int32_t sse_sadbw_unrolled4_sumsignedbytes(int8_t* array, size_t size) {
const __m128i zero = _mm_setzero_si128();
__m128i positive = zero;
__m128i negative = zero;
for (size_t i=0; i < size; i += 16*4) {
const __m128i v0 = _mm_loadu_si128((__m128i*)(array + i + 0*16));
const __m128i v1 = _mm_loadu_si128((__m128i*)(array + i + 1*16));
const __m128i v2 = _mm_loadu_si128((__m128i*)(array + i + 2*16));
const __m128i v3 = _mm_loadu_si128((__m128i*)(array + i + 3*16));
{
const __m128i v = v0;
const __m128i m = _mm_cmplt_epi8(v, zero);
const __m128i t0 = _mm_sad_epu8(_mm_andnot_si128(m, v), zero);
const __m128i va = _mm_abs_epi8(v);
const __m128i t1 = _mm_sad_epu8(_mm_and_si128(m, va), zero);
positive = _mm_add_epi32(positive, t0);
negative = _mm_sub_epi32(negative, t1);
}
{
const __m128i v = v1;
const __m128i m = _mm_cmplt_epi8(v, zero);
const __m128i t0 = _mm_sad_epu8(_mm_andnot_si128(m, v), zero);
const __m128i va = _mm_abs_epi8(v);
const __m128i t1 = _mm_sad_epu8(_mm_and_si128(m, va), zero);
positive = _mm_add_epi32(positive, t0);
negative = _mm_sub_epi32(negative, t1);
}
{
const __m128i v = v2;
const __m128i m = _mm_cmplt_epi8(v, zero);
const __m128i t0 = _mm_sad_epu8(_mm_andnot_si128(m, v), zero);
const __m128i va = _mm_abs_epi8(v);
const __m128i t1 = _mm_sad_epu8(_mm_and_si128(m, va), zero);
positive = _mm_add_epi32(positive, t0);
negative = _mm_sub_epi32(negative, t1);
}
{
const __m128i v = v3;
const __m128i m = _mm_cmplt_epi8(v, zero);
const __m128i t0 = _mm_sad_epu8(_mm_andnot_si128(m, v), zero);
const __m128i va = _mm_abs_epi8(v);
const __m128i t1 = _mm_sad_epu8(_mm_and_si128(m, va), zero);
positive = _mm_add_epi32(positive, t0);
negative = _mm_sub_epi32(negative, t1);
}
}
const __m128i accumulator = _mm_add_epi32(positive, negative);
return int32_t(_mm_extract_epi32(accumulator, 0)) +
int32_t(_mm_extract_epi32(accumulator, 2));
}
int32_t sse_sadbw_sumsignedbytes(int8_t* array, size_t size) {
const __m128i zero = _mm_setzero_si128();
__m128i positive = zero;
__m128i negative = zero;
for (size_t i=0; i < size; i += 16) {
const __m128i v = _mm_loadu_si128((__m128i*)(array + i));
const __m128i m = _mm_cmplt_epi8(v, zero);
const __m128i va = _mm_abs_epi8(v);
// sum just positive numbers
const __m128i t0 = _mm_sad_epu8(_mm_andnot_si128(m, v), zero);
// sum just negative numbers
const __m128i t1 = _mm_sad_epu8(_mm_and_si128(m, va), zero);
positive = _mm_add_epi32(positive, t0);
negative = _mm_sub_epi32(negative, t1);
}
const __m128i accumulator = _mm_add_epi32(positive, negative);
return int32_t(_mm_extract_epi32(accumulator, 0)) +
int32_t(_mm_extract_epi32(accumulator, 2));
}
__m128i test_mm_cmplt_epi8(__m128i A, __m128i B) {
// DAG-LABEL: test_mm_cmplt_epi8
// DAG: icmp sgt <16 x i8>
//
// ASM-LABEL: test_mm_cmplt_epi8
// ASM: pcmpgtb
return _mm_cmplt_epi8(A, B);
}
SIMDValue SIMDInt8x16Operation::OpLessThan(const SIMDValue& aValue, const SIMDValue& bValue)
{
X86SIMDValue x86Result;
X86SIMDValue tmpaValue = X86SIMDValue::ToX86SIMDValue(aValue);
X86SIMDValue tmpbValue = X86SIMDValue::ToX86SIMDValue(bValue);
x86Result.m128i_value = _mm_cmplt_epi8(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a < b?
