1u};

force_inline
void vtranspose4(float32x4_t &x0, float32x4_t &x1, float32x4_t &x2, float32x4_t &x3) noexcept
{
float32x4x2_t t0_{vzipq_f32(x0, x2)};
float32x4x2_t t1_{vzipq_f32(x1, x3)};
float32x4x2_t u0_{vzipq_f32(t0_.val[0], t1_.val[0])};
float32x4x2_t u1_{vzipq_f32(t0_.val[1], t1_.val[1])};
x0 = u0_.val[0];
x1 = u0_.val[1];
x2 = u1_.val[0];
x3 = u1_.val[1];
}

inline float32x4_t set_f4(float l0, float l1, float l2, float l3)
{
float32x4_t ret{vmovq_n_f32(l0)};
Expand Down Expand Up @@ -150,42 +165,42 @@ void Resample_<LerpTag,NEONTag>(const InterpState*, const float *src, uint frac,
{
ASSUME(frac < MixerFracOne);

const int32x4_t increment4 = vdupq_n_s32(static_cast<int>(increment*4));
const uint32x4_t increment4 = vdupq_n_u32(increment*4u);
const float32x4_t fracOne4 = vdupq_n_f32(1.0f/MixerFracOne);
const int32x4_t fracMask4 = vdupq_n_s32(MixerFracMask);
const uint32x4_t fracMask4 = vdupq_n_u32(MixerFracMask);

alignas(16) std::array<uint,4> pos_, frac_;
InitPosArrays(frac, increment, al::span{frac_}, al::span{pos_});
int32x4_t frac4 = vld1q_s32(reinterpret_cast<int*>(frac_.data()));
int32x4_t pos4 = vld1q_s32(reinterpret_cast<int*>(pos_.data()));
uint32x4_t frac4 = vld1q_u32(frac_.data());
uint32x4_t pos4 = vld1q_u32(pos_.data());

auto dst_iter = dst.begin();
for(size_t todo{dst.size()>>2};todo;--todo)
{
const int pos0{vgetq_lane_s32(pos4, 0)};
const int pos1{vgetq_lane_s32(pos4, 1)};
const int pos2{vgetq_lane_s32(pos4, 2)};
const int pos3{vgetq_lane_s32(pos4, 3)};
const uint pos0{vgetq_lane_u32(pos4, 0)};
const uint pos1{vgetq_lane_u32(pos4, 1)};
const uint pos2{vgetq_lane_u32(pos4, 2)};
const uint pos3{vgetq_lane_u32(pos4, 3)};
const float32x4_t val1{set_f4(src[pos0], src[pos1], src[pos2], src[pos3])};
const float32x4_t val2{set_f4(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
const float32x4_t val2{set_f4(src[pos0+1_uz], src[pos1+1_uz], src[pos2+1_uz], src[pos3+1_uz])};

/* val1 + (val2-val1)*mu */
const float32x4_t r0{vsubq_f32(val2, val1)};
const float32x4_t mu{vmulq_f32(vcvtq_f32_s32(frac4), fracOne4)};
const float32x4_t mu{vmulq_f32(vcvtq_f32_u32(frac4), fracOne4)};
const float32x4_t out{vmlaq_f32(val1, mu, r0)};

vst1q_f32(dst_iter, out);
dst_iter += 4;

frac4 = vaddq_s32(frac4, increment4);
pos4 = vaddq_s32(pos4, vshrq_n_s32(frac4, MixerFracBits));
frac4 = vandq_s32(frac4, fracMask4);
frac4 = vaddq_u32(frac4, increment4);
pos4 = vaddq_u32(pos4, vshrq_n_u32(frac4, MixerFracBits));
frac4 = vandq_u32(frac4, fracMask4);
}

if(size_t todo{dst.size()&3})
{
src += static_cast<uint>(vgetq_lane_s32(pos4, 0));
frac = static_cast<uint>(vgetq_lane_s32(frac4, 0));
src += vgetq_lane_u32(pos4, 0);
frac = vgetq_lane_u32(frac4, 0);

do {
*(dst_iter++) = lerpf(src[0], src[1], static_cast<float>(frac) * (1.0f/MixerFracOne));
Expand All @@ -205,29 +220,86 @@ void Resample_<CubicTag,NEONTag>(const InterpState *state, const float *src, uin

const auto *filter = al::assume_aligned<16>(std::get<CubicState>(*state).filter);

