void test_vmaxQf32 (void)
{
float32x4_t out_float32x4_t;
float32x4_t arg0_float32x4_t;
float32x4_t arg1_float32x4_t;
out_float32x4_t = vmaxq_f32 (arg0_float32x4_t, arg1_float32x4_t);
}
inline v_float32x4 v_sqrt(const v_float32x4& x)
{
float32x4_t x1 = vmaxq_f32(x.val, vdupq_n_f32(FLT_MIN));
float32x4_t e = vrsqrteq_f32(x1);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
e = vmulq_f32(vrsqrtsq_f32(vmulq_f32(x1, e), e), e);
return v_float32x4(vmulq_f32(x.val, e));
}
void qcms_transform_data_rgba_out_lut_neon(qcms_transform *transform,
unsigned char *src,
unsigned char *dest,
size_t length)
{
size_t i;
unsigned char alpha;
float32_t (*mat)[4] = transform->matrix;
const float32_t *igtbl_r = (float32_t*)transform->input_gamma_table_r;
const float32_t *igtbl_g = (float32_t*)transform->input_gamma_table_g;
const float32_t *igtbl_b = (float32_t*)transform->input_gamma_table_b;
const uint8_t *otdata_r = &transform->output_table_r->data[0];
const uint8_t *otdata_g = &transform->output_table_g->data[0];
const uint8_t *otdata_b = &transform->output_table_b->data[0];
const float32x4_t mat0 = vld1q_f32(mat[0]);
const float32x4_t mat1 = vld1q_f32(mat[1]);
const float32x4_t mat2 = vld1q_f32(mat[2]);
const float32x4_t max = vld1q_dup_f32(&clampMaxValue);
const float32x4_t min = vld1q_dup_f32(&zero);
const float32x4_t scale = vld1q_dup_f32(&floatScale);
float32x4_t vec_r, vec_g, vec_b;
int32x4_t result;
/* CYA */
if (!length)
return;
for (i = 0; i < length; i++) {
/* setup for transforming the pixel */
vec_r = vld1q_dup_f32(&igtbl_r[*src++]);
vec_g = vld1q_dup_f32(&igtbl_g[*src++]);
vec_b = vld1q_dup_f32(&igtbl_b[*src++]);
alpha = *src++;
/* gamma * matrix */
vec_r = vmulq_f32(vec_r, mat0);
vec_g = vmulq_f32(vec_g, mat1);
vec_b = vmulq_f32(vec_b, mat2);
/* crunch, crunch, crunch */
vec_r = vaddq_f32(vec_r, vaddq_f32(vec_g, vec_b));
vec_r = vmaxq_f32(min, vec_r);
vec_r = vminq_f32(max, vec_r);
result = vcvtq_s32_f32(vmulq_f32(vec_r, scale));
/* use calc'd indices to output RGB values */
*dest++ = otdata_r[vgetq_lane_s32(result, 0)];
*dest++ = otdata_g[vgetq_lane_s32(result, 1)];
*dest++ = otdata_b[vgetq_lane_s32(result, 2)];
*dest++ = alpha;
}
}
template <bool align> SIMD_INLINE void HogDirectionHistograms(const float32x4_t & dx, const float32x4_t & dy, Buffer & buffer, size_t col)
{
float32x4_t bestDot = vdupq_n_f32(0);
int32x4_t bestIndex = vdupq_n_s32(0);
for(int i = 0; i < buffer.size; ++i)
{
float32x4_t dot = vaddq_f32(vmulq_f32(dx, buffer.cos[i]), vmulq_f32(dy, buffer.sin[i]));
uint32x4_t mask = vcgtq_f32(dot, bestDot);
bestDot = vmaxq_f32(dot, bestDot);
bestIndex = vbslq_s32(mask, buffer.pos[i], bestIndex);
dot = vnegq_f32(dot);
mask = vcgtq_f32(dot, bestDot);
bestDot = vmaxq_f32(dot, bestDot);
bestIndex = vbslq_s32(mask, buffer.neg[i], bestIndex);
}
Store<align>(buffer.index + col, bestIndex);
