void srslte_tdec_sse_decision(srslte_tdec_sse_t * h, uint8_t *output, uint32_t long_cb)
{
__m128i zero = _mm_set1_epi16(0);
__m128i lsb_mask = _mm_set1_epi16(1);
__m128i *appPtr = (__m128i*) h->app1;
__m128i *outPtr = (__m128i*) output;
__m128i ap, out, out0, out1;
for (uint32_t i = 0; i < long_cb/16; i++) {
ap = _mm_load_si128(appPtr); appPtr++;
out0 = _mm_and_si128(_mm_cmpgt_epi16(ap, zero), lsb_mask);
ap = _mm_load_si128(appPtr); appPtr++;
out1 = _mm_and_si128(_mm_cmpgt_epi16(ap, zero), lsb_mask);
out = _mm_packs_epi16(out0, out1);
_mm_store_si128(outPtr, out);
outPtr++;
}
if (long_cb%16) {
for (int i=0;i<8;i++) {
output[long_cb-8+i] = h->app1[long_cb-8+i]>0?1:0;
}
}
}
void conv_Short1ToFloat2(void* dst, const void* s, s32 numSamples)
{
LSfloat* d = reinterpret_cast<LSfloat*>(dst);
const LSshort* src = reinterpret_cast<const LSshort*>(s);
s32 num = numSamples >> 3; //8個のshortをまとめて処理
s32 offset = num << 3;
s32 rem = numSamples - offset;
const __m128i izero = _mm_setzero_si128();
const __m128 fcoff = _mm_set1_ps(1.0f/32767.0f);
const LSshort* p = src;
LSfloat* q = d;
for(s32 i=0; i<num; ++i){
__m128i t = _mm_loadu_si128((const __m128i*)p);
__m128i s16_0 = _mm_unpackhi_epi16(t, t);
__m128i s16_1 = _mm_unpacklo_epi16(t, t);
__m128i t1 = _mm_cmpgt_epi16(izero, s16_0);
__m128i t2 = _mm_cmpgt_epi16(izero, s16_1);
__m128i s32_0 = _mm_unpackhi_epi16(s16_0, t1);
__m128i s32_1 = _mm_unpacklo_epi16(s16_0, t1);
__m128i s32_2 = _mm_unpackhi_epi16(s16_1, t2);
__m128i s32_3 = _mm_unpacklo_epi16(s16_1, t2);
//32bit浮動小数点に変換
__m128 f32_0 = _mm_mul_ps(_mm_cvtepi32_ps(s32_0), fcoff);
__m128 f32_1 = _mm_mul_ps(_mm_cvtepi32_ps(s32_1), fcoff);
__m128 f32_2 = _mm_mul_ps(_mm_cvtepi32_ps(s32_2), fcoff);
__m128 f32_3 = _mm_mul_ps(_mm_cvtepi32_ps(s32_3), fcoff);
_mm_storeu_ps((q+0), f32_3);
_mm_storeu_ps((q+4), f32_2);
_mm_storeu_ps((q+8), f32_1);
_mm_storeu_ps((q+12), f32_0);
p += 8;
q += 16;
}
for(s32 i=0; i<rem; ++i){
s32 j = i<<1;
q[j+0] = toFloat(p[i]);
q[j+1] = toFloat(p[i]);
}
}
// 16-bit pixels clip with bd (10/12)
static void highbd_clip(__m128i *p, int numVecs, int bd) {
const __m128i zero = _mm_setzero_si128();
const __m128i one = _mm_set1_epi16(1);
const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
__m128i clamped, mask;
int i;
for (i = 0; i < numVecs; i++) {
mask = _mm_cmpgt_epi16(p[i], max);
clamped = _mm_andnot_si128(mask, p[i]);
mask = _mm_and_si128(mask, max);
clamped = _mm_or_si128(mask, clamped);
mask = _mm_cmpgt_epi16(clamped, zero);
p[i] = _mm_and_si128(clamped, mask);
}
}
// Predictors13: ClampedAddSubtractHalf
static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
int num_pixels, uint32_t* out) {
int i;
const __m128i zero = _mm_setzero_si128();
for (i = 0; i + 2 <= num_pixels; i += 2) {
// we can only process two pixels at a time
const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]);
const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]);
const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]);
const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]);
const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
const __m128i sum = _mm_add_epi16(T_lo, L_lo);
const __m128i avg = _mm_srli_epi16(sum, 1);
const __m128i A1 = _mm_sub_epi16(avg, TL_lo);
const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg);
const __m128i A2 = _mm_sub_epi16(A1, bit_fix);
const __m128i A3 = _mm_srai_epi16(A2, 1);
const __m128i A4 = _mm_add_epi16(avg, A3);
const __m128i pred = _mm_packus_epi16(A4, A4);
const __m128i res = _mm_sub_epi8(src, pred);
_mm_storel_epi64((__m128i*)&out[i], res);
}
if (i != num_pixels) {
VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i);
}
}
//----------------------------------------------------------------------------
void conv_Short1ToFloat1(void* dst, const void* s, s32 numSamples)
{
LSfloat* d = reinterpret_cast<LSfloat*>(dst);
const LSshort* src = reinterpret_cast<const LSshort*>(s);
s32 num = numSamples >> 3; //8個のshortをまとめて処理
s32 offset = num << 3;
s32 rem = numSamples - offset;
const __m128i izero = _mm_setzero_si128();
const __m128 fcoff = _mm_set1_ps(1.0f/32767.0f);
const LSshort* p = src;
LSfloat* q = d;
for(s32 i=0; i<num; ++i){
//32bit浮動小数点r0, r1に変換
__m128i t0 = _mm_loadu_si128((const __m128i*)p);
__m128i t1 = _mm_cmpgt_epi16(izero, t0);
__m128 r0 = _mm_cvtepi32_ps(_mm_unpackhi_epi16(t0, t1));
__m128 r1 = _mm_cvtepi32_ps(_mm_unpacklo_epi16(t0, t1));
r0 = _mm_mul_ps(r0, fcoff);
r1 = _mm_mul_ps(r1, fcoff);
_mm_storeu_ps((q+0), r1);
_mm_storeu_ps((q+4), r0);
p += 8;
q += 8;
}
for(s32 i=0; i<rem; ++i){
q[i] = toFloat(p[i]);
}
}
__m128i test_mm_cmpgt_epi16(__m128i A, __m128i B) {
// DAG-LABEL: test_mm_cmpgt_epi16
// DAG: icmp sgt <8 x i16>
//
// ASM-LABEL: test_mm_cmpgt_epi16
// ASM: pcmpgtw
return _mm_cmpgt_epi16(A, B);
}
SIMDValue SIMDInt16x8Operation::OpGreaterThan(const SIMDValue& aValue, const SIMDValue& bValue)
{
X86SIMDValue x86Result;
X86SIMDValue tmpaValue = X86SIMDValue::ToX86SIMDValue(aValue);
X86SIMDValue tmpbValue = X86SIMDValue::ToX86SIMDValue(bValue);
x86Result.m128i_value = _mm_cmpgt_epi16(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a > b?
return X86SIMDValue::ToSIMDValue(x86Result);
}
SIMDValue SIMDInt16x8Operation::OpGreaterThanOrEqual(const SIMDValue& aValue, const SIMDValue& bValue)
{
X86SIMDValue x86Result, x86Result1, x86Result2;
X86SIMDValue tmpaValue = X86SIMDValue::ToX86SIMDValue(aValue);
X86SIMDValue tmpbValue = X86SIMDValue::ToX86SIMDValue(bValue);
x86Result1.m128i_value = _mm_cmpgt_epi16(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a > b?
x86Result2.m128i_value = _mm_cmpeq_epi16(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a == b?
x86Result.m128i_value = _mm_or_si128(x86Result1.m128i_value, x86Result2.m128i_value);
return X86SIMDValue::ToSIMDValue(x86Result);
}
__m64 _m_pcmpgtw(__m64 _MM1, __m64 _MM2)
{
__m128i lhs = {0}, rhs = {0};
lhs.m128i_i64[0] = _MM1.m64_i64;
rhs.m128i_i64[0] = _MM2.m64_i64;
lhs = _mm_cmpgt_epi16(lhs, rhs);
_MM1.m64_i64 = lhs.m128i_i64[0];
return _MM1;
}
static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
uint32_t c2) {
const __m128i zero = _mm_setzero_si128();
const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
const __m128i avg = _mm_add_epi16(C1, C0);
const __m128i A0 = _mm_srli_epi16(avg, 1);
const __m128i A1 = _mm_sub_epi16(A0, B0);
const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
const __m128i A2 = _mm_sub_epi16(A1, BgtA);
const __m128i A3 = _mm_srai_epi16(A2, 1);
const __m128i A4 = _mm_add_epi16(A0, A3);
const __m128i A5 = _mm_packus_epi16(A4, A4);
const uint32_t output = _mm_cvtsi128_si32(A5);
return output;
}
void conv_Short2ToShort1(void* dst, const void* s, s32 numSamples)
{
LSshort* d = reinterpret_cast<LSshort*>(dst);
const LSshort* src = reinterpret_cast<const LSshort*>(s);
s32 num = numSamples >> 2; //8個のshortをまとめて処理
s32 offset = num << 2;
s32 rem = numSamples - offset;
const __m128i izero = _mm_setzero_si128();
__declspec(align(16)) LSshort tmp[8];
const LSshort* p = src;
LSshort* q = d;
for(s32 i=0; i<num; ++i){
//32bit整数r0, r1に変換
__m128i t0 = _mm_loadu_si128((const __m128i*)p);
__m128i t1 = _mm_cmpgt_epi16(izero, t0);
__m128i r0 = _mm_unpackhi_epi16(t0, t1);
__m128i r1 = _mm_unpacklo_epi16(t0, t1);
__m128i r2 = _mm_add_epi32(r0, _mm_shuffle_epi32(r0, _MM_SHUFFLE(2, 3, 0, 1)));
__m128i r3 = _mm_add_epi32(r1, _mm_shuffle_epi32(r1, _MM_SHUFFLE(2, 3, 0, 1)));
r2 = _mm_srai_epi32(r2, 1);
r3 = _mm_srai_epi32(r3, 1);
__m128i r4 = _mm_packs_epi32(r3, r2);
_mm_store_si128((__m128i*)tmp, r4);
q[0] = tmp[0];
q[1] = tmp[2];
q[2] = tmp[4];
q[3] = tmp[6];
p += 8;
q += 4;
}
for(s32 i=0; i<rem; ++i){
s32 j = i<<1;
s32 t = (p[j+0] + p[j+1]) >> 1;
q[i] = static_cast<LSshort>(t);
}
}
void conv_Short2ToFloat1(void* dst, const void* s, s32 numSamples)
{
LSfloat* d = reinterpret_cast<LSfloat*>(dst);
const LSshort* src = reinterpret_cast<const LSshort*>(s);
s32 num = numSamples >> 2; //8個のshortをまとめて処理
s32 offset = num << 2;
s32 rem = numSamples - offset;
const __m128i izero = _mm_setzero_si128();
const __m128 fcoff = _mm_set1_ps(0.5f/32767.0f); //half
const LSshort* p = src;
LSfloat* q = d;
for(s32 i=0; i<num; ++i){
//32bit整数に変換
__m128i t0 = _mm_loadu_si128((const __m128i*)p);
__m128i t1 = _mm_cmpgt_epi16(izero, t0);
__m128i s32_0 = _mm_unpackhi_epi16(t0, t1);
__m128i s32_1 = _mm_unpacklo_epi16(t0, t1);
//32bit浮動小数点に変換
__m128 f32_0 = _mm_mul_ps(_mm_cvtepi32_ps(s32_0), fcoff);
__m128 f32_1 = _mm_mul_ps(_mm_cvtepi32_ps(s32_1), fcoff);
__m128 f32_2 = _mm_add_ps(f32_0, _mm_shuffle_ps(f32_0, f32_0, _MM_SHUFFLE(2, 3, 0, 1)));
__m128 f32_3 = _mm_add_ps(f32_1, _mm_shuffle_ps(f32_1, f32_1, _MM_SHUFFLE(2, 3, 0, 1)));
__m128 r = _mm_shuffle_ps(f32_3, f32_2, _MM_SHUFFLE(2, 0, 2, 0));
_mm_storeu_ps(q, r);
p += 8;
q += 4;
}
for(s32 i=0; i<rem; ++i){
s32 j = i<<1;
q[i] = 0.5f*(toFloat(p[j+0]) + toFloat(p[j+1]));
}
}
static void
thresh_16s( const Mat& _src, Mat& _dst, short thresh, short maxval, int type )
{
int i, j;
Size roi = _src.size();
roi.width *= _src.channels();
const short* src = _src.ptr<short>();
short* dst = _dst.ptr<short>();
size_t src_step = _src.step/sizeof(src[0]);
size_t dst_step = _dst.step/sizeof(dst[0]);
#if CV_SSE2
volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE);
#endif
if( _src.isContinuous() && _dst.isContinuous() )
{
roi.width *= roi.height;
roi.height = 1;
src_step = dst_step = roi.width;
}
#ifdef HAVE_TEGRA_OPTIMIZATION
if (tegra::thresh_16s(_src, _dst, roi.width, roi.height, thresh, maxval, type))
return;
#endif
#if defined(HAVE_IPP)
CV_IPP_CHECK()
{
IppiSize sz = { roi.width, roi.height };
CV_SUPPRESS_DEPRECATED_START
switch( type )
{
case THRESH_TRUNC:
#ifndef HAVE_IPP_ICV_ONLY
if (_src.data == _dst.data && ippiThreshold_GT_16s_C1IR(dst, (int)dst_step*sizeof(dst[0]), sz, thresh) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