const uint32x4_t increment4{vdupq_n_u32(increment*4u)};
const uint32x4_t fracMask4{vdupq_n_u32(MixerFracMask)};
const float32x4_t fracDiffOne4{vdupq_n_f32(1.0f/CubicPhaseDiffOne)};
const uint32x4_t fracDiffMask4{vdupq_n_u32(CubicPhaseDiffMask)};

alignas(16) std::array<uint,4> pos_, frac_;
InitPosArrays(frac, increment, al::span{frac_}, al::span{pos_});
uint32x4_t frac4{vld1q_u32(frac_.data())};
uint32x4_t pos4{vld1q_u32(pos_.data())};

src -= 1;
std::generate(dst.begin(), dst.end(), [&src,&frac,increment,filter]() -> float
auto dst_iter = dst.begin();
for(size_t todo{dst.size()>>2};todo;--todo)
{
const uint pi{frac >> CubicPhaseDiffBits};
const float pf{static_cast<float>(frac&CubicPhaseDiffMask) * (1.0f/CubicPhaseDiffOne)};
const float32x4_t pf4{vdupq_n_f32(pf)};
const uint pos0{vgetq_lane_u32(pos4, 0)};
const uint pos1{vgetq_lane_u32(pos4, 1)};
const uint pos2{vgetq_lane_u32(pos4, 2)};
const uint pos3{vgetq_lane_u32(pos4, 3)};
const float32x4_t val0{vld1q_f32(src+pos0)};
const float32x4_t val1{vld1q_f32(src+pos1)};
const float32x4_t val2{vld1q_f32(src+pos2)};
const float32x4_t val3{vld1q_f32(src+pos3)};

const uint32x4_t pi4{vshrq_n_u32(frac4, CubicPhaseDiffBits)};
const uint pi0{vgetq_lane_u32(pi4, 0)};
const uint pi1{vgetq_lane_u32(pi4, 1)};
const uint pi2{vgetq_lane_u32(pi4, 2)};
const uint pi3{vgetq_lane_u32(pi4, 3)};

const float32x4_t pf4{vmulq_f32(vcvtq_f32_u32(vandq_u32(frac4, fracDiffMask4)),
fracDiffOne4)};

float32x4_t r0{vmulq_f32(val0,
vmlaq_f32(vld1q_f32(filter[pi0].mCoeffs.data()), vdupq_lane_f32(vget_low_f32(pf4), 0),
vld1q_f32(filter[pi0].mDeltas.data())))};
float32x4_t r1{vmulq_f32(val1,
vmlaq_f32(vld1q_f32(filter[pi1].mCoeffs.data()), vdupq_lane_f32(vget_low_f32(pf4), 1),
vld1q_f32(filter[pi1].mDeltas.data())))};
float32x4_t r2{vmulq_f32(val2,
vmlaq_f32(vld1q_f32(filter[pi2].mCoeffs.data()), vdupq_lane_f32(vget_high_f32(pf4), 0),
vld1q_f32(filter[pi2].mDeltas.data())))};
float32x4_t r3{vmulq_f32(val3,
vmlaq_f32(vld1q_f32(filter[pi3].mCoeffs.data()), vdupq_lane_f32(vget_high_f32(pf4), 1),
vld1q_f32(filter[pi3].mDeltas.data())))};

vtranspose4(r0, r1, r2, r3);
r0 = vaddq_f32(vaddq_f32(r0, r1), vaddq_f32(r2, r3));

vst1q_f32(dst_iter, r0);
dst_iter += 4;

/* Apply the phase interpolated filter. */
frac4 = vaddq_u32(frac4, increment4);
pos4 = vaddq_u32(pos4, vshrq_n_u32(frac4, MixerFracBits));
frac4 = vandq_u32(frac4, fracMask4);
}

/* f = fil + pf*phd */
const float32x4_t f4 = vmlaq_f32(vld1q_f32(filter[pi].mCoeffs.data()), pf4,
vld1q_f32(filter[pi].mDeltas.data()));
/* r = f*src */
float32x4_t r4{vmulq_f32(f4, vld1q_f32(src))};
if(const size_t todo{dst.size()&3})
{
src += vgetq_lane_u32(pos4, 0);
frac = vgetq_lane_u32(frac4, 0);

r4 = vaddq_f32(r4, vrev64q_f32(r4));
const float output{vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0)};
std::generate(dst.end()-todo, dst.end(), [&src,&frac,increment,filter]() -> float
{
const uint pi{frac >> CubicPhaseDiffBits};
const float pf{static_cast<float>(frac&CubicPhaseDiffMask) * (1.0f/CubicPhaseDiffOne)};
const float32x4_t pf4{vdupq_n_f32(pf)};