Store<align>(buffer.value + col, Sqrt<SIMD_NEON_RCP_ITER>(vaddq_f32(vmulq_f32(dx, dx), vmulq_f32(dy, dy))));
}
dp::math::Box3f ManagerBitSet::calculateBoundingBox( const GroupSharedPtr& group ) const
{
#if defined(SSE)
if ( useSSE )
{
GroupBitSetSharedPtr groupImpl = std::static_pointer_cast<GroupBitSet>(group);
__m128 minValue = _mm_set1_ps( std::numeric_limits<float>::signaling_NaN() );
__m128 maxValue = _mm_set1_ps( std::numeric_limits<float>::signaling_NaN() );
char const* basePtr = reinterpret_cast<char const*>(groupImpl->getMatrices());
for ( size_t index = 0;index < groupImpl->getObjectCount(); ++index )
{
ObjectBitSetSharedPtr objectImpl = std::static_pointer_cast<ObjectBitSet>(groupImpl->getObject( index ));
dp::math::sse::Mat44f const& modelView = *reinterpret_cast<dp::math::sse::Mat44f const*>(basePtr + objectImpl->getTransformIndex() * groupImpl->getMatricesStride());
dp::math::Vec4f const& extent = objectImpl->getExtent();
dp::math::sse::Vec4f vectors[8];
vectors[0] = *reinterpret_cast<dp::math::sse::Vec4f const*>(&objectImpl->getLowerLeft()) * modelView;
dp::math::sse::Vec4f x( extent[0] * modelView[0] );
dp::math::sse::Vec4f y( extent[1] * modelView[1] );
dp::math::sse::Vec4f z( extent[2] * modelView[2] );
vectors[1] = vectors[0] + x;
vectors[2] = vectors[0] + y;
vectors[3] = vectors[1] + y;
vectors[4] = vectors[0] + z;
vectors[5] = vectors[1] + z;
vectors[6] = vectors[2] + z;
vectors[7] = vectors[3] + z;
for ( unsigned int i = 0;i < 8; ++i )
{
minValue = _mm_min_ps( minValue, vectors[i].sse() );
maxValue = _mm_max_ps( maxValue, vectors[i].sse() );
}
}
dp::math::Vec3f minVec, maxVec;
_MM_EXTRACT_FLOAT( minVec[0], minValue, 0);
_MM_EXTRACT_FLOAT( minVec[1], minValue, 1);
_MM_EXTRACT_FLOAT( minVec[2], minValue, 2);
_MM_EXTRACT_FLOAT( maxVec[0], maxValue, 0);
_MM_EXTRACT_FLOAT( maxVec[1], maxValue, 1);
_MM_EXTRACT_FLOAT( maxVec[2], maxValue, 2);
return dp::math::Box3f( minVec, maxVec );
}
else
#elif defined(NEON)
if ( useNEON )
{
GroupBitSetSharedPtr groupImpl = std::static_pointer_cast<GroupBitSet>(group);
float32x4_t minValue = vdupq_n_f32( std::numeric_limits<float>::max() );
float32x4_t maxValue = vdupq_n_f32( -std::numeric_limits<float>::max() );
char const* basePtr = reinterpret_cast<char const*>(groupImpl->getMatrices());
for ( size_t index = 0;index < groupImpl->getObjectCount(); ++index )
{
const ObjectBitSetSharedPtr objectImpl = std::static_pointer_cast<ObjectBitSet>(groupImpl->getObject( index ));
dp::math::neon::Mat44f const& modelView = *reinterpret_cast<dp::math::neon::Mat44f const*>(basePtr + objectImpl->getTransformIndex() * groupImpl->getMatricesStride());
dp::math::Vec4f const& extent = objectImpl->getExtent();
dp::math::neon::Vec4f vectors[8];
vectors[0] = *reinterpret_cast<dp::math::neon::Vec4f const*>(&objectImpl->getLowerLeft()) * modelView;