#endif
if (ippiThreshold_GT_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
setIppErrorStatus();
break;
case THRESH_TOZERO:
#ifndef HAVE_IPP_ICV_ONLY
if (_src.data == _dst.data && ippiThreshold_LTVal_16s_C1IR(dst, (int)dst_step*sizeof(dst[0]), sz, thresh + 1, 0) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
#endif
if (ippiThreshold_LTVal_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh+1, 0) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
setIppErrorStatus();
break;
case THRESH_TOZERO_INV:
#ifndef HAVE_IPP_ICV_ONLY
if (_src.data == _dst.data && ippiThreshold_GTVal_16s_C1IR(dst, (int)dst_step*sizeof(dst[0]), sz, thresh, 0) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
#endif
if (ippiThreshold_GTVal_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh, 0) >= 0)
{
CV_IMPL_ADD(CV_IMPL_IPP);
return;
}
setIppErrorStatus();
break;
}
CV_SUPPRESS_DEPRECATED_END
}
#endif
switch( type )
{
case THRESH_BINARY:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh), maxval8 = _mm_set1_epi16(maxval);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_cmpgt_epi16( v0, thresh8 );
v1 = _mm_cmpgt_epi16( v1, thresh8 );
v0 = _mm_and_si128( v0, maxval8 );
v1 = _mm_and_si128( v1, maxval8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#elif CV_NEON
int16x8_t v_thresh = vdupq_n_s16(thresh), v_maxval = vdupq_n_s16(maxval);
for( ; j <= roi.width - 8; j += 8 )
{
uint16x8_t v_mask = vcgtq_s16(vld1q_s16(src + j), v_thresh);
vst1q_s16(dst + j, vandq_s16(vreinterpretq_s16_u16(v_mask), v_maxval));
}
#endif
for( ; j < roi.width; j++ )
dst[j] = src[j] > thresh ? maxval : 0;
}
break;
case THRESH_BINARY_INV:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh), maxval8 = _mm_set1_epi16(maxval);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_cmpgt_epi16( v0, thresh8 );
v1 = _mm_cmpgt_epi16( v1, thresh8 );
v0 = _mm_andnot_si128( v0, maxval8 );
v1 = _mm_andnot_si128( v1, maxval8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#elif CV_NEON
int16x8_t v_thresh = vdupq_n_s16(thresh), v_maxval = vdupq_n_s16(maxval);
for( ; j <= roi.width - 8; j += 8 )
{
uint16x8_t v_mask = vcleq_s16(vld1q_s16(src + j), v_thresh);
vst1q_s16(dst + j, vandq_s16(vreinterpretq_s16_u16(v_mask), v_maxval));
}
#endif
for( ; j < roi.width; j++ )
dst[j] = src[j] <= thresh ? maxval : 0;
}
break;
case THRESH_TRUNC:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_min_epi16( v0, thresh8 );
v1 = _mm_min_epi16( v1, thresh8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#elif CV_NEON
int16x8_t v_thresh = vdupq_n_s16(thresh);
for( ; j <= roi.width - 8; j += 8 )
vst1q_s16(dst + j, vminq_s16(vld1q_s16(src + j), v_thresh));
#endif
for( ; j < roi.width; j++ )
dst[j] = std::min(src[j], thresh);
}
break;
case THRESH_TOZERO:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_and_si128(v0, _mm_cmpgt_epi16(v0, thresh8));
v1 = _mm_and_si128(v1, _mm_cmpgt_epi16(v1, thresh8));
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#elif CV_NEON
int16x8_t v_thresh = vdupq_n_s16(thresh);
for( ; j <= roi.width - 8; j += 8 )
{
int16x8_t v_src = vld1q_s16(src + j);
uint16x8_t v_mask = vcgtq_s16(v_src, v_thresh);
vst1q_s16(dst + j, vandq_s16(vreinterpretq_s16_u16(v_mask), v_src));
}
#endif
for( ; j < roi.width; j++ )
{
short v = src[j];
dst[j] = v > thresh ? v : 0;
}
}
break;
case THRESH_TOZERO_INV:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_andnot_si128(_mm_cmpgt_epi16(v0, thresh8), v0);
v1 = _mm_andnot_si128(_mm_cmpgt_epi16(v1, thresh8), v1);
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#elif CV_NEON
int16x8_t v_thresh = vdupq_n_s16(thresh);
for( ; j <= roi.width - 8; j += 8 )
{
int16x8_t v_src = vld1q_s16(src + j);
uint16x8_t v_mask = vcleq_s16(v_src, v_thresh);
vst1q_s16(dst + j, vandq_s16(vreinterpretq_s16_u16(v_mask), v_src));
}
#endif
for( ; j < roi.width; j++ )
{
short v = src[j];
dst[j] = v <= thresh ? v : 0;
}
}
break;
default:
return CV_Error( CV_StsBadArg, "" );
}
}
int
global_sse2_word(int queryLength,
unsigned short *profile,
const unsigned char *dbSeq,
int dbLength,
unsigned short gapOpen,
unsigned short gapExtend,
unsigned short ceiling,
struct f_struct *f_str)
{
int i, j;
int score;
int scale;
int temp;
int distance;
int offset;
int position;
int cmp;
int iter;
__m128i *pvH;
__m128i *pvE;
__m128i vE, vF, vH;
__m128i vHNext;
__m128i vFPrev;
__m128i vGapOpen;
__m128i vGapExtend;
__m128i vCeiling;
__m128i vScale;
__m128i vScaleAmt;
__m128i vScaleTmp;
__m128i vTemp;
__m128i vNull;
__m128i *pvScore;
scale = 0;
iter = (queryLength + 7) / 8;
offset = (queryLength - 1) % iter;
position = 7 - (queryLength - 1) / iter;
pvH = (__m128i *)f_str->workspace;
pvE = pvH + iter;
/* Load gap opening penalty to all elements of a constant */
vGapOpen = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vGapOpen = _mm_insert_epi16 (vGapOpen, gapOpen, 0);
vGapOpen = _mm_shufflelo_epi16 (vGapOpen, 0);
vGapOpen = _mm_shuffle_epi32 (vGapOpen, 0);
/* Load gap extension penalty to all elements of a constant */
vGapExtend = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vGapExtend = _mm_insert_epi16 (vGapExtend, gapExtend, 0);
vGapExtend = _mm_shufflelo_epi16 (vGapExtend, 0);
vGapExtend = _mm_shuffle_epi32 (vGapExtend, 0);
/* Generate the ceiling before scaling */
vTemp = _mm_setzero_si128(); /* transfered from Apple Devel smith_waterman_sse2.c fix */
vTemp = _mm_insert_epi16 (vTemp, ceiling, 0);
vTemp = _mm_shufflelo_epi16 (vTemp, 0);
vTemp = _mm_shuffle_epi32 (vTemp, 0);
vCeiling = _mm_cmpeq_epi16 (vTemp, vTemp);
vCeiling = _mm_srli_epi16 (vCeiling, 1);
vCeiling = _mm_subs_epi16 (vCeiling, vTemp);
vCeiling = _mm_subs_epi16 (vCeiling, vGapOpen);
vNull = _mm_cmpeq_epi16 (vTemp, vTemp);
vNull = _mm_slli_epi16 (vNull, 15);
vScaleAmt = _mm_xor_si128 (vNull, vNull);
/* Zero out the storage vector */
vTemp = _mm_adds_epi16 (vNull, vGapOpen);
for (i = 0; i < iter; i++) {
_mm_store_si128 (pvH + i, vTemp);
_mm_store_si128 (pvE + i, vNull);
}
/* initialize F */
vF = vNull;
vFPrev = vNull;
/* load and scale H for the next round */
vTemp = _mm_srli_si128 (vGapOpen, 14);
vH = _mm_load_si128 (pvH + iter - 1);
vH = _mm_adds_epi16 (vH, vTemp);
for (i = 0; i < dbLength; ++i) {
/* fetch first data asap. */
pvScore = (__m128i *) profile + dbSeq[i] * iter;
vF = vNull;
vH = _mm_max_epi16 (vH, vFPrev);
for (j = 0; j < iter; j++) {
/* correct H from the previous columns F */
vHNext = _mm_load_si128 (pvH + j);
vHNext = _mm_max_epi16 (vHNext, vFPrev);
/* load and correct E value */
vE = _mm_load_si128 (pvE + j);
vTemp = _mm_subs_epi16 (vHNext, vGapOpen);
vE = _mm_max_epi16 (vE, vTemp);
_mm_store_si128 (pvE + j, vE);
/* add score to vH */
vH = _mm_adds_epi16 (vH, *pvScore++);
/* get max from vH, vE and vF */
vH = _mm_max_epi16 (vH, vE);
vH = _mm_max_epi16 (vH, vF);
_mm_store_si128 (pvH + j, vH);
/* update vF value */
vH = _mm_subs_epi16 (vH, vGapOpen);
vF = _mm_max_epi16 (vF, vH);
/* load the next h values */
vH = vHNext;
}
/* check if we need to scale before the next round */
vTemp = _mm_cmpgt_epi16 (vF, vCeiling);
cmp = _mm_movemask_epi8 (vTemp);
/* broadcast F values */
vF = _mm_xor_si128 (vF, vNull);
vTemp = _mm_slli_si128 (vF, 2);
vTemp = _mm_subs_epu16 (vTemp, vScaleAmt);
vF = max_epu16 (vF, vTemp);
vTemp = _mm_slli_si128 (vF, 4);
vScaleTmp = _mm_slli_si128 (vScaleAmt, 2);
vScaleTmp = _mm_adds_epu16 (vScaleTmp, vScaleAmt);
vTemp = _mm_subs_epu16 (vTemp, vScaleTmp);
vF = max_epu16 (vF, vTemp);
vTemp = _mm_slli_si128 (vScaleTmp, 4);
vScaleTmp = _mm_adds_epu16 (vScaleTmp, vTemp);
vTemp = _mm_slli_si128 (vF, 8);
vTemp = _mm_subs_epu16 (vTemp, vScaleTmp);
vF = max_epu16 (vF, vTemp);
/* scale if necessary */
if (cmp != 0x0000) {
__m128i vScale1;
__m128i vScale2;
vScale = _mm_slli_si128 (vF, 2);
vScale = _mm_subs_epu16 (vScale, vGapOpen);
vScale = _mm_subs_epu16 (vScale, vScaleAmt);
vTemp = _mm_slli_si128 (vScale, 2);
vTemp = _mm_subs_epu16 (vScale, vTemp);
vScaleAmt = _mm_adds_epu16 (vScaleAmt, vTemp);
vTemp = _mm_slli_si128 (vScale, 2);
vTemp = _mm_subs_epu16 (vTemp, vScale);
vScaleAmt = _mm_subs_epu16 (vScaleAmt, vTemp);
/* rescale the previous F */
vF = _mm_subs_epu16 (vF, vScale);
/* check if we can continue in signed 16-bits */
vTemp = _mm_xor_si128 (vF, vNull);
vTemp = _mm_cmpgt_epi16 (vTemp, vCeiling);
cmp = _mm_movemask_epi8 (vTemp);
if (cmp != 0x0000) {
return OVERFLOW_SCORE;
}
vTemp = _mm_adds_epi16 (vCeiling, vCeiling);
vScale1 = _mm_subs_epu16 (vScale, vTemp);
vScale2 = _mm_subs_epu16 (vScale, vScale1);
/* scale all the vectors */
for (j = 0; j < iter; j++) {
/* load H and E */
vH = _mm_load_si128 (pvH + j);
vE = _mm_load_si128 (pvE + j);
/* get max from vH, vE and vF */
vH = _mm_subs_epi16 (vH, vScale1);
vH = _mm_subs_epi16 (vH, vScale2);
vE = _mm_subs_epi16 (vE, vScale1);
vE = _mm_subs_epi16 (vE, vScale2);
/* save the H and E */
_mm_store_si128 (pvH + j, vH);
_mm_store_si128 (pvE + j, vE);
}
vScale = vScaleAmt;
for (j = 0; j < position; ++j) {
vScale = _mm_slli_si128 (vScale, 2);
}
/* calculate the final scaling amount */
vTemp = _mm_xor_si128 (vTemp, vTemp);
vScale1 = _mm_unpacklo_epi16 (vScale, vTemp);
vScale2 = _mm_unpackhi_epi16 (vScale, vTemp);
vScale = _mm_add_epi32 (vScale1, vScale2);
vTemp = _mm_srli_si128 (vScale, 8);
vScale = _mm_add_epi32 (vScale, vTemp);
vTemp = _mm_srli_si128 (vScale, 4);
vScale = _mm_add_epi32 (vScale, vTemp);
scale = (int) (unsigned short) _mm_extract_epi16 (vScale, 0);
temp = (int) (unsigned short) _mm_extract_epi16 (vScale, 1);
scale = scale + (temp << 16);
}
/* scale the F value for the next round */
vFPrev = _mm_slli_si128 (vF, 2);
vFPrev = _mm_subs_epu16 (vFPrev, vScaleAmt);
vFPrev = _mm_xor_si128 (vFPrev, vNull);
/* load and scale H for the next round */
vH = _mm_load_si128 (pvH + iter - 1);