frac += increment;
src += frac>>MixerFracBits;
frac &= MixerFracMask;
return output;
});
const float32x4_t f4{vmlaq_f32(vld1q_f32(filter[pi].mCoeffs.data()), pf4,
vld1q_f32(filter[pi].mDeltas.data()))};
float32x4_t r4{vmulq_f32(f4, vld1q_f32(src))};

r4 = vaddq_f32(r4, vrev64q_f32(r4));
const float output{vget_lane_f32(vadd_f32(vget_low_f32(r4), vget_high_f32(r4)), 0)};

frac += increment;
src += frac>>MixerFracBits;
frac &= MixerFracMask;
return output;
});
}
}

template<>
Expand Down
124 changes: 118 additions & 6 deletions core/mixer/mixer_sse2.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,17 +28,32 @@

#include "alnumeric.h"
#include "alspan.h"
#include "core/cubic_defs.h"
#include "defs.h"
#include "opthelpers.h"

struct SSE2Tag;
struct LerpTag;
struct CubicTag;


#if defined(__GNUC__) && !defined(__clang__) && !defined(__SSE2__)
#pragma GCC target("sse2")
#endif

using uint = unsigned int;

namespace {

constexpr uint CubicPhaseDiffBits{MixerFracBits - CubicPhaseBits};
constexpr uint CubicPhaseDiffOne{1 << CubicPhaseDiffBits};
constexpr uint CubicPhaseDiffMask{CubicPhaseDiffOne - 1u};

force_inline __m128 vmadd(const __m128 x, const __m128 y, const __m128 z) noexcept
{ return _mm_add_ps(x, _mm_mul_ps(y, z)); }

} // namespace

template<>
void Resample_<LerpTag,SSE2Tag>(const InterpState*, const float *src, uint frac,
const uint increment, const al::span<float> dst)
Expand All @@ -59,12 +74,12 @@ void Resample_<LerpTag,SSE2Tag>(const InterpState*, const float *src, uint frac,
auto dst_iter = dst.begin();
for(size_t todo{dst.size()>>2};todo;--todo)
{
const int pos0{_mm_cvtsi128_si32(pos4)};
const int pos1{_mm_cvtsi128_si32(_mm_srli_si128(pos4, 4))};
const int pos2{_mm_cvtsi128_si32(_mm_srli_si128(pos4, 8))};
const int pos3{_mm_cvtsi128_si32(_mm_srli_si128(pos4, 12))};
const __m128 val1{_mm_setr_ps(src[pos0 ], src[pos1 ], src[pos2 ], src[pos3 ])};
const __m128 val2{_mm_setr_ps(src[pos0+1], src[pos1+1], src[pos2+1], src[pos3+1])};
const auto pos0 = static_cast<uint>(_mm_cvtsi128_si32(pos4));
const auto pos1 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 4)));
const auto pos2 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 8)));
const auto pos3 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 12)));
const __m128 val1{_mm_setr_ps(src[pos0], src[pos1], src[pos2], src[pos3])};
const __m128 val2{_mm_setr_ps(src[pos0+1_uz], src[pos1+1_uz], src[pos2+1_uz], src[pos3+1_uz])};

/* val1 + (val2-val1)*mu */
const __m128 r0{_mm_sub_ps(val2, val1)};
Expand Down Expand Up @@ -93,3 +108,100 @@ void Resample_<LerpTag,SSE2Tag>(const InterpState*, const float *src, uint frac,
} while(--todo);
}
}

template<>
void Resample_<CubicTag,SSE2Tag>(const InterpState *state, const float *src, uint frac,
const uint increment, const al::span<float> dst)
{
ASSUME(frac < MixerFracOne);

const auto *filter = al::assume_aligned<16>(std::get<CubicState>(*state).filter);

const __m128i increment4{_mm_set1_epi32(static_cast<int>(increment*4))};
const __m128i fracMask4{_mm_set1_epi32(MixerFracMask)};
const __m128 fracDiffOne4{_mm_set1_ps(1.0f/CubicPhaseDiffOne)};
const __m128i fracDiffMask4{_mm_set1_epi32(CubicPhaseDiffMask)};