dp::math::neon::Vec4f x( extent[0] * modelView[0] );
dp::math::neon::Vec4f y( extent[1] * modelView[1] );
dp::math::neon::Vec4f z( extent[2] * modelView[2] );
vectors[1] = vectors[0] + x;
vectors[2] = vectors[0] + y;
vectors[3] = vectors[1] + y;
vectors[4] = vectors[0] + z;
vectors[5] = vectors[1] + z;
vectors[6] = vectors[2] + z;
vectors[7] = vectors[3] + z;
for ( unsigned int i = 0;i < 8; ++i )
{
minValue = vminq_f32( minValue, vectors[i].neon() );
maxValue = vmaxq_f32( maxValue, vectors[i].neon() );
}
}
dp::math::Vec3f minVec, maxVec;
vst1q_lane_f32( &minVec[0], minValue, 0);
vst1q_lane_f32( &minVec[1], minValue, 1);
vst1q_lane_f32( &minVec[2], minValue, 2);
vst1q_lane_f32( &maxVec[0], maxValue, 0);
vst1q_lane_f32( &maxVec[1], maxValue, 1);
vst1q_lane_f32( &maxVec[2], maxValue, 2);
return dp::math::Box3f( minVec, maxVec );
}
else
#endif
// CPU fallback
{
GroupBitSetSharedPtr groupImpl = std::static_pointer_cast<GroupBitSet>(group);
dp::math::Box4f boundingBox;
char const* basePtr = reinterpret_cast<char const*>(groupImpl->getMatrices());
for ( size_t index = 0;index < groupImpl->getObjectCount(); ++index )
{
const ObjectBitSetSharedPtr objectImpl = std::static_pointer_cast<ObjectBitSet>(groupImpl->getObject( index ));
dp::math::Mat44f const& modelView = reinterpret_cast<dp::math::Mat44f const&>(*(basePtr + objectImpl->getTransformIndex() * groupImpl->getMatricesStride()));
dp::math::Vec4f const& extent = objectImpl->getExtent();
dp::math::Vec4f vectors[8];
vectors[0] = (objectImpl->getLowerLeft() * modelView);
dp::math::Vec4f x( extent[0] * modelView.getPtr()[0], extent[0] * modelView.getPtr()[1], extent[0] * modelView.getPtr()[2], extent[0] * modelView.getPtr()[3] );
dp::math::Vec4f y( extent[1] * modelView.getPtr()[4], extent[1] * modelView.getPtr()[5], extent[1] * modelView.getPtr()[6], extent[1] * modelView.getPtr()[7] );
dp::math::Vec4f z( extent[2] * modelView.getPtr()[8], extent[2] * modelView.getPtr()[9], extent[2] * modelView.getPtr()[10], extent[2] * modelView.getPtr()[11] );
vectors[1] = vectors[0] + x;
vectors[2] = vectors[0] + y;
vectors[3] = vectors[1] + y;
vectors[4] = vectors[0] + z;
vectors[5] = vectors[1] + z;
vectors[6] = vectors[2] + z;
vectors[7] = vectors[3] + z;
for ( unsigned int i = 0;i < 8; ++i )
{
boundingBox.update( vectors[i] );
}
}
dp::math::Vec4f lower = boundingBox.getLower();
dp::math::Vec4f upper = boundingBox.getUpper();
return dp::math::Box3f( dp::math::Vec3f( lower[0], lower[1], lower[2]), dp::math::Vec3f( upper[0], upper[1], upper[2]));
}
}
v4sf max_ps(v4sf a, v4sf b) {
return vmaxq_f32(a, b);
}
float32x4_t test_vmaxq_f32(float32x4_t a1, float32x4_t a2) {
// CHECK: test_vmaxq_f32
return vmaxq_f32(a1, a2);
// CHECK llvm.aarch64.neon.fmax.v4f32
}
static forcedinline ParallelType max (ParallelType a, ParallelType b) noexcept { return vmaxq_f32 (a, b); }