vH = _mm_xor_si128 (vH, vNull);
vH = _mm_slli_si128 (vH, 2);
vH = _mm_subs_epu16 (vH, vScaleAmt);
vH = _mm_insert_epi16 (vH, gapOpen, 0);
vH = _mm_xor_si128 (vH, vNull);
}
vH = _mm_load_si128 (pvH + offset);
vH = _mm_max_epi16 (vH, vFPrev);
for (j = 0; j < position; ++j) {
vH = _mm_slli_si128 (vH, 2);
}
score = (int) (signed short) _mm_extract_epi16 (vH, 7);
score = score + SHORT_BIAS;
/* return largest score */
distance = (queryLength + dbLength) * gapExtend;
score = score - (gapOpen * 2) - distance + scale;
return score;
}
static void GF_FUNC_ALIGN VS_CC
proc_9_10_sse2(convolution_t *ch, uint8_t *buff, int bstride, int width,
int height, int stride, uint8_t *d, const uint8_t *s)
{
const uint16_t *srcp = (uint16_t *)s;
uint16_t *dstp = (uint16_t *)d;
stride /= 2;
bstride /= 2;
uint16_t *p0 = (uint16_t *)buff + 8;
uint16_t *p1 = p0 + bstride;
uint16_t *p2 = p1 + bstride;
uint16_t *p3 = p2 + bstride;
uint16_t *p4 = p3 + bstride;
uint16_t *orig = p0, *end = p4;
line_copy16(p0, srcp + 2 * stride, width, 2);
line_copy16(p1, srcp + stride, width, 2);
line_copy16(p2, srcp, width, 2);
srcp += stride;
line_copy16(p3, srcp, width, 2);
__m128i zero = _mm_setzero_si128();
__m128 rdiv = _mm_set1_ps((float)ch->rdiv);
__m128 bias = _mm_set1_ps((float)ch->bias);
__m128i matrix[25];
for (int i = 0; i < 25; i++) {
matrix[i] = _mm_unpacklo_epi16(_mm_set1_epi16((int16_t)ch->m[i]), zero);
}
for (int y = 0; y < height; y++) {
srcp += stride * (y < height - 2 ? 1 : -1);
line_copy16(p4, srcp, width, 2);
uint16_t *array[] = {
p0 - 2, p0 - 1, p0, p0 + 1, p0 + 2,
p1 - 2, p1 - 1, p1, p1 + 1, p1 + 2,
p2 - 2, p2 - 1, p2, p2 + 1, p2 + 2,
p3 - 2, p3 - 1, p3, p3 + 1, p3 + 2,
p4 - 2, p4 - 1, p4, p4 + 1, p4 + 2
};
for (int x = 0; x < width; x += 8) {
__m128i sum[2] = { zero, zero };
for (int i = 0; i < 25; i++) {
__m128i xmm0, xmm1;
xmm0 = _mm_loadu_si128((__m128i *)(array[i] + x));
xmm1 = _mm_unpackhi_epi16(xmm0, zero);
xmm0 = _mm_unpacklo_epi16(xmm0, zero);
sum[0] = _mm_add_epi32(sum[0], _mm_madd_epi16(xmm0, matrix[i]));
sum[1] = _mm_add_epi32(sum[1], _mm_madd_epi16(xmm1, matrix[i]));
}
for (int i = 0; i < 2; i++) {
__m128 sumfp = _mm_cvtepi32_ps(sum[i]);
sumfp = _mm_mul_ps(sumfp, rdiv);
sumfp = _mm_add_ps(sumfp, bias);
if (!ch->saturate) {
sumfp = mm_abs_ps(sumfp);
}
sum[i] = _mm_cvttps_epi32(sumfp);
}
sum[0] = _mm_packs_epi32(sum[0], sum[1]);
__m128i mask = _mm_cmpgt_epi16(sum[0], zero);
sum[0] = _mm_and_si128(sum[0], mask);
_mm_store_si128((__m128i *)(dstp + x), sum[0]);
}
dstp += stride;
p0 = p1;
p1 = p2;
p2 = p3;
p3 = p4;
p4 = (p4 == end) ? orig : p4 + bstride;
}
}
static void
thresh_16s( const Mat& _src, Mat& _dst, short thresh, short maxval, int type )
{
int i, j;
Size roi = _src.size();
roi.width *= _src.channels();
const short* src = (const short*)_src.data;
short* dst = (short*)_dst.data;
size_t src_step = _src.step/sizeof(src[0]);
size_t dst_step = _dst.step/sizeof(dst[0]);
#if CV_SSE2
volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE);
#endif
if( _src.isContinuous() && _dst.isContinuous() )
{
roi.width *= roi.height;
roi.height = 1;
src_step = dst_step = roi.width;
}
#ifdef HAVE_TEGRA_OPTIMIZATION
if (tegra::thresh_16s(_src, _dst, roi.width, roi.height, thresh, maxval, type))
return;
#endif
#if defined(HAVE_IPP)
IppiSize sz = { roi.width, roi.height };
switch( type )
{
case THRESH_TRUNC:
if (0 <= ippiThreshold_GT_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh))
return;
setIppErrorStatus();
break;
case THRESH_TOZERO:
if (0 <= ippiThreshold_LTVal_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh+1, 0))
return;
setIppErrorStatus();
break;
case THRESH_TOZERO_INV:
if (0 <= ippiThreshold_GTVal_16s_C1R(src, (int)src_step*sizeof(src[0]), dst, (int)dst_step*sizeof(dst[0]), sz, thresh, 0))
return;
setIppErrorStatus();
break;
}
#endif
switch( type )
{
case THRESH_BINARY:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh), maxval8 = _mm_set1_epi16(maxval);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_cmpgt_epi16( v0, thresh8 );
v1 = _mm_cmpgt_epi16( v1, thresh8 );
v0 = _mm_and_si128( v0, maxval8 );
v1 = _mm_and_si128( v1, maxval8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#endif
for( ; j < roi.width; j++ )
dst[j] = src[j] > thresh ? maxval : 0;
}
break;
case THRESH_BINARY_INV:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh), maxval8 = _mm_set1_epi16(maxval);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_cmpgt_epi16( v0, thresh8 );
v1 = _mm_cmpgt_epi16( v1, thresh8 );
v0 = _mm_andnot_si128( v0, maxval8 );
v1 = _mm_andnot_si128( v1, maxval8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#endif
for( ; j < roi.width; j++ )
dst[j] = src[j] <= thresh ? maxval : 0;
}
break;
case THRESH_TRUNC:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_min_epi16( v0, thresh8 );
v1 = _mm_min_epi16( v1, thresh8 );
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#endif
for( ; j < roi.width; j++ )
dst[j] = std::min(src[j], thresh);
}
break;
case THRESH_TOZERO:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_and_si128(v0, _mm_cmpgt_epi16(v0, thresh8));
v1 = _mm_and_si128(v1, _mm_cmpgt_epi16(v1, thresh8));
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#endif
for( ; j < roi.width; j++ )
{
short v = src[j];
dst[j] = v > thresh ? v : 0;
}
}
break;
case THRESH_TOZERO_INV:
for( i = 0; i < roi.height; i++, src += src_step, dst += dst_step )
{
j = 0;
#if CV_SSE2
if( useSIMD )
{
__m128i thresh8 = _mm_set1_epi16(thresh);
for( ; j <= roi.width - 16; j += 16 )
{
__m128i v0, v1;
v0 = _mm_loadu_si128( (const __m128i*)(src + j) );
v1 = _mm_loadu_si128( (const __m128i*)(src + j + 8) );
v0 = _mm_andnot_si128(_mm_cmpgt_epi16(v0, thresh8), v0);
v1 = _mm_andnot_si128(_mm_cmpgt_epi16(v1, thresh8), v1);
_mm_storeu_si128((__m128i*)(dst + j), v0 );
_mm_storeu_si128((__m128i*)(dst + j + 8), v1 );
}
}
#endif
for( ; j < roi.width; j++ )
{
short v = src[j];
dst[j] = v <= thresh ? v : 0;
}
}
break;
default:
return CV_Error( CV_StsBadArg, "" );
}
}
// Simple quantization
static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
int n, const VP8Matrix* const mtx) {
const __m128i max_coeff_2047 = _mm_set1_epi16(2047);
const __m128i zero = _mm_set1_epi16(0);
__m128i sign0, sign8;
__m128i coeff0, coeff8;
__m128i out0, out8;
__m128i packed_out;
// Load all inputs.
// TODO(cduvivier): Make variable declarations and allocations aligned so that
// we can use _mm_load_si128 instead of _mm_loadu_si128.
__m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
__m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
const __m128i sharpen0 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[0]);
const __m128i sharpen8 = _mm_loadu_si128((__m128i*)&mtx->sharpen_[8]);
const __m128i iq0 = _mm_loadu_si128((__m128i*)&mtx->iq_[0]);
const __m128i iq8 = _mm_loadu_si128((__m128i*)&mtx->iq_[8]);
const __m128i bias0 = _mm_loadu_si128((__m128i*)&mtx->bias_[0]);
const __m128i bias8 = _mm_loadu_si128((__m128i*)&mtx->bias_[8]);
const __m128i q0 = _mm_loadu_si128((__m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((__m128i*)&mtx->q_[8]);
const __m128i zthresh0 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[0]);
const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
// sign(in) = in >> 15 (0x0000 if positive, 0xffff if negative)
sign0 = _mm_srai_epi16(in0, 15);
sign8 = _mm_srai_epi16(in8, 15);
// coeff = abs(in) = (in ^ sign) - sign
coeff0 = _mm_xor_si128(in0, sign0);
coeff8 = _mm_xor_si128(in8, sign8);
coeff0 = _mm_sub_epi16(coeff0, sign0);
coeff8 = _mm_sub_epi16(coeff8, sign8);
// coeff = abs(in) + sharpen
coeff0 = _mm_add_epi16(coeff0, sharpen0);
coeff8 = _mm_add_epi16(coeff8, sharpen8);
// if (coeff > 2047) coeff = 2047
coeff0 = _mm_min_epi16(coeff0, max_coeff_2047);
coeff8 = _mm_min_epi16(coeff8, max_coeff_2047);
// out = (coeff * iQ + B) >> QFIX;
{
// doing calculations with 32b precision (QFIX=17)
// out = (coeff * iQ)
__m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
__m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
__m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
__m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
__m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
__m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
// expand bias from 16b to 32b
__m128i bias_00 = _mm_unpacklo_epi16(bias0, zero);
__m128i bias_04 = _mm_unpackhi_epi16(bias0, zero);
__m128i bias_08 = _mm_unpacklo_epi16(bias8, zero);
__m128i bias_12 = _mm_unpackhi_epi16(bias8, zero);
// out = (coeff * iQ + B)
out_00 = _mm_add_epi32(out_00, bias_00);
out_04 = _mm_add_epi32(out_04, bias_04);
out_08 = _mm_add_epi32(out_08, bias_08);
out_12 = _mm_add_epi32(out_12, bias_12);
// out = (coeff * iQ + B) >> QFIX;
out_00 = _mm_srai_epi32(out_00, QFIX);
out_04 = _mm_srai_epi32(out_04, QFIX);
out_08 = _mm_srai_epi32(out_08, QFIX);
out_12 = _mm_srai_epi32(out_12, QFIX);
// pack result as 16b
out0 = _mm_packs_epi32(out_00, out_04);
out8 = _mm_packs_epi32(out_08, out_12);
}
// get sign back (if (sign[j]) out_n = -out_n)
out0 = _mm_xor_si128(out0, sign0);
out8 = _mm_xor_si128(out8, sign8);
out0 = _mm_sub_epi16(out0, sign0);
out8 = _mm_sub_epi16(out8, sign8);
// in = out * Q
in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8);
// if (coeff <= mtx->zthresh_) {in=0; out=0;}
{
__m128i cmp0 = _mm_cmpgt_epi16(coeff0, zthresh0);
__m128i cmp8 = _mm_cmpgt_epi16(coeff8, zthresh8);
in0 = _mm_and_si128(in0, cmp0);
in8 = _mm_and_si128(in8, cmp8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
out0 = _mm_and_si128(out0, cmp0);
out8 = _mm_and_si128(out8, cmp8);
}
// zigzag the output before storing it.
//
// The zigzag pattern can almost be reproduced with a small sequence of
// shuffles. After it, we only need to swap the 7th (ending up in third
// position instead of twelfth) and 8th values.