alignas(16) std::array<uint,4> pos_, frac_;
InitPosArrays(frac, increment, al::span{frac_}, al::span{pos_});
__m128i frac4{_mm_setr_epi32(static_cast<int>(frac_[0]), static_cast<int>(frac_[1]),
static_cast<int>(frac_[2]), static_cast<int>(frac_[3]))};
__m128i pos4{_mm_setr_epi32(static_cast<int>(pos_[0]), static_cast<int>(pos_[1]),
static_cast<int>(pos_[2]), static_cast<int>(pos_[3]))};

src -= 1;
auto dst_iter = dst.begin();
for(size_t todo{dst.size()>>2};todo;--todo)
{
const auto pos0 = static_cast<uint>(_mm_cvtsi128_si32(pos4));
const auto pos1 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 4)));
const auto pos2 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 8)));
const auto pos3 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pos4, 12)));
const __m128 val0{_mm_loadu_ps(src+pos0)};
const __m128 val1{_mm_loadu_ps(src+pos1)};
const __m128 val2{_mm_loadu_ps(src+pos2)};
const __m128 val3{_mm_loadu_ps(src+pos3)};

const __m128i pi4{_mm_srli_epi32(frac4, CubicPhaseDiffBits)};
const auto pi0 = static_cast<uint>(_mm_cvtsi128_si32(pi4));
const auto pi1 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pi4, 4)));
const auto pi2 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pi4, 8)));
const auto pi3 = static_cast<uint>(_mm_cvtsi128_si32(_mm_srli_si128(pi4, 12)));

const __m128 pf4{_mm_mul_ps(_mm_cvtepi32_ps(_mm_and_si128(frac4, fracDiffMask4)),
fracDiffOne4)};

__m128 r0{_mm_mul_ps(val0,
vmadd(_mm_load_ps(filter[pi0].mCoeffs.data()),
_mm_shuffle_ps(pf4, pf4, _MM_SHUFFLE(0, 0, 0, 0)),
_mm_load_ps(filter[pi0].mDeltas.data())))};
__m128 r1{_mm_mul_ps(val1,
vmadd(_mm_load_ps(filter[pi1].mCoeffs.data()),
_mm_shuffle_ps(pf4, pf4, _MM_SHUFFLE(1, 1, 1, 1)),
_mm_load_ps(filter[pi1].mDeltas.data())))};
__m128 r2{_mm_mul_ps(val2,
vmadd(_mm_load_ps(filter[pi2].mCoeffs.data()),
_mm_shuffle_ps(pf4, pf4, _MM_SHUFFLE(2, 2, 2, 2)),
_mm_load_ps(filter[pi2].mDeltas.data())))};
__m128 r3{_mm_mul_ps(val3,
vmadd(_mm_load_ps(filter[pi3].mCoeffs.data()),
_mm_shuffle_ps(pf4, pf4, _MM_SHUFFLE(3, 3, 3, 3)),
_mm_load_ps(filter[pi3].mDeltas.data())))};

_MM_TRANSPOSE4_PS(r0, r1, r2, r3);
r0 = _mm_add_ps(_mm_add_ps(r0, r1), _mm_add_ps(r2, r3));

_mm_store_ps(dst_iter, r0);
dst_iter += 4;

frac4 = _mm_add_epi32(frac4, increment4);
pos4 = _mm_add_epi32(pos4, _mm_srli_epi32(frac4, MixerFracBits));
frac4 = _mm_and_si128(frac4, fracMask4);
}

if(const size_t todo{dst.size()&3})
{
src += static_cast<uint>(_mm_cvtsi128_si32(pos4));
frac = static_cast<uint>(_mm_cvtsi128_si32(frac4));

std::generate(dst.end()-todo, dst.end(), [&src,&frac,increment,filter]() -> float
{
const uint pi{frac >> CubicPhaseDiffBits};
const float pf{static_cast<float>(frac&CubicPhaseDiffMask) * (1.0f/CubicPhaseDiffOne)};
const __m128 pf4{_mm_set1_ps(pf)};

const __m128 f4 = vmadd(_mm_load_ps(filter[pi].mCoeffs.data()), pf4,
_mm_load_ps(filter[pi].mDeltas.data()));
__m128 r4{_mm_mul_ps(f4, _mm_loadu_ps(src))};

r4 = _mm_add_ps(r4, _mm_shuffle_ps(r4, r4, _MM_SHUFFLE(0, 1, 2, 3)));
r4 = _mm_add_ps(r4, _mm_movehl_ps(r4, r4));
const float output{_mm_cvtss_f32(r4)};

frac += increment;
src += frac>>MixerFracBits;
frac &= MixerFracMask;
return output;
});
}
}
Loading

0 comments on commit ed75f54

Please sign in to comment.