inline float32x4_t vmaxq(const float32x4_t & v0, const float32x4_t & v1) { return vmaxq_f32(v0, v1); }
float32x4_t
test_vmaxq_f32 (float32x4_t __a, float32x4_t __b)
{
return vmaxq_f32(__a, __b);
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1],
int* extreme_filter_divergence) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const float32x4_t vec_15 = vdupq_n_f32(WebRtcAec_kMinFarendPSD);
float32x4_t vec_sdSum = vdupq_n_f32(0.0f);
float32x4_t vec_seSum = vdupq_n_f32(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_dfw0 = vld1q_f32(&dfw[0][i]);
const float32x4_t vec_dfw1 = vld1q_f32(&dfw[1][i]);
const float32x4_t vec_efw0 = vld1q_f32(&efw[0][i]);
const float32x4_t vec_efw1 = vld1q_f32(&efw[1][i]);
const float32x4_t vec_xfw0 = vld1q_f32(&xfw[0][i]);
const float32x4_t vec_xfw1 = vld1q_f32(&xfw[1][i]);
float32x4_t vec_sd = vmulq_n_f32(vld1q_f32(&aec->sd[i]), ptrGCoh[0]);
float32x4_t vec_se = vmulq_n_f32(vld1q_f32(&aec->se[i]), ptrGCoh[0]);
float32x4_t vec_sx = vmulq_n_f32(vld1q_f32(&aec->sx[i]), ptrGCoh[0]);
float32x4_t vec_dfw_sumsq = vmulq_f32(vec_dfw0, vec_dfw0);
float32x4_t vec_efw_sumsq = vmulq_f32(vec_efw0, vec_efw0);
float32x4_t vec_xfw_sumsq = vmulq_f32(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = vmlaq_f32(vec_dfw_sumsq, vec_dfw1, vec_dfw1);
vec_efw_sumsq = vmlaq_f32(vec_efw_sumsq, vec_efw1, vec_efw1);
vec_xfw_sumsq = vmlaq_f32(vec_xfw_sumsq, vec_xfw1, vec_xfw1);
vec_xfw_sumsq = vmaxq_f32(vec_xfw_sumsq, vec_15);
vec_sd = vmlaq_n_f32(vec_sd, vec_dfw_sumsq, ptrGCoh[1]);
vec_se = vmlaq_n_f32(vec_se, vec_efw_sumsq, ptrGCoh[1]);
vec_sx = vmlaq_n_f32(vec_sx, vec_xfw_sumsq, ptrGCoh[1]);
vst1q_f32(&aec->sd[i], vec_sd);
vst1q_f32(&aec->se[i], vec_se);
vst1q_f32(&aec->sx[i], vec_sx);
{
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4_t vec_dfwefw0011 = vmulq_f32(vec_dfw0, vec_efw0);
float32x4_t vec_dfwefw0110 = vmulq_f32(vec_dfw0, vec_efw1);
vec_sde.val[0] = vmulq_n_f32(vec_sde.val[0], ptrGCoh[0]);
vec_sde.val[1] = vmulq_n_f32(vec_sde.val[1], ptrGCoh[0]);
vec_dfwefw0011 = vmlaq_f32(vec_dfwefw0011, vec_dfw1, vec_efw1);
vec_dfwefw0110 = vmlsq_f32(vec_dfwefw0110, vec_dfw1, vec_efw0);
vec_sde.val[0] = vmlaq_n_f32(vec_sde.val[0], vec_dfwefw0011, ptrGCoh[1]);
vec_sde.val[1] = vmlaq_n_f32(vec_sde.val[1], vec_dfwefw0110, ptrGCoh[1]);
vst2q_f32(&aec->sde[i][0], vec_sde);
}
{
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_dfwxfw0011 = vmulq_f32(vec_dfw0, vec_xfw0);
float32x4_t vec_dfwxfw0110 = vmulq_f32(vec_dfw0, vec_xfw1);
vec_sxd.val[0] = vmulq_n_f32(vec_sxd.val[0], ptrGCoh[0]);
vec_sxd.val[1] = vmulq_n_f32(vec_sxd.val[1], ptrGCoh[0]);
vec_dfwxfw0011 = vmlaq_f32(vec_dfwxfw0011, vec_dfw1, vec_xfw1);
vec_dfwxfw0110 = vmlsq_f32(vec_dfwxfw0110, vec_dfw1, vec_xfw0);
vec_sxd.val[0] = vmlaq_n_f32(vec_sxd.val[0], vec_dfwxfw0011, ptrGCoh[1]);
vec_sxd.val[1] = vmlaq_n_f32(vec_sxd.val[1], vec_dfwxfw0110, ptrGCoh[1]);