{
__m128i outZ0, outZ8;
outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0));
outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1));
outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
_mm_storeu_si128((__m128i*)&out[0], outZ0);
_mm_storeu_si128((__m128i*)&out[8], outZ8);
packed_out = _mm_packs_epi16(outZ0, outZ8);
}
{
const int16_t outZ_12 = out[12];
const int16_t outZ_3 = out[3];
out[3] = outZ_12;
out[12] = outZ_3;
}
// detect if all 'out' values are zeroes or not
{
int32_t tmp[4];
_mm_storeu_si128((__m128i*)tmp, packed_out);
if (n) {
tmp[0] &= ~0xff;
}
return (tmp[3] || tmp[2] || tmp[1] || tmp[0]);
}
}
void EmitColorIndices_Intrinsics( const byte *colorBlock, const byte *minColor, const byte *maxColor, byte *&outData )
{
ALIGN16( byte color0[16] );
ALIGN16( byte color1[16] );
ALIGN16( byte color2[16] );
ALIGN16( byte color3[16] );
ALIGN16( byte result[16] );
// mov esi, maxColor
// mov edi, minColor
__m128i t0, t1, t2, t3, t4, t5, t6, t7;
t7 = _mm_setzero_si128();
//t7 = _mm_xor_si128(t7, t7);
_mm_store_si128 ( (__m128i*) &result, t7 );
//t0 = _mm_load_si128 ( (__m128i*) maxColor );
t0 = _mm_cvtsi32_si128( *(int*)maxColor);
// Bitwise AND
__m128i tt = _mm_load_si128 ( (__m128i*) SIMD_SSE2_byte_colorMask );
t0 = _mm_and_si128(t0, tt);
t0 = _mm_unpacklo_epi8(t0, t7);
t4 = _mm_shufflelo_epi16( t0, R_SHUFFLE_D( 0, 3, 2, 3 ));
t5 = _mm_shufflelo_epi16( t0, R_SHUFFLE_D( 3, 1, 3, 3 ));
t4 = _mm_srli_epi16(t4, 5);
t5 = _mm_srli_epi16(t5, 6);
// Bitwise Logical OR
t0 = _mm_or_si128(t0, t4);
t0 = _mm_or_si128(t0, t5); // t0 contains color0 in 565
//t1 = _mm_load_si128 ( (__m128i*) minColor );
t1 = _mm_cvtsi32_si128( *(int*)minColor);
t1 = _mm_and_si128(t1, tt);
t1 = _mm_unpacklo_epi8(t1, t7);
t4 = _mm_shufflelo_epi16( t1, R_SHUFFLE_D( 0, 3, 2, 3 ));
t5 = _mm_shufflelo_epi16( t1, R_SHUFFLE_D( 3, 1, 3, 3 ));
t4 = _mm_srli_epi16(t4, 5);
t5 = _mm_srli_epi16(t5, 6);
t1 = _mm_or_si128(t1, t4);
t1 = _mm_or_si128(t1, t5); // t1 contains color1 in 565
t2 = t0;
t2 = _mm_packus_epi16(t2, t7);
t2 = _mm_shuffle_epi32( t2, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color0, t2 );
t6 = t0;
t6 = _mm_add_epi16(t6, t0);
t6 = _mm_add_epi16(t6, t1);
// Multiply Packed Signed Integers and Store High Result
__m128i tw3 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_div_by_3 );
t6 = _mm_mulhi_epi16(t6, tw3);
t6 = _mm_packus_epi16(t6, t7);
t6 = _mm_shuffle_epi32( t6, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color2, t6 );
t3 = t1;
t3 = _mm_packus_epi16(t3, t7);
t3 = _mm_shuffle_epi32( t3, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color1, t3 );
t1 = _mm_add_epi16(t1, t1);
t0 = _mm_add_epi16(t0, t1);
t0 = _mm_mulhi_epi16(t0, tw3);
t0 = _mm_packus_epi16(t0, t7);
t0 = _mm_shuffle_epi32( t0, R_SHUFFLE_D( 0, 1, 0, 1 ));
_mm_store_si128 ( (__m128i*) &color3, t0 );
__m128i w0 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_0);
__m128i w1 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_1);
__m128i w2 = _mm_load_si128 ( (__m128i*) SIMD_SSE2_word_2);
// mov eax, 32
// mov esi, colorBlock
int x = 32;
//const byte *c = colorBlock;
while (x >= 0)
{
t3 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+0));
t3 = _mm_shuffle_epi32( t3, R_SHUFFLE_D( 0, 2, 1, 3 ));
t5 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+8));
t5 = _mm_shuffle_epi32( t5, R_SHUFFLE_D( 0, 2, 1, 3 ));
t0 = t3;
t6 = t5;
// Compute Sum of Absolute Difference
__m128i c0 = _mm_load_si128 ( (__m128i*) color0 );
t0 = _mm_sad_epu8(t0, c0);
t6 = _mm_sad_epu8(t6, c0);
// Pack with Signed Saturation
t0 = _mm_packs_epi32 (t0, t6);
t1 = t3;
t6 = t5;
__m128i c1 = _mm_load_si128 ( (__m128i*) color1 );
t1 = _mm_sad_epu8(t1, c1);
t6 = _mm_sad_epu8(t6, c1);
t1 = _mm_packs_epi32 (t1, t6);
t2 = t3;
t6 = t5;
__m128i c2 = _mm_load_si128 ( (__m128i*) color2 );
t2 = _mm_sad_epu8(t2, c2);
t6 = _mm_sad_epu8(t6, c2);
t2 = _mm_packs_epi32 (t2, t6);
__m128i c3 = _mm_load_si128 ( (__m128i*) color3 );
t3 = _mm_sad_epu8(t3, c3);
t5 = _mm_sad_epu8(t5, c3);
t3 = _mm_packs_epi32 (t3, t5);
t4 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+16));
t4 = _mm_shuffle_epi32( t4, R_SHUFFLE_D( 0, 2, 1, 3 ));
t5 = _mm_loadl_epi64( (__m128i*) (colorBlock+x+24));
t5 = _mm_shuffle_epi32( t5, R_SHUFFLE_D( 0, 2, 1, 3 ));
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c0);
t7 = _mm_sad_epu8(t7, c0);
t6 = _mm_packs_epi32 (t6, t7);
t0 = _mm_packs_epi32 (t0, t6); // d0
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c1);
t7 = _mm_sad_epu8(t7, c1);
t6 = _mm_packs_epi32 (t6, t7);
t1 = _mm_packs_epi32 (t1, t6); // d1
t6 = t4;
t7 = t5;
t6 = _mm_sad_epu8(t6, c2);
t7 = _mm_sad_epu8(t7, c2);
t6 = _mm_packs_epi32 (t6, t7);
t2 = _mm_packs_epi32 (t2, t6); // d2
t4 = _mm_sad_epu8(t4, c3);
t5 = _mm_sad_epu8(t5, c3);
t4 = _mm_packs_epi32 (t4, t5);
t3 = _mm_packs_epi32 (t3, t4); // d3
t7 = _mm_load_si128 ( (__m128i*) result );
t7 = _mm_slli_epi32( t7, 16);
t4 = t0;
t5 = t1;
// Compare Packed Signed Integers for Greater Than
t0 = _mm_cmpgt_epi16(t0, t3); // b0
t1 = _mm_cmpgt_epi16(t1, t2); // b1
t4 = _mm_cmpgt_epi16(t4, t2); // b2
t5 = _mm_cmpgt_epi16(t5, t3); // b3
t2 = _mm_cmpgt_epi16(t2, t3); // b4
t4 = _mm_and_si128(t4, t1); // x0
t5 = _mm_and_si128(t5, t0); // x1
t2 = _mm_and_si128(t2, t0); // x2
t4 = _mm_or_si128(t4, t5);
t2 = _mm_and_si128(t2, w1);
t4 = _mm_and_si128(t4, w2);
t2 = _mm_or_si128(t2, t4);
t5 = _mm_shuffle_epi32( t2, R_SHUFFLE_D( 2, 3, 0, 1 ));
// Unpack Low Data
t2 = _mm_unpacklo_epi16 ( t2, w0);
t5 = _mm_unpacklo_epi16 ( t5, w0);
//t5 = _mm_slli_si128 ( t5, 8);
t5 = _mm_slli_epi32( t5, 8);
t7 = _mm_or_si128(t7, t5);
t7 = _mm_or_si128(t7, t2);
_mm_store_si128 ( (__m128i*) &result, t7 );
x -=32;
}
t4 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 1, 2, 3, 0 ));
t5 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 2, 3, 0, 1 ));
t6 = _mm_shuffle_epi32( t7, R_SHUFFLE_D( 3, 0, 1, 2 ));
t4 = _mm_slli_epi32 ( t4, 2);
t5 = _mm_slli_epi32 ( t5, 4);
t6 = _mm_slli_epi32 ( t6, 6);
t7 = _mm_or_si128(t7, t4);
t7 = _mm_or_si128(t7, t5);
t7 = _mm_or_si128(t7, t6);
//_mm_store_si128 ( (__m128i*) outData, t7 );
int r = _mm_cvtsi128_si32 (t7);
memcpy(outData, &r, 4); // Anything better ?
outData += 4;
}
__m128i test_mm_cmpgt_epi16(__m128i A, __m128i B) {
// CHECK-LABEL: test_mm_cmpgt_epi16
// CHECK: icmp sgt <8 x i16>
return _mm_cmpgt_epi16(A, B);
}
void vpx_quantize_b_sse2(const tran_low_t* coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t* zbin_ptr,
const int16_t* round_ptr, const int16_t* quant_ptr,
const int16_t* quant_shift_ptr, tran_low_t* qcoeff_ptr,
tran_low_t* dqcoeff_ptr, const int16_t* dequant_ptr,
uint16_t* eob_ptr, const int16_t* scan_ptr,
const int16_t* iscan_ptr) {
__m128i zero;
(void)scan_ptr;
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
qcoeff_ptr += n_coeffs;
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
if (!skip_block) {
__m128i eob;
__m128i zbin;
__m128i round, quant, dequant, shift;
{
__m128i coeff0, coeff1;
// Setup global values
{
__m128i pw_1;
zbin = _mm_load_si128((const __m128i*)zbin_ptr);
round = _mm_load_si128((const __m128i*)round_ptr);
quant = _mm_load_si128((const __m128i*)quant_ptr);
pw_1 = _mm_set1_epi16(1);
zbin = _mm_sub_epi16(zbin, pw_1);
dequant = _mm_load_si128((const __m128i*)dequant_ptr);
shift = _mm_load_si128((const __m128i*)quant_shift_ptr);
}
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
// Do DC and first 15 AC
coeff0 = load_coefficients(coeff_ptr + n_coeffs);
coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
shift = _mm_unpackhi_epi64(shift, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
// Mask out zbin threshold coeffs
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
}
// AC only loop
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
__m128i cmp_mask0, cmp_mask1;
coeff0 = load_coefficients(coeff_ptr + n_coeffs);
coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
// Mask out zbin threshold coeffs
qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob0, eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob0 = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob0 = _mm_max_epi16(eob0, eob1);
eob = _mm_max_epi16(eob, eob0);
}
n_coeffs += 8 * 2;
}
// Accumulate EOB
{
__m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
eob = _mm_max_epi16(eob, eob_shuffled);
*eob_ptr = _mm_extract_epi16(eob, 1);
}
} else {
do {
store_coefficients(zero, dqcoeff_ptr + n_coeffs);
store_coefficients(zero, dqcoeff_ptr + n_coeffs + 8);
store_coefficients(zero, qcoeff_ptr + n_coeffs);
store_coefficients(zero, qcoeff_ptr + n_coeffs + 8);
n_coeffs += 8 * 2;
} while (n_coeffs < 0);
*eob_ptr = 0;
}
}
mlib_status
__mlib_VectorSumAbsDiff_S16_Sat(
mlib_d64 *z,
const mlib_s16 *x,
const mlib_s16 *y,
mlib_s32 n)
{
if (n <= 0)
return (MLIB_FAILURE);
mlib_s32 i, nstep, ax, ay, n1, n2, n3, xval, sum = 0;
mlib_s16 *px = (mlib_s16 *)x, *py = (mlib_s16 *)y;
__m128i zero, xbuf, ybuf, zbuf32, zbuf64, xlo, xhi, mext;
zero = _mm_setzero_si128();
zbuf64 = zero;
nstep = 16 / sizeof (mlib_s16);
ax = (mlib_addr)x & 15;
ay = (mlib_addr)y & 15;
n1 = ((16 - ax) & 15) / sizeof (mlib_s16);
n2 = (n - n1) / nstep;
n3 = n - n1 - n2 * nstep;
if (n2 < 1) {
for (i = 0; i < n; i++) {
xval = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(xval);
}
*z = sum;
} else {
for (i = 0; i < n1; i++) {
xval = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(xval);
}
mlib_s32 nblock = n2 >> 12;
mlib_s32 tail = n2 & 4095;
mlib_s32 k;
if (ax == ay) {
for (k = 0; k < nblock; k++) {
zbuf32 = zero;
for (i = 0; i < 4096; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_load_si128((__m128i *)py);
mext = _mm_cmpgt_epi16(ybuf, xbuf);
xbuf = _mm_sub_epi16(xbuf, ybuf);
xbuf = _mm_xor_si128(xbuf, mext);
xbuf = _mm_sub_epi16(xbuf, mext);
xlo = _mm_unpacklo_epi16(xbuf, zero);
xhi = _mm_unpackhi_epi16(xbuf, zero);
zbuf32 = _mm_add_epi32(zbuf32, xlo);
zbuf32 = _mm_add_epi32(zbuf32, xhi);
px += nstep;
py += nstep;
}
xlo = _mm_unpacklo_epi32(zbuf32, zero);
xhi = _mm_unpackhi_epi32(zbuf32, zero);
zbuf64 = _mm_add_epi64(zbuf64, xlo);
zbuf64 = _mm_add_epi64(zbuf64, xhi);
}
zbuf32 = zero;
for (i = 0; i < tail; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_load_si128((__m128i *)py);
mext = _mm_cmpgt_epi16(ybuf, xbuf);
xbuf = _mm_sub_epi16(xbuf, ybuf);
xbuf = _mm_xor_si128(xbuf, mext);
xbuf = _mm_sub_epi16(xbuf, mext);
xlo = _mm_unpacklo_epi16(xbuf, zero);
xhi = _mm_unpackhi_epi16(xbuf, zero);
zbuf32 = _mm_add_epi32(zbuf32, xlo);
zbuf32 = _mm_add_epi32(zbuf32, xhi);
px += nstep;
py += nstep;
}
xlo = _mm_unpacklo_epi32(zbuf32, zero);
xhi = _mm_unpackhi_epi32(zbuf32, zero);
zbuf64 = _mm_add_epi64(zbuf64, xlo);
zbuf64 = _mm_add_epi64(zbuf64, xhi);
} else { /* not aligned */
for (k = 0; k < nblock; k++) {
zbuf32 = zero;
for (i = 0; i < 4096; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_loadu_si128((__m128i *)py);
mext = _mm_cmpgt_epi16(ybuf, xbuf);
xbuf = _mm_sub_epi16(xbuf, ybuf);
xbuf = _mm_xor_si128(xbuf, mext);
xbuf = _mm_sub_epi16(xbuf, mext);
xlo = _mm_unpacklo_epi16(xbuf, zero);
xhi = _mm_unpackhi_epi16(xbuf, zero);
zbuf32 = _mm_add_epi32(zbuf32, xlo);
zbuf32 = _mm_add_epi32(zbuf32, xhi);
px += nstep;
py += nstep;
}
xlo = _mm_unpacklo_epi32(zbuf32, zero);
xhi = _mm_unpackhi_epi32(zbuf32, zero);
zbuf64 = _mm_add_epi64(zbuf64, xlo);
zbuf64 = _mm_add_epi64(zbuf64, xhi);
}
zbuf32 = zero;
for (i = 0; i < tail; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_loadu_si128((__m128i *)py);
mext = _mm_cmpgt_epi16(ybuf, xbuf);
xbuf = _mm_sub_epi16(xbuf, ybuf);
xbuf = _mm_xor_si128(xbuf, mext);
xbuf = _mm_sub_epi16(xbuf, mext);
xlo = _mm_unpacklo_epi16(xbuf, zero);
xhi = _mm_unpackhi_epi16(xbuf, zero);
zbuf32 = _mm_add_epi32(zbuf32, xlo);
zbuf32 = _mm_add_epi32(zbuf32, xhi);
px += nstep;
py += nstep;
}
xlo = _mm_unpacklo_epi32(zbuf32, zero);
xhi = _mm_unpackhi_epi32(zbuf32, zero);
zbuf64 = _mm_add_epi64(zbuf64, xlo);
zbuf64 = _mm_add_epi64(zbuf64, xhi);
}
for (i = 0; i < n3; i++) {
xval = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(xval);
}
mlib_d64 dsum = sum;
long long pz[2];
_mm_storeu_si128((__m128i *)pz, zbuf64);
dsum += pz[0];
dsum += pz[1];
*z = dsum;
}
return (MLIB_SUCCESS);
}
static WEBP_INLINE int DoQuantizeBlock(int16_t in[16], int16_t out[16],
const uint16_t* const sharpen,
const VP8Matrix* const mtx) {
const __m128i max_coeff_2047 = _mm_set1_epi16(MAX_LEVEL);
const __m128i zero = _mm_setzero_si128();
__m128i coeff0, coeff8;
__m128i out0, out8;
__m128i packed_out;
// Load all inputs.
// TODO(cduvivier): Make variable declarations and allocations aligned so that
// we can use _mm_load_si128 instead of _mm_loadu_si128.
__m128i in0 = _mm_loadu_si128((__m128i*)&in[0]);
__m128i in8 = _mm_loadu_si128((__m128i*)&in[8]);
const __m128i iq0 = _mm_loadu_si128((const __m128i*)&mtx->iq_[0]);
const __m128i iq8 = _mm_loadu_si128((const __m128i*)&mtx->iq_[8]);
const __m128i q0 = _mm_loadu_si128((const __m128i*)&mtx->q_[0]);
const __m128i q8 = _mm_loadu_si128((const __m128i*)&mtx->q_[8]);
// extract sign(in) (0x0000 if positive, 0xffff if negative)
const __m128i sign0 = _mm_cmpgt_epi16(zero, in0);
const __m128i sign8 = _mm_cmpgt_epi16(zero, in8);
// coeff = abs(in) = (in ^ sign) - sign
coeff0 = _mm_xor_si128(in0, sign0);
coeff8 = _mm_xor_si128(in8, sign8);
coeff0 = _mm_sub_epi16(coeff0, sign0);
coeff8 = _mm_sub_epi16(coeff8, sign8);
// coeff = abs(in) + sharpen
if (sharpen != NULL) {
const __m128i sharpen0 = _mm_loadu_si128((const __m128i*)&sharpen[0]);
const __m128i sharpen8 = _mm_loadu_si128((const __m128i*)&sharpen[8]);
coeff0 = _mm_add_epi16(coeff0, sharpen0);
coeff8 = _mm_add_epi16(coeff8, sharpen8);
}
// out = (coeff * iQ + B) >> QFIX
{
// doing calculations with 32b precision (QFIX=17)
// out = (coeff * iQ)
const __m128i coeff_iQ0H = _mm_mulhi_epu16(coeff0, iq0);
const __m128i coeff_iQ0L = _mm_mullo_epi16(coeff0, iq0);
const __m128i coeff_iQ8H = _mm_mulhi_epu16(coeff8, iq8);
const __m128i coeff_iQ8L = _mm_mullo_epi16(coeff8, iq8);
__m128i out_00 = _mm_unpacklo_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_04 = _mm_unpackhi_epi16(coeff_iQ0L, coeff_iQ0H);
__m128i out_08 = _mm_unpacklo_epi16(coeff_iQ8L, coeff_iQ8H);
__m128i out_12 = _mm_unpackhi_epi16(coeff_iQ8L, coeff_iQ8H);
// out = (coeff * iQ + B)
const __m128i bias_00 = _mm_loadu_si128((const __m128i*)&mtx->bias_[0]);
const __m128i bias_04 = _mm_loadu_si128((const __m128i*)&mtx->bias_[4]);
const __m128i bias_08 = _mm_loadu_si128((const __m128i*)&mtx->bias_[8]);
const __m128i bias_12 = _mm_loadu_si128((const __m128i*)&mtx->bias_[12]);
out_00 = _mm_add_epi32(out_00, bias_00);
out_04 = _mm_add_epi32(out_04, bias_04);
out_08 = _mm_add_epi32(out_08, bias_08);
out_12 = _mm_add_epi32(out_12, bias_12);
// out = QUANTDIV(coeff, iQ, B, QFIX)
out_00 = _mm_srai_epi32(out_00, QFIX);
out_04 = _mm_srai_epi32(out_04, QFIX);
out_08 = _mm_srai_epi32(out_08, QFIX);
out_12 = _mm_srai_epi32(out_12, QFIX);
// pack result as 16b
out0 = _mm_packs_epi32(out_00, out_04);
out8 = _mm_packs_epi32(out_08, out_12);
// if (coeff > 2047) coeff = 2047
out0 = _mm_min_epi16(out0, max_coeff_2047);
out8 = _mm_min_epi16(out8, max_coeff_2047);
}
// get sign back (if (sign[j]) out_n = -out_n)
out0 = _mm_xor_si128(out0, sign0);
out8 = _mm_xor_si128(out8, sign8);
out0 = _mm_sub_epi16(out0, sign0);
out8 = _mm_sub_epi16(out8, sign8);
// in = out * Q
in0 = _mm_mullo_epi16(out0, q0);
in8 = _mm_mullo_epi16(out8, q8);
_mm_storeu_si128((__m128i*)&in[0], in0);
_mm_storeu_si128((__m128i*)&in[8], in8);
// zigzag the output before storing it.
//
// The zigzag pattern can almost be reproduced with a small sequence of
// shuffles. After it, we only need to swap the 7th (ending up in third
// position instead of twelfth) and 8th values.
{
__m128i outZ0, outZ8;
outZ0 = _mm_shufflehi_epi16(out0, _MM_SHUFFLE(2, 1, 3, 0));
outZ0 = _mm_shuffle_epi32 (outZ0, _MM_SHUFFLE(3, 1, 2, 0));
outZ0 = _mm_shufflehi_epi16(outZ0, _MM_SHUFFLE(3, 1, 0, 2));
outZ8 = _mm_shufflelo_epi16(out8, _MM_SHUFFLE(3, 0, 2, 1));
outZ8 = _mm_shuffle_epi32 (outZ8, _MM_SHUFFLE(3, 1, 2, 0));
outZ8 = _mm_shufflelo_epi16(outZ8, _MM_SHUFFLE(1, 3, 2, 0));
_mm_storeu_si128((__m128i*)&out[0], outZ0);
_mm_storeu_si128((__m128i*)&out[8], outZ8);
packed_out = _mm_packs_epi16(outZ0, outZ8);
}
{
const int16_t outZ_12 = out[12];
const int16_t outZ_3 = out[3];
out[3] = outZ_12;
out[12] = outZ_3;
}
// detect if all 'out' values are zeroes or not
return (_mm_movemask_epi8(_mm_cmpeq_epi8(packed_out, zero)) != 0xffff);
}
/*
********************************************************************************
*
* @brief This function performs a 4x4 inverse hadamard transform on the 4x4 DC coefficients
* of a 16x16 intra prediction macroblock, and then performs scaling.
* prediction buffer
*
* @par Description:
* The DC coefficients pass through a 2-stage inverse hadamard transform.
* This inverse transformed content is scaled to based on Qp value.
*
* @param[in] pi2_src
* input 4x4 block of DC coefficients
*
* @param[out] pi2_out
* output 4x4 block
*
* @param[in] pu2_iscal_mat
* pointer to scaling list
*
* @param[in] pu2_weigh_mat
* pointer to weight matrix
*
* @param[in] u4_qp_div_6
* Floor (qp/6)
*
* @param[in] pi4_tmp
* temporary buffer of size 1*16
*
* @returns none
*
* @remarks none
*
*******************************************************************************
*/
void ih264_ihadamard_scaling_4x4_ssse3(WORD16* pi2_src,
WORD16* pi2_out,
const UWORD16 *pu2_iscal_mat,
const UWORD16 *pu2_weigh_mat,
UWORD32 u4_qp_div_6,
WORD32* pi4_tmp)
{
int val = 0xFFFF;
__m128i src_r0_r1, src_r2_r3, sign_reg, zero_8x16b = _mm_setzero_si128();
__m128i src_r0, src_r1, src_r2, src_r3;
__m128i temp0, temp1, temp2, temp3;
__m128i add_rshift = _mm_set1_epi32((1 << (5 - u4_qp_div_6)));
__m128i mult_val = _mm_set1_epi32(pu2_iscal_mat[0] * pu2_weigh_mat[0]);
__m128i mask = _mm_set1_epi32(val);
UNUSED (pi4_tmp);
mult_val = _mm_and_si128(mult_val, mask);
src_r0_r1 = _mm_loadu_si128((__m128i *) (pi2_src)); //a00 a01 a02 a03 a10 a11 a12 a13 -- the source matrix 0th,1st row
src_r2_r3 = _mm_loadu_si128((__m128i *) (pi2_src + 8)); //a20 a21 a22 a23 a30 a31 a32 a33 -- the source matrix 2nd,3rd row
sign_reg = _mm_cmpgt_epi16(zero_8x16b, src_r0_r1);
src_r0 = _mm_unpacklo_epi16(src_r0_r1, sign_reg);
src_r1 = _mm_unpackhi_epi16(src_r0_r1, sign_reg);
sign_reg = _mm_cmpgt_epi16(zero_8x16b, src_r2_r3);
src_r2 = _mm_unpacklo_epi16(src_r2_r3, sign_reg);
src_r3 = _mm_unpackhi_epi16(src_r2_r3, sign_reg);
/* Perform Inverse transform */
/*-------------------------------------------------------------*/
/* IDCT [ Horizontal transformation ] */
/*-------------------------------------------------------------*/
// Matrix transpose
/*
* a0 a1 a2 a3
* b0 b1 b2 b3
* c0 c1 c2 c3
* d0 d1 d2 d3
*/
temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 b0 a1 b1
temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //c0 d0 c1 d1
temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3
temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3
src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 b0 c0 d0
src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1
src_r2 = _mm_unpacklo_epi64(temp1, temp3); //a2 b2 c2 d2
src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3
temp0 = _mm_add_epi32(src_r0, src_r3);
temp1 = _mm_add_epi32(src_r1, src_r2);
temp2 = _mm_sub_epi32(src_r1, src_r2);
temp3 = _mm_sub_epi32(src_r0, src_r3);
src_r0 = _mm_add_epi32(temp0, temp1);
src_r1 = _mm_add_epi32(temp2, temp3);
src_r2 = _mm_sub_epi32(temp0, temp1);
src_r3 = _mm_sub_epi32(temp3, temp2);
/*-------------------------------------------------------------*/
/* IDCT [ Vertical transformation ] */
/*-------------------------------------------------------------*/
// Matrix transpose
/*
* a0 b0 c0 d0
* a1 b1 c1 d1
* a2 b2 c2 d2
* a3 b3 c3 d3
*/
temp0 = _mm_unpacklo_epi32(src_r0, src_r1); //a0 a1 b0 b1
temp2 = _mm_unpacklo_epi32(src_r2, src_r3); //a2 a3 b2 b3
temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1
temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3
src_r0 = _mm_unpacklo_epi64(temp0, temp2); //a0 a1 a2 a3
src_r1 = _mm_unpackhi_epi64(temp0, temp2); //b0 b1 b2 b3
src_r2 = _mm_unpacklo_epi64(temp1, temp3); //c0 c1 c2 c3
src_r3 = _mm_unpackhi_epi64(temp1, temp3); //d0 d1 d2 d3
temp0 = _mm_add_epi32(src_r0, src_r3);
temp1 = _mm_add_epi32(src_r1, src_r2);
temp2 = _mm_sub_epi32(src_r1, src_r2);
temp3 = _mm_sub_epi32(src_r0, src_r3);
src_r0 = _mm_add_epi32(temp0, temp1);
src_r1 = _mm_add_epi32(temp2, temp3);
src_r2 = _mm_sub_epi32(temp0, temp1);
src_r3 = _mm_sub_epi32(temp3, temp2);
src_r0 = _mm_and_si128(src_r0, mask);
src_r1 = _mm_and_si128(src_r1, mask);
src_r2 = _mm_and_si128(src_r2, mask);
src_r3 = _mm_and_si128(src_r3, mask);
src_r0 = _mm_madd_epi16(src_r0, mult_val);
src_r1 = _mm_madd_epi16(src_r1, mult_val);
src_r2 = _mm_madd_epi16(src_r2, mult_val);
src_r3 = _mm_madd_epi16(src_r3, mult_val);
//Scaling
if(u4_qp_div_6 >= 6)
{
src_r0 = _mm_slli_epi32(src_r0, u4_qp_div_6 - 6);
src_r1 = _mm_slli_epi32(src_r1, u4_qp_div_6 - 6);
src_r2 = _mm_slli_epi32(src_r2, u4_qp_div_6 - 6);
src_r3 = _mm_slli_epi32(src_r3, u4_qp_div_6 - 6);
}
else
{
temp0 = _mm_add_epi32(src_r0, add_rshift);
temp1 = _mm_add_epi32(src_r1, add_rshift);
temp2 = _mm_add_epi32(src_r2, add_rshift);
temp3 = _mm_add_epi32(src_r3, add_rshift);
src_r0 = _mm_srai_epi32(temp0, 6 - u4_qp_div_6);
src_r1 = _mm_srai_epi32(temp1, 6 - u4_qp_div_6);
src_r2 = _mm_srai_epi32(temp2, 6 - u4_qp_div_6);
src_r3 = _mm_srai_epi32(temp3, 6 - u4_qp_div_6);
}
src_r0_r1 = _mm_packs_epi32(src_r0, src_r1);
src_r2_r3 = _mm_packs_epi32(src_r2, src_r3);
_mm_storeu_si128((__m128i *) (&pi2_out[0]), src_r0_r1);
_mm_storeu_si128((__m128i *) (&pi2_out[8]), src_r2_r3);
}
static FORCE_INLINE void warp_mmword_u8_sse2(const uint8_t *srcp, const uint8_t *edgep, uint8_t *dstp, int src_stride, int edge_stride, int height, int x, int y, const __m128i &depth, const __m128i &zero, const __m128i &x_limit_min, const __m128i &x_limit_max, const __m128i &y_limit_min, const __m128i &y_limit_max, const __m128i &word_64, const __m128i &word_127, const __m128i &word_128, const __m128i &word_255, const __m128i &one_stride) {
int SMAG = 1 << SMAGL;
// calculate displacement
__m128i above = _mm_loadl_epi64((const __m128i *)(edgep + x - (y ? edge_stride : 0)));
__m128i below = _mm_loadl_epi64((const __m128i *)(edgep + x + (y < height - 1 ? edge_stride : 0)));
__m128i left = _mm_loadl_epi64((const __m128i *)(edgep + x - 1));
__m128i right = _mm_loadl_epi64((const __m128i *)(edgep + x + 1));
above = _mm_unpacklo_epi8(above, zero);
below = _mm_unpacklo_epi8(below, zero);
left = _mm_unpacklo_epi8(left, zero);
right = _mm_unpacklo_epi8(right, zero);
__m128i h = _mm_sub_epi16(left, right);
__m128i v = _mm_sub_epi16(above, below);
h = _mm_slli_epi16(h, 7);
v = _mm_slli_epi16(v, 7);
h = _mm_mulhi_epi16(h, depth);
v = _mm_mulhi_epi16(v, depth);
v = _mm_max_epi16(v, y_limit_min);
v = _mm_min_epi16(v, y_limit_max);
__m128i remainder_h = h;
__m128i remainder_v = v;
if (SMAGL) {
remainder_h = _mm_slli_epi16(remainder_h, SMAGL);
remainder_v = _mm_slli_epi16(remainder_v, SMAGL);
}
remainder_h = _mm_and_si128(remainder_h, word_127);
remainder_v = _mm_and_si128(remainder_v, word_127);
h = _mm_srai_epi16(h, 7 - SMAGL);
v = _mm_srai_epi16(v, 7 - SMAGL);
__m128i xx = _mm_set1_epi32(x << SMAGL);
xx = _mm_packs_epi32(xx, xx);
h = _mm_adds_epi16(h, xx);
remainder_h = _mm_and_si128(remainder_h, _mm_cmpgt_epi16(x_limit_max, h));
remainder_h = _mm_andnot_si128(_mm_cmpgt_epi16(x_limit_min, h), remainder_h);
h = _mm_max_epi16(h, x_limit_min);
h = _mm_min_epi16(h, x_limit_max);
// h and v contain the displacement now.