vst2q_f32(&aec->sxd[i][0], vec_sxd);
}
vec_sdSum = vaddq_f32(vec_sdSum, vec_sd);
vec_seSum = vaddq_f32(vec_seSum, vec_se);
}
{
float32x2_t vec_sdSum_total;
float32x2_t vec_seSum_total;
// A B C D
vec_sdSum_total = vpadd_f32(vget_low_f32(vec_sdSum),
vget_high_f32(vec_sdSum));
vec_seSum_total = vpadd_f32(vget_low_f32(vec_seSum),
vget_high_f32(vec_seSum));
// A+B C+D
vec_sdSum_total = vpadd_f32(vec_sdSum_total, vec_sdSum_total);
vec_seSum_total = vpadd_f32(vec_seSum_total, vec_seSum_total);
// A+B+C+D A+B+C+D
sdSum = vget_lane_f32(vec_sdSum_total, 0);
seSum = vget_lane_f32(vec_seSum_total, 0);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard update.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
// Signal extreme filter divergence if the error is significantly larger
// than the nearend (13 dB).
*extreme_filter_divergence = (seSum > (19.95f * sdSum));
}
static float32x4_t vpowq_f32(float32x4_t a, float32x4_t b) {
// a^b = exp2(b * log2(a))
// exp2(x) and log2(x) are calculated using polynomial approximations.
float32x4_t log2_a, b_log2_a, a_exp_b;
// Calculate log2(x), x = a.
{
// To calculate log2(x), we decompose x like this:
// x = y * 2^n
// n is an integer
// y is in the [1.0, 2.0) range
//
// log2(x) = log2(y) + n
// n can be evaluated by playing with float representation.
// log2(y) in a small range can be approximated, this code uses an order
// five polynomial approximation. The coefficients have been
// estimated with the Remez algorithm and the resulting
// polynomial has a maximum relative error of 0.00086%.
// Compute n.
// This is done by masking the exponent, shifting it into the top bit of
// the mantissa, putting eight into the biased exponent (to shift/
// compensate the fact that the exponent has been shifted in the top/
// fractional part and finally getting rid of the implicit leading one
// from the mantissa by substracting it out.
const uint32x4_t vec_float_exponent_mask = vdupq_n_u32(0x7F800000);
const uint32x4_t vec_eight_biased_exponent = vdupq_n_u32(0x43800000);
const uint32x4_t vec_implicit_leading_one = vdupq_n_u32(0x43BF8000);
const uint32x4_t two_n = vandq_u32(vreinterpretq_u32_f32(a),
vec_float_exponent_mask);
const uint32x4_t n_1 = vshrq_n_u32(two_n, kShiftExponentIntoTopMantissa);
const uint32x4_t n_0 = vorrq_u32(n_1, vec_eight_biased_exponent);
const float32x4_t n =
vsubq_f32(vreinterpretq_f32_u32(n_0),
vreinterpretq_f32_u32(vec_implicit_leading_one));
// Compute y.
const uint32x4_t vec_mantissa_mask = vdupq_n_u32(0x007FFFFF);
const uint32x4_t vec_zero_biased_exponent_is_one = vdupq_n_u32(0x3F800000);
const uint32x4_t mantissa = vandq_u32(vreinterpretq_u32_f32(a),
vec_mantissa_mask);
const float32x4_t y =
vreinterpretq_f32_u32(vorrq_u32(mantissa,
vec_zero_biased_exponent_is_one));
// Approximate log2(y) ~= (y - 1) * pol5(y).