__m128i disp_lo = _mm_unpacklo_epi16(v, h);
__m128i disp_hi = _mm_unpackhi_epi16(v, h);
disp_lo = _mm_madd_epi16(disp_lo, one_stride);
disp_hi = _mm_madd_epi16(disp_hi, one_stride);
__m128i line0 = _mm_setzero_si128();
__m128i line1 = _mm_setzero_si128();
int offset = _mm_cvtsi128_si32(disp_lo);
disp_lo = _mm_srli_si128(disp_lo, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset), 0);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride), 0);
offset = _mm_cvtsi128_si32(disp_lo);
disp_lo = _mm_srli_si128(disp_lo, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 1 * SMAG), 1);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 1 * SMAG), 1);
offset = _mm_cvtsi128_si32(disp_lo);
disp_lo = _mm_srli_si128(disp_lo, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 2 * SMAG), 2);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 2 * SMAG), 2);
offset = _mm_cvtsi128_si32(disp_lo);
disp_lo = _mm_srli_si128(disp_lo, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 3 * SMAG), 3);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 3 * SMAG), 3);
offset = _mm_cvtsi128_si32(disp_hi);
disp_hi = _mm_srli_si128(disp_hi, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 4 * SMAG), 4);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 4 * SMAG), 4);
offset = _mm_cvtsi128_si32(disp_hi);
disp_hi = _mm_srli_si128(disp_hi, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 5 * SMAG), 5);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 5 * SMAG), 5);
offset = _mm_cvtsi128_si32(disp_hi);
disp_hi = _mm_srli_si128(disp_hi, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 6 * SMAG), 6);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 6 * SMAG), 6);
offset = _mm_cvtsi128_si32(disp_hi);
disp_hi = _mm_srli_si128(disp_hi, 4);
line0 = _mm_insert_epi16(line0, *(int16_t *)(srcp + offset + 7 * SMAG), 7);
line1 = _mm_insert_epi16(line1, *(int16_t *)(srcp + offset + src_stride + 7 * SMAG), 7);
__m128i left0 = _mm_and_si128(line0, word_255);
__m128i left1 = _mm_and_si128(line1, word_255);
__m128i right0 = _mm_srli_epi16(line0, 8);
__m128i right1 = _mm_srli_epi16(line1, 8);
left0 = _mm_mullo_epi16(left0, _mm_sub_epi16(word_128, remainder_h));
left1 = _mm_mullo_epi16(left1, _mm_sub_epi16(word_128, remainder_h));
right0 = _mm_mullo_epi16(right0, remainder_h);
right1 = _mm_mullo_epi16(right1, remainder_h);
line0 = _mm_add_epi16(left0, right0);
line1 = _mm_add_epi16(left1, right1);
line0 = _mm_add_epi16(line0, word_64);
line1 = _mm_add_epi16(line1, word_64);
line0 = _mm_srai_epi16(line0, 7);
line1 = _mm_srai_epi16(line1, 7);
line0 = _mm_mullo_epi16(line0, _mm_sub_epi16(word_128, remainder_v));
line1 = _mm_mullo_epi16(line1, remainder_v);
__m128i result = _mm_add_epi16(line0, line1);
result = _mm_add_epi16(result, word_64);
result = _mm_srai_epi16(result, 7);
result = _mm_packus_epi16(result, result);
_mm_storel_epi64((__m128i *)(dstp + x), result);
}
static alignment_end* sw_sse2_word (const int8_t* ref,
int8_t ref_dir, // 0: forward ref; 1: reverse ref
int32_t refLen,
int32_t readLen,
const uint8_t weight_gapO, /* will be used as - */
const uint8_t weight_gapE, /* will be used as - */
const __m128i* vProfile,
uint16_t terminate,
int32_t maskLen) {
#define max8(m, vm) (vm) = _mm_max_epi16((vm), _mm_srli_si128((vm), 8)); \
(vm) = _mm_max_epi16((vm), _mm_srli_si128((vm), 4)); \
(vm) = _mm_max_epi16((vm), _mm_srli_si128((vm), 2)); \
(m) = _mm_extract_epi16((vm), 0)
uint16_t max = 0; /* the max alignment score */
int32_t end_read = readLen - 1;
int32_t end_ref = 0; /* 1_based best alignment ending point; Initialized as isn't aligned - 0. */
int32_t segLen = (readLen + 7) / 8; /* number of segment */
/* array to record the largest score of each reference position */
uint16_t* maxColumn = (uint16_t*) calloc(refLen, 2);
/* array to record the alignment read ending position of the largest score of each reference position */
int32_t* end_read_column = (int32_t*) calloc(refLen, sizeof(int32_t));
/* Define 16 byte 0 vector. */
__m128i vZero = _mm_set1_epi32(0);
__m128i* pvHStore = (__m128i*) calloc(segLen, sizeof(__m128i));
__m128i* pvHLoad = (__m128i*) calloc(segLen, sizeof(__m128i));
__m128i* pvE = (__m128i*) calloc(segLen, sizeof(__m128i));
__m128i* pvHmax = (__m128i*) calloc(segLen, sizeof(__m128i));
int32_t i, j, k;
/* 16 byte insertion begin vector */
__m128i vGapO = _mm_set1_epi16(weight_gapO);
/* 16 byte insertion extension vector */
__m128i vGapE = _mm_set1_epi16(weight_gapE);
__m128i vMaxScore = vZero; /* Trace the highest score of the whole SW matrix. */
__m128i vMaxMark = vZero; /* Trace the highest score till the previous column. */
__m128i vTemp;
int32_t edge, begin = 0, end = refLen, step = 1;
/* outer loop to process the reference sequence */
if (ref_dir == 1) {
begin = refLen - 1;
end = -1;
step = -1;
}
for (i = begin; LIKELY(i != end); i += step) {
int32_t cmp;
__m128i e, vF = vZero; /* Initialize F value to 0.
Any errors to vH values will be corrected in the Lazy_F loop.
*/
__m128i vH = pvHStore[segLen - 1];
vH = _mm_slli_si128 (vH, 2); /* Shift the 128-bit value in vH left by 2 byte. */
/* Swap the 2 H buffers. */
__m128i* pv = pvHLoad;
__m128i vMaxColumn = vZero; /* vMaxColumn is used to record the max values of column i. */
const __m128i* vP = vProfile + ref[i] * segLen; /* Right part of the vProfile */
pvHLoad = pvHStore;
pvHStore = pv;
/* inner loop to process the query sequence */
for (j = 0; LIKELY(j < segLen); j ++) {
vH = _mm_adds_epi16(vH, _mm_load_si128(vP + j));
/* Get max from vH, vE and vF. */
e = _mm_load_si128(pvE + j);
vH = _mm_max_epi16(vH, e);
vH = _mm_max_epi16(vH, vF);
vMaxColumn = _mm_max_epi16(vMaxColumn, vH);
/* Save vH values. */
_mm_store_si128(pvHStore + j, vH);
/* Update vE value. */
vH = _mm_subs_epu16(vH, vGapO); /* saturation arithmetic, result >= 0 */
e = _mm_max_epi16(e, vH);
e = _mm_subs_epu16(e, vGapE);
_mm_store_si128(pvE + j, e);
/* Update vF value. */
vF = _mm_max_epi16(vF, vH);
vF = _mm_subs_epu16(vF, vGapE);
/* Load the next vH. */
vH = _mm_load_si128(pvHLoad + j);
}
/* Lazy_F loop: has been revised to disallow adjecent insertion and then deletion, so don't update E(i, j), learn from SWPS3 */
for (k = 0; LIKELY(k < 8); ++k) {
vF = _mm_slli_si128 (vF, 2);
for (j = 0; LIKELY(j < segLen); ++j) {
vH = _mm_load_si128(pvHStore + j);
vH = _mm_max_epi16(vH, vF);
_mm_store_si128(pvHStore + j, vH);
vH = _mm_subs_epu16(vH, vGapO);
vF = _mm_subs_epu16(vF, vGapE);
if (UNLIKELY(! _mm_movemask_epi8(_mm_cmpgt_epi16(vF, vH)))) goto end;
}
}
end:
vMaxScore = _mm_max_epi16(vMaxScore, vMaxColumn);
vTemp = _mm_cmpeq_epi16(vMaxMark, vMaxScore);
cmp = _mm_movemask_epi8(vTemp);
if (cmp != 0xffff) {
uint16_t temp;
vMaxMark = vMaxScore;
max8(temp, vMaxScore);
vMaxScore = vMaxMark;
if (LIKELY(temp > max)) {
max = temp;
end_ref = i;
for (j = 0; LIKELY(j < segLen); ++j) pvHmax[j] = pvHStore[j];
}
}
/* Record the max score of current column. */
max8(maxColumn[i], vMaxColumn);
if (maxColumn[i] == terminate) break;
}
/* Trace the alignment ending position on read. */
uint16_t *t = (uint16_t*)pvHmax;
int32_t column_len = segLen * 8;
for (i = 0; LIKELY(i < column_len); ++i, ++t) {
int32_t temp;
if (*t == max) {
temp = i / 8 + i % 8 * segLen;
if (temp < end_read) end_read = temp;
}
}
free(pvHmax);
free(pvE);
free(pvHLoad);
free(pvHStore);
/* Find the most possible 2nd best alignment. */
alignment_end* bests = (alignment_end*) calloc(2, sizeof(alignment_end));
bests[0].score = max;
bests[0].ref = end_ref;
bests[0].read = end_read;
bests[1].score = 0;
bests[1].ref = 0;
bests[1].read = 0;
edge = (end_ref - maskLen) > 0 ? (end_ref - maskLen) : 0;
for (i = 0; i < edge; i ++) {
if (maxColumn[i] > bests[1].score) {
bests[1].score = maxColumn[i];
bests[1].ref = i;
}
}
edge = (end_ref + maskLen) > refLen ? refLen : (end_ref + maskLen);
for (i = edge; i < refLen; i ++) {
if (maxColumn[i] > bests[1].score) {
bests[1].score = maxColumn[i];
bests[1].ref = i;
}
}
free(maxColumn);
free(end_read_column);
return bests;
}
int
smith_waterman_sse2_word(const unsigned char * query_sequence,
unsigned short * query_profile_word,
const int query_length,
const unsigned char * db_sequence,
const int db_length,
unsigned short gap_open,
unsigned short gap_extend,
struct f_struct * f_str)
{
int i, j, k;
short score;
int cmp;
int iter = (query_length + 7) / 8;
__m128i *p;
__m128i *workspace = (__m128i *) f_str->workspace;
__m128i E, F, H;
__m128i v_maxscore;
__m128i v_gapopen;
__m128i v_gapextend;
__m128i v_min;
__m128i v_minimums;
__m128i v_temp;
__m128i *pHLoad, *pHStore;
__m128i *pE;
__m128i *pScore;
/* Load gap opening penalty to all elements of a constant */
v_gapopen = _mm_setzero_si128(); /* Apple Devel */
v_gapopen = _mm_insert_epi16 (v_gapopen, gap_open, 0);
v_gapopen = _mm_shufflelo_epi16 (v_gapopen, 0);
v_gapopen = _mm_shuffle_epi32 (v_gapopen, 0);
/* Load gap extension penalty to all elements of a constant */
v_gapextend = _mm_setzero_si128(); /* Apple Devel */
v_gapextend = _mm_insert_epi16 (v_gapextend, gap_extend, 0);
v_gapextend = _mm_shufflelo_epi16 (v_gapextend, 0);
v_gapextend = _mm_shuffle_epi32 (v_gapextend, 0);
/* load v_maxscore with the zeros. since we are using signed */
/* math, we will bias the maxscore to -32768 so we have the */
/* full range of the short. */
v_maxscore = _mm_setzero_si128(); /* Apple Devel */
v_maxscore = _mm_cmpeq_epi16 (v_maxscore, v_maxscore);
v_maxscore = _mm_slli_epi16 (v_maxscore, 15);
v_minimums = _mm_shuffle_epi32 (v_maxscore, 0);
v_min = _mm_shuffle_epi32 (v_maxscore, 0);
v_min = _mm_srli_si128 (v_min, 14);
/* Zero out the storage vector */
k = 2 * iter;
p = workspace;
for (i = 0; i < k; i++)
{
_mm_store_si128 (p++, v_maxscore);
}
pE = workspace;
pHStore = pE + iter;
pHLoad = pHStore + iter;
for (i = 0; i < db_length; ++i)
{
/* fetch first data asap. */
pScore = (__m128i *) query_profile_word + db_sequence[i] * iter;
/* bias all elements in F to -32768 */
F = _mm_setzero_si128(); /* Apple Devel */
F = _mm_cmpeq_epi16 (F, F);
F = _mm_slli_epi16 (F, 15);
/* load the next h value */
H = _mm_load_si128 (pHStore + iter - 1);
H = _mm_slli_si128 (H, 2);
H = _mm_or_si128 (H, v_min);
p = pHLoad;