// pol5(y) = C5 * y^5 + C4 * y^4 + C3 * y^3 + C2 * y^2 + C1 * y + C0
const float32x4_t C5 = vdupq_n_f32(-3.4436006e-2f);
const float32x4_t C4 = vdupq_n_f32(3.1821337e-1f);
const float32x4_t C3 = vdupq_n_f32(-1.2315303f);
const float32x4_t C2 = vdupq_n_f32(2.5988452f);
const float32x4_t C1 = vdupq_n_f32(-3.3241990f);
const float32x4_t C0 = vdupq_n_f32(3.1157899f);
float32x4_t pol5_y = C5;
pol5_y = vmlaq_f32(C4, y, pol5_y);
pol5_y = vmlaq_f32(C3, y, pol5_y);
pol5_y = vmlaq_f32(C2, y, pol5_y);
pol5_y = vmlaq_f32(C1, y, pol5_y);
pol5_y = vmlaq_f32(C0, y, pol5_y);
const float32x4_t y_minus_one =
vsubq_f32(y, vreinterpretq_f32_u32(vec_zero_biased_exponent_is_one));
const float32x4_t log2_y = vmulq_f32(y_minus_one, pol5_y);
// Combine parts.
log2_a = vaddq_f32(n, log2_y);
}
// b * log2(a)
b_log2_a = vmulq_f32(b, log2_a);
// Calculate exp2(x), x = b * log2(a).
{
// To calculate 2^x, we decompose x like this:
// x = n + y
// n is an integer, the value of x - 0.5 rounded down, therefore
// y is in the [0.5, 1.5) range
//
// 2^x = 2^n * 2^y
// 2^n can be evaluated by playing with float representation.
// 2^y in a small range can be approximated, this code uses an order two
// polynomial approximation. The coefficients have been estimated
// with the Remez algorithm and the resulting polynomial has a
// maximum relative error of 0.17%.
// To avoid over/underflow, we reduce the range of input to ]-127, 129].
const float32x4_t max_input = vdupq_n_f32(129.f);
const float32x4_t min_input = vdupq_n_f32(-126.99999f);
const float32x4_t x_min = vminq_f32(b_log2_a, max_input);
const float32x4_t x_max = vmaxq_f32(x_min, min_input);
// Compute n.
const float32x4_t half = vdupq_n_f32(0.5f);
const float32x4_t x_minus_half = vsubq_f32(x_max, half);
const int32x4_t x_minus_half_floor = vcvtq_s32_f32(x_minus_half);
// Compute 2^n.
const int32x4_t float_exponent_bias = vdupq_n_s32(127);
const int32x4_t two_n_exponent =
vaddq_s32(x_minus_half_floor, float_exponent_bias);
const float32x4_t two_n =
vreinterpretq_f32_s32(vshlq_n_s32(two_n_exponent, kFloatExponentShift));
// Compute y.
const float32x4_t y = vsubq_f32(x_max, vcvtq_f32_s32(x_minus_half_floor));
// Approximate 2^y ~= C2 * y^2 + C1 * y + C0.
const float32x4_t C2 = vdupq_n_f32(3.3718944e-1f);
const float32x4_t C1 = vdupq_n_f32(6.5763628e-1f);
const float32x4_t C0 = vdupq_n_f32(1.0017247f);
float32x4_t exp2_y = C2;
exp2_y = vmlaq_f32(C1, y, exp2_y);
exp2_y = vmlaq_f32(C0, y, exp2_y);
// Combine parts.
a_exp_b = vmulq_f32(exp2_y, two_n);
}
return a_exp_b;
}