pHLoad = pHStore;
pHStore = p;
for (j = 0; j < iter; j++)
{
/* load E values */
E = _mm_load_si128 (pE + j);
/* add score to H */
H = _mm_adds_epi16 (H, *pScore++);
/* Update highest score encountered this far */
v_maxscore = _mm_max_epi16 (v_maxscore, H);
/* get max from H, E and F */
H = _mm_max_epi16 (H, E);
H = _mm_max_epi16 (H, F);
/* save H values */
_mm_store_si128 (pHStore + j, H);
/* subtract the gap open penalty from H */
H = _mm_subs_epi16 (H, v_gapopen);
/* update E value */
E = _mm_subs_epi16 (E, v_gapextend);
E = _mm_max_epi16 (E, H);
/* update F value */
F = _mm_subs_epi16 (F, v_gapextend);
F = _mm_max_epi16 (F, H);
/* save E values */
_mm_store_si128 (pE + j, E);
/* load the next h value */
H = _mm_load_si128 (pHLoad + j);
}
/* reset pointers to the start of the saved data */
j = 0;
H = _mm_load_si128 (pHStore + j);
/* the computed F value is for the given column. since */
/* we are at the end, we need to shift the F value over */
/* to the next column. */
F = _mm_slli_si128 (F, 2);
F = _mm_or_si128 (F, v_min);
v_temp = _mm_subs_epi16 (H, v_gapopen);
v_temp = _mm_cmpgt_epi16 (F, v_temp);
cmp = _mm_movemask_epi8 (v_temp);
while (cmp != 0x0000)
{
E = _mm_load_si128 (pE + j);
H = _mm_max_epi16 (H, F);
/* save H values */
_mm_store_si128 (pHStore + j, H);
/* update E in case the new H value would change it */
H = _mm_subs_epi16 (H, v_gapopen);
E = _mm_max_epi16 (E, H);
_mm_store_si128 (pE + j, E);
/* update F value */
F = _mm_subs_epi16 (F, v_gapextend);
j++;
if (j >= iter)
{
j = 0;
F = _mm_slli_si128 (F, 2);
F = _mm_or_si128 (F, v_min);
}
H = _mm_load_si128 (pHStore + j);
v_temp = _mm_subs_epi16 (H, v_gapopen);
v_temp = _mm_cmpgt_epi16 (F, v_temp);
cmp = _mm_movemask_epi8 (v_temp);
}
}
/* find largest score in the v_maxscore vector */
v_temp = _mm_srli_si128 (v_maxscore, 8);
v_maxscore = _mm_max_epi16 (v_maxscore, v_temp);
v_temp = _mm_srli_si128 (v_maxscore, 4);
v_maxscore = _mm_max_epi16 (v_maxscore, v_temp);
v_temp = _mm_srli_si128 (v_maxscore, 2);
v_maxscore = _mm_max_epi16 (v_maxscore, v_temp);
/* extract the largest score */
score = _mm_extract_epi16 (v_maxscore, 0);
/* return largest score biased by 32768 */
/* fix for Mac OSX clang 4.1 */
/*
#ifdef __clang__
if (score < 0) score += 32768;
return score;
#else
*/
return score + 32768;
/* #endif */
}
void vp9_quantize_fp_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) {
__m128i zero;
__m128i thr;
int16_t nzflag;
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
coeff_ptr += n_coeffs;
iscan_ptr += n_coeffs;
qcoeff_ptr += n_coeffs;
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
if (!skip_block) {
__m128i eob;
__m128i round, quant, dequant;
{
__m128i coeff0, coeff1;
// Setup global values
{
round = _mm_load_si128((const __m128i *)round_ptr);
quant = _mm_load_si128((const __m128i *)quant_ptr);
dequant = _mm_load_si128((const __m128i *)dequant_ptr);
}
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
// Do DC and first 15 AC
coeff0 = load_tran_low(coeff_ptr + n_coeffs);
coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
}
thr = _mm_srai_epi16(dequant, 1);
// AC only loop
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
coeff0 = load_tran_low(coeff_ptr + n_coeffs);
coeff1 = load_tran_low(coeff_ptr + n_coeffs + 8);
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
_mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
if (nzflag) {
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
} else {
store_zero_tran_low(qcoeff_ptr + n_coeffs);
store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
store_zero_tran_low(dqcoeff_ptr + n_coeffs);
store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
}
}
if (nzflag) {
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob0, eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob0 = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob0 = _mm_max_epi16(eob0, eob1);
eob = _mm_max_epi16(eob, eob0);
}
n_coeffs += 8 * 2;
}
// Accumulate EOB
{
__m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
eob = _mm_max_epi16(eob, eob_shuffled);
*eob_ptr = _mm_extract_epi16(eob, 1);
}
} else {
do {
store_zero_tran_low(qcoeff_ptr + n_coeffs);
store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
store_zero_tran_low(dqcoeff_ptr + n_coeffs);
store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
n_coeffs += 8 * 2;
} while (n_coeffs < 0);
*eob_ptr = 0;
}
}
static void GF_FUNC_ALIGN VS_CC
proc_9_10_sse2(convolution_hv_t *ch, uint8_t *buff, int bstride, int width,
int height, int stride, uint8_t *d, const uint8_t *s)
{
const uint16_t *srcp = (uint16_t *)s;
uint16_t *dstp = (uint16_t *)d;
stride /= 2;
bstride /= 2;
uint16_t *p0 = (uint16_t *)buff + 8;
uint16_t *p1 = p0 + bstride;
uint16_t *p2 = p1 + bstride;
uint16_t *p3 = p2 + bstride;
uint16_t *p4 = p3 + bstride;
uint16_t *orig = p0, *end = p4;
line_copy16(p0, srcp + 2 * stride, width, 2);
line_copy16(p1, srcp + stride, width, 2);
line_copy16(p2, srcp, width, 2);
srcp += stride;
line_copy16(p3, srcp, width, 2);
__m128i zero = _mm_setzero_si128();
__m128i all1 = _mm_cmpeq_epi32(zero, zero);
__m128i one = _mm_srli_epi16(all1, 15);
__m128 rdiv_h = _mm_set1_ps((float)ch->rdiv_h);
__m128 rdiv_v = _mm_set1_ps((float)ch->rdiv_v);
__m128 bias = _mm_set1_ps((float)ch->bias);
__m128i matrix_h[5];
__m128i matrix_v[5];
for (int i = 0; i < 5; i++) {
matrix_h[i] = _mm_unpacklo_epi16(_mm_set1_epi16((int16_t)ch->m_h[i]), zero);
matrix_v[i] = _mm_unpacklo_epi16(_mm_set1_epi16((int16_t)ch->m_v[i]), zero);
}
for (int y = 0; y < height; y++) {
srcp += stride * (y < height - 2 ? 1 : -1);
line_copy16(p4, srcp, width, 2);
for (int x = 0; x < width; x += 8) {
uint16_t *array[] = {
p0 + x, p1 + x, p2 + x, p3 + x, p4 + x,
p2 + x - 2, p2 + x - 1, dstp + x, p2 + x + 1, p2 + x + 2
};
for (int j = 0; j < 2; j++) {
__m128i *matrix = j == 0 ? matrix_v : matrix_h;
__m128i sum[2];
sum[0] = _mm_setzero_si128();
sum[1] = _mm_setzero_si128();
for (int i = 0; i < 5; i++) {
__m128i xmm0, xmm1;
xmm0 = _mm_loadu_si128((__m128i *)array[i + j * 5]);
xmm1 = _mm_unpackhi_epi16(xmm0, zero);
xmm0 = _mm_unpacklo_epi16(xmm0, zero);
sum[0] = _mm_add_epi32(sum[0], _mm_madd_epi16(xmm0, matrix[i]));
sum[1] = _mm_add_epi32(sum[1], _mm_madd_epi16(xmm1, matrix[i]));
}
for (int i = 0; i < 2; i++) {
__m128 sumfp = _mm_cvtepi32_ps(sum[i]);
sumfp = _mm_mul_ps(sumfp, j == 0 ? rdiv_v : rdiv_h);
if (j == 1) {
sumfp = _mm_add_ps(sumfp, bias);
}
sum[i] = _mm_cvttps_epi32(sumfp);
}
sum[0] = _mm_packs_epi32(sum[0], sum[1]);
__m128i mask = _mm_cmpgt_epi16(sum[0], zero);
if (ch->saturate) {
sum[0] = _mm_and_si128(sum[0], mask);
} else {
__m128i temp = _mm_add_epi16(one, _mm_xor_si128(sum[0], all1));
temp = _mm_andnot_si128(mask, temp);
sum[0] = _mm_and_si128(sum[0], mask);
sum[0] = _mm_or_si128(sum[0], temp);
}
_mm_store_si128((__m128i *)(dstp + x), sum[0]);
}
}
dstp += stride;
p0 = p1;
p1 = p2;
p2 = p3;
p3 = p4;
p4 = (p4 == end) ? orig : p4 + bstride;
}
}
int viterbi_stream_word_partitioned(DATA_STREAM* dstream, float* opt_res, int thrid)
{
// if (NTHREADS > 1) pthread_barrier_wait(&dstream->barrier);
return 0;
int L = dstream->L;
P7_PROFILE* gm = dstream->gm;
ESL_DSQ** ddsq = dstream->seqs;
int M = gm->M, i, k, v, t, j;
const int PARTITION = dstream->partition;
__m128i** oprmsc = (__m128i**) dstream->rsc_msc;
__m128i* xmxEv = dstream->xmxE;
__m128i xmxB, xmxE, xmxC, moveC, Vinf = _mm_set1_epi16(-WORDMAX);
__m128i dmx[PARTITION];
__m128i mmx[PARTITION];
__m128i imx[PARTITION];
__m128i xmm[24];
__m128i *mscore[8];
__m128i overflowlimit, overflows;
overflowlimit = overflows = Vinf;
if (thrid == NTHREADS-1)
{ overflowlimit = _mm_set1_epi16(WORDMAX-1);
overflows= _mm_xor_si128(overflows,overflows); // zero out
}
t = ((dstream->Npartitions+thrid)%NTHREADS)*PARTITION;
tprintf("START viterbiThr %d in %d L %d | Seq %d\n", thrid, t, L, 0); // ccount[thrid]++);
xmxC = Vinf;
moveC = _mm_set1_epi16(discretize(dstream->scale, gm->xsc[p7P_C][p7P_MOVE]));
xmxB = _mm_set1_epi16(dstream->wordoffset + discretize(dstream->scale, gm->xsc[p7P_N][p7P_MOVE]));
for ( ; t < M; t += NTHREADS*PARTITION)
{
volatile uchar* synchflags1 = dstream->synchflags[t/PARTITION];
volatile uchar* synchflags2 = dstream->synchflags[t/PARTITION+1];
int t8 = t/8;
for (k = 0; k < PARTITION; k++)
dmx[k] = mmx[k] = imx[k] = Vinf;
for (i = 1; i <= L; i++)
{
// tprintf("Iter Thr %d t %d: I %d\n", thrid, t, i);
__m128i sc, dcv, temp, mpv, ipv, dpv;
__m128i *ttsc = dstream->tsc_all + t*8;
v = i-1;
ttsc += 3;
if (t == 0)
xmxE = mpv = dpv = ipv = sc = dcv = Vinf;
else {
if (NTHREADS > 1)
while (!synchflags1[v]) sched_yield();
xmxE = xmxEv[v];
dcv = dstream->pdcv[v];
sc = dstream->psc[v];
}
for (j = 0; j < 8; j++)
mscore[j] = oprmsc[ddsq[j][i]] + t8;
for (k = 0; k < PARTITION && t+k < M; )
{
#if 0
#define EMLOAD(i) xmm[i+24] = _mm_load_si128(mscore[i]);
EMLOAD(0) EMLOAD(1)
EMLOAD(2) EMLOAD(3)
EMLOAD(4) EMLOAD(5)
EMLOAD(6) EMLOAD(7)
#define MIX16(i,r,range) \
xmm[r ] = _mm_unpacklo_epi##range(xmm[24+i], xmm[24+i+1]); \
xmm[r+1] = _mm_unpackhi_epi##range(xmm[24+i], xmm[24+i+1]);
MIX16(0,0,16) MIX16(2,2,16)
MIX16(4,4,16) MIX16(6,6,16)
#else
#define MMLOAD(a,b) \
xmm[a] = _mm_unpacklo_epi16(*mscore[a], *mscore[b]); \
xmm[b] = _mm_unpackhi_epi16(*mscore[a], *mscore[b]);
MMLOAD(0,1) MMLOAD(2,3)
MMLOAD(4,5) MMLOAD(6,7)
#endif
#define MIX(i,r,range) \
xmm[r ] = _mm_unpacklo_epi##range(xmm[i], xmm[i+2]); \
xmm[r+1] = _mm_unpackhi_epi##range(xmm[i], xmm[i+2]);
MIX(0, 8,32) MIX(1,12,32)
MIX(4,10,32) MIX(5,14,32)
MIX( 8,16,64) MIX( 9,18,64)
MIX(12,20,64) MIX(13,22,64)
#define TRIPLETCOMPUTE(k,j) \
{ /* Calculate new M(k), delay store */ \
sc = _mm_max_epi16(sc, _mm_adds_epi16(xmxB, *ttsc)); ttsc++; \
sc = _mm_adds_epi16(sc, xmm[j]); \
/* Update E */ \
xmxE = _mm_max_epi16(xmxE, sc); \
\
/* Pre-emptive load of M, D, I */ \
dpv = dmx[k]; \
ipv = imx[k]; \
mpv = mmx[k]; \
\
/* Calculate current I(k) */ \
temp = _mm_adds_epi16(mpv, *ttsc); ttsc++; \
imx[k] = _mm_max_epi16(temp, _mm_adds_epi16(ipv, *ttsc)); ttsc++;\
\
/* Delayed stores of M and D */ \
mmx[k] = sc; \
dmx[k] = dcv; \
\
/* Calculate next D, D(k+1) */ \
sc = _mm_adds_epi16(sc, *ttsc); ttsc++; \
dcv = _mm_max_epi16(sc, _mm_adds_epi16(dcv, *ttsc));ttsc++; \
\
/* Pre-emptive partial calculation of M(k+1) */ \
sc = _mm_adds_epi16(mpv, *ttsc); ttsc++; \
sc = _mm_max_epi16(sc, _mm_adds_epi16(ipv, *ttsc)); ttsc++; \
sc = _mm_max_epi16(sc, _mm_adds_epi16(dpv, *ttsc)); ttsc++; \
k++; \
}
TRIPLETCOMPUTE(k,16+0) TRIPLETCOMPUTE(k,16+1)
TRIPLETCOMPUTE(k,16+2) TRIPLETCOMPUTE(k,16+3)
TRIPLETCOMPUTE(k,16+4) TRIPLETCOMPUTE(k,16+5)
TRIPLETCOMPUTE(k,16+6) TRIPLETCOMPUTE(k,16+7)
mscore[0]++; mscore[1]++; mscore[2]++; mscore[3]++;
mscore[4]++; mscore[5]++; mscore[6]++; mscore[7]++;
}
if (t+k < M)
{ v = i-1;
xmxEv[v] = xmxE;
dstream->pdcv[v] = dcv;
dstream->psc [v] = sc;
if (NTHREADS > 1) synchflags2[v] = 1;
}
else // executed only by main thread (NTHRS-1)
{
__m128i overfs = _mm_cmpgt_epi16(xmxE, overflowlimit);
overflows = _mm_or_si128(overflows, overfs); // select the overflowed channels
xmxC = _mm_max_epi16(xmxC, xmxE);
}
}
}
xmxC = _mm_adds_epi16(xmxC, moveC);
if (opt_res != NULL)
{
float offset = (float) dstream->wordoffset;
int16_t res[8] __attribute__ ((aligned (16)));
int16_t ovs[8] __attribute__ ((aligned (16)));
memmove(res, &xmxC, sizeof(xmxC));
memmove(ovs, &overflows, sizeof(overflows));
for (i = 0; i < 8; i++)
if (ovs[i]) opt_res[i] = eslINFINITY; // signal overflow
else opt_res[i] = ((float) res[i] - offset) / dstream->scale - 2.0;
// 2.0 nat approximation, UNILOCAL mode
}
tprintf("END viterbi Thr %d - t %d\n", thrid, t);
return eslOK;
}
void vp9_fdct8x8_quant_ssse3(
const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs,
int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr,
const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
uint16_t *eob_ptr, const int16_t *scan_ptr, const int16_t *iscan_ptr) {
__m128i zero;
int pass;
// Constants
// When we use them, in one case, they are all the same. In all others
// it's a pair of them that we need to repeat four times. This is done
// by constructing the 32 bit constant corresponding to that pair.
const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170);
const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
// Load input
__m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
__m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
__m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
__m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
__m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
__m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
__m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
__m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
__m128i *in[8];
int index = 0;
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
(void)coeff_ptr;
// Pre-condition input (shift by two)
in0 = _mm_slli_epi16(in0, 2);
in1 = _mm_slli_epi16(in1, 2);
in2 = _mm_slli_epi16(in2, 2);
in3 = _mm_slli_epi16(in3, 2);
in4 = _mm_slli_epi16(in4, 2);
in5 = _mm_slli_epi16(in5, 2);
in6 = _mm_slli_epi16(in6, 2);
in7 = _mm_slli_epi16(in7, 2);
in[0] = &in0;
in[1] = &in1;
in[2] = &in2;
in[3] = &in3;
in[4] = &in4;
in[5] = &in5;
in[6] = &in6;
in[7] = &in7;
// We do two passes, first the columns, then the rows. The results of the
// first pass are transposed so that the same column code can be reused. The
// results of the second pass are also transposed so that the rows (processed
// as columns) are put back in row positions.
for (pass = 0; pass < 2; pass++) {
// To store results of each pass before the transpose.
__m128i res0, res1, res2, res3, res4, res5, res6, res7;
// Add/subtract
const __m128i q0 = _mm_add_epi16(in0, in7);
const __m128i q1 = _mm_add_epi16(in1, in6);
const __m128i q2 = _mm_add_epi16(in2, in5);
const __m128i q3 = _mm_add_epi16(in3, in4);
const __m128i q4 = _mm_sub_epi16(in3, in4);
const __m128i q5 = _mm_sub_epi16(in2, in5);
const __m128i q6 = _mm_sub_epi16(in1, in6);
const __m128i q7 = _mm_sub_epi16(in0, in7);
// Work on first four results
{
// Add/subtract
const __m128i r0 = _mm_add_epi16(q0, q3);
const __m128i r1 = _mm_add_epi16(q1, q2);
const __m128i r2 = _mm_sub_epi16(q1, q2);
const __m128i r3 = _mm_sub_epi16(q0, q3);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res0 = _mm_packs_epi32(w0, w1);
res4 = _mm_packs_epi32(w2, w3);
res2 = _mm_packs_epi32(w4, w5);
res6 = _mm_packs_epi32(w6, w7);
}
// Work on next four results
{
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i d0 = _mm_sub_epi16(q6, q5);
const __m128i d1 = _mm_add_epi16(q6, q5);
const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16);
const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16);
// Add/subtract
const __m128i x0 = _mm_add_epi16(q4, r0);
const __m128i x1 = _mm_sub_epi16(q4, r0);
const __m128i x2 = _mm_sub_epi16(q7, r1);
const __m128i x3 = _mm_add_epi16(q7, r1);
// Interleave to do the multiply by constants which gets us into 32bits
const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
// dct_const_round_shift
const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
// Combine
res1 = _mm_packs_epi32(w0, w1);
res7 = _mm_packs_epi32(w2, w3);
res5 = _mm_packs_epi32(w4, w5);
res3 = _mm_packs_epi32(w6, w7);
}
// Transpose the 8x8.
{
// 00 01 02 03 04 05 06 07
// 10 11 12 13 14 15 16 17
// 20 21 22 23 24 25 26 27
// 30 31 32 33 34 35 36 37
// 40 41 42 43 44 45 46 47
// 50 51 52 53 54 55 56 57
// 60 61 62 63 64 65 66 67
// 70 71 72 73 74 75 76 77
const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
// 00 10 01 11 02 12 03 13
// 20 30 21 31 22 32 23 33
// 04 14 05 15 06 16 07 17
// 24 34 25 35 26 36 27 37
// 40 50 41 51 42 52 43 53
// 60 70 61 71 62 72 63 73
// 54 54 55 55 56 56 57 57
// 64 74 65 75 66 76 67 77
const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
// 00 10 20 30 01 11 21 31
// 40 50 60 70 41 51 61 71
// 02 12 22 32 03 13 23 33
// 42 52 62 72 43 53 63 73
// 04 14 24 34 05 15 21 36
// 44 54 64 74 45 55 61 76
// 06 16 26 36 07 17 27 37
// 46 56 66 76 47 57 67 77
in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
// 00 10 20 30 40 50 60 70
// 01 11 21 31 41 51 61 71
// 02 12 22 32 42 52 62 72
// 03 13 23 33 43 53 63 73
// 04 14 24 34 44 54 64 74
// 05 15 25 35 45 55 65 75
// 06 16 26 36 46 56 66 76
// 07 17 27 37 47 57 67 77
}
}
// Post-condition output and store it
{
// Post-condition (division by two)
// division of two 16 bits signed numbers using shifts
// n / 2 = (n - (n >> 15)) >> 1
const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
in0 = _mm_sub_epi16(in0, sign_in0);
in1 = _mm_sub_epi16(in1, sign_in1);
in2 = _mm_sub_epi16(in2, sign_in2);
in3 = _mm_sub_epi16(in3, sign_in3);
in4 = _mm_sub_epi16(in4, sign_in4);
in5 = _mm_sub_epi16(in5, sign_in5);
in6 = _mm_sub_epi16(in6, sign_in6);
in7 = _mm_sub_epi16(in7, sign_in7);
in0 = _mm_srai_epi16(in0, 1);
in1 = _mm_srai_epi16(in1, 1);
in2 = _mm_srai_epi16(in2, 1);
in3 = _mm_srai_epi16(in3, 1);
in4 = _mm_srai_epi16(in4, 1);
in5 = _mm_srai_epi16(in5, 1);
in6 = _mm_srai_epi16(in6, 1);
in7 = _mm_srai_epi16(in7, 1);
}
iscan_ptr += n_coeffs;
qcoeff_ptr += n_coeffs;
dqcoeff_ptr += n_coeffs;
n_coeffs = -n_coeffs;
zero = _mm_setzero_si128();
if (!skip_block) {
__m128i eob;
__m128i round, quant, dequant, thr;
int16_t nzflag;
{
__m128i coeff0, coeff1;
// Setup global values
{
round = _mm_load_si128((const __m128i *)round_ptr);
quant = _mm_load_si128((const __m128i *)quant_ptr);
dequant = _mm_load_si128((const __m128i *)dequant_ptr);
}
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
// Do DC and first 15 AC
coeff0 = *in[0];
coeff1 = *in[1];
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
round = _mm_unpackhi_epi64(round, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
quant = _mm_unpackhi_epi64(quant, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
dequant = _mm_unpackhi_epi64(dequant, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
}
{
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob = _mm_max_epi16(eob, eob1);
}
n_coeffs += 8 * 2;
}
// AC only loop
index = 2;
thr = _mm_srai_epi16(dequant, 1);
while (n_coeffs < 0) {
__m128i coeff0, coeff1;
{
__m128i coeff0_sign, coeff1_sign;
__m128i qcoeff0, qcoeff1;
__m128i qtmp0, qtmp1;
assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
coeff0 = *in[index];
coeff1 = *in[index + 1];
// Poor man's sign extract
coeff0_sign = _mm_srai_epi16(coeff0, 15);
coeff1_sign = _mm_srai_epi16(coeff1, 15);
qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
_mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
if (nzflag) {
qcoeff0 = _mm_adds_epi16(qcoeff0, round);
qcoeff1 = _mm_adds_epi16(qcoeff1, round);
qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
// Reinsert signs
qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
store_tran_low(qcoeff0, qcoeff_ptr + n_coeffs);
store_tran_low(qcoeff1, qcoeff_ptr + n_coeffs + 8);
coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
store_tran_low(coeff0, dqcoeff_ptr + n_coeffs);
store_tran_low(coeff1, dqcoeff_ptr + n_coeffs + 8);
} else {
// Maybe a more efficient way to store 0?
store_zero_tran_low(qcoeff_ptr + n_coeffs);
store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
store_zero_tran_low(dqcoeff_ptr + n_coeffs);
store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
}
}
if (nzflag) {
// Scan for eob
__m128i zero_coeff0, zero_coeff1;
__m128i nzero_coeff0, nzero_coeff1;
__m128i iscan0, iscan1;
__m128i eob0, eob1;
zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
// Add one to convert from indices to counts
iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
eob0 = _mm_and_si128(iscan0, nzero_coeff0);
eob1 = _mm_and_si128(iscan1, nzero_coeff1);
eob0 = _mm_max_epi16(eob0, eob1);
eob = _mm_max_epi16(eob, eob0);
}
n_coeffs += 8 * 2;
index += 2;
}
// Accumulate EOB
{
__m128i eob_shuffled;
eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
eob = _mm_max_epi16(eob, eob_shuffled);
eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
eob = _mm_max_epi16(eob, eob_shuffled);
*eob_ptr = _mm_extract_epi16(eob, 1);
}
} else {
do {
store_zero_tran_low(dqcoeff_ptr + n_coeffs);
store_zero_tran_low(dqcoeff_ptr + n_coeffs + 8);
store_zero_tran_low(qcoeff_ptr + n_coeffs);
store_zero_tran_low(qcoeff_ptr + n_coeffs + 8);
n_coeffs += 8 * 2;
} while (n_coeffs < 0);
*eob_ptr = 0;
}
}