return X86SIMDValue::ToSIMDValue(x86Result);
}
inline __m128i load_aligned_int32(const int8_t* src)
{
__m128i tmp = _mm_loadl_epi64((const __m128i*)src);
#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE4_1_VERSION
__m128i res = _mm_cvtepi8_epi32(tmp);
#else
__m128i mask = _mm_cmplt_epi8(tmp, _mm_set1_epi8(0));
__m128i tmp1 = _mm_unpacklo_epi8(tmp, mask);
mask = _mm_cmplt_epi16(tmp1, _mm_set1_epi16(0));
__m128i res = _mm_unpacklo_epi16(tmp1, mask);
#endif
return res;
}
int main()
{
ssereg a(0xffffffff);
ssereg b(0x00000000);
ssereg c = _mm_cmplt_epi8(a, b);
a.print();
b.print();
c.print();
return 0;
}
SIMDValue SIMDUint8x16Operation::OpLessThan(const SIMDValue& aValue, const SIMDValue& bValue)
{
X86SIMDValue x86Result;
X86SIMDValue tmpaValue = X86SIMDValue::ToX86SIMDValue(aValue);
X86SIMDValue tmpbValue = X86SIMDValue::ToX86SIMDValue(bValue);
X86SIMDValue signBits = { {0x80808080,0x80808080, 0x80808080, 0x80808080} };
// Signed comparison of unsigned ints can be done if the ints have the "sign" bit xored with 1
tmpaValue.m128i_value = _mm_xor_si128(tmpaValue.m128i_value, signBits.m128i_value);
tmpbValue.m128i_value = _mm_xor_si128(tmpbValue.m128i_value, signBits.m128i_value);
x86Result.m128i_value = _mm_cmplt_epi8(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a < b?
return X86SIMDValue::ToSIMDValue(x86Result);
}
/// any (*p > 2) is set to be 3
COREARRAY_DLL_DEFAULT void vec_u8_geno_valid(C_UInt8 *p, size_t n)
{
#if defined(COREARRAY_SIMD_SSE2)
// header 1, 16-byte aligned
size_t h = (16 - ((size_t)p & 0x0F)) & 0x0F;
for (; (n > 0) && (h > 0); n--, h--, p++)
if (*p > 3) *p = 3;
const __m128i zero = _mm_setzero_si128();
const __m128i three = _mm_set1_epi8(3);
for (; n >= 16; n-=16, p+=16)
{
__m128i v = _mm_load_si128((__m128i*)p);
__m128i mask = _mm_or_si128(_mm_cmplt_epi8(v, zero),
_mm_cmplt_epi8(three, v));
if (_mm_movemask_epi8(mask) > 0)
_mm_store_si128((__m128i*)p, _mm_min_epu8(v, three));
}
#endif
for (; n > 0; n--, p++) if (*p > 3) *p = 3;
}
//
// If we assumed AVX512, this routine could be implemented more efficiently
// and straightforwardly with _mm_mask_storeu_epi8().
//
inline void _assembler_store_partial(__m128i *dst, int s, int p, __m128i x0, __m128i x1, __m128i x2, __m128i x3, __m128i x4, __m128i x5, __m128i x6, __m128i x7)
{
static const __m128i r = _mm_set_epi8(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0);
__m128i mask = _mm_cmplt_epi8(r, _mm_set1_epi8(p));
x0 = _mm_and_si128(mask, x0);
x1 = _mm_and_si128(mask, x1);
x2 = _mm_and_si128(mask, x2);
x3 = _mm_and_si128(mask, x3);
x4 = _mm_and_si128(mask, x4);
x5 = _mm_and_si128(mask, x5);
x6 = _mm_and_si128(mask, x6);
x7 = _mm_and_si128(mask, x7);
__m128i y0 = _mm_loadu_si128(dst);
__m128i y1 = _mm_loadu_si128(dst+s);
__m128i y2 = _mm_loadu_si128(dst+2*s);
__m128i y3 = _mm_loadu_si128(dst+3*s);
__m128i y4 = _mm_loadu_si128(dst+4*s);
__m128i y5 = _mm_loadu_si128(dst+5*s);
__m128i y6 = _mm_loadu_si128(dst+6*s);
__m128i y7 = _mm_loadu_si128(dst+7*s);
y0 = _mm_andnot_si128(mask, y0);
y1 = _mm_andnot_si128(mask, y1);
y2 = _mm_andnot_si128(mask, y2);
y3 = _mm_andnot_si128(mask, y3);
y4 = _mm_andnot_si128(mask, y4);
y5 = _mm_andnot_si128(mask, y5);
y6 = _mm_andnot_si128(mask, y6);
y7 = _mm_andnot_si128(mask, y7);
y0 = _mm_or_si128(x0, y0);
y1 = _mm_or_si128(x1, y1);
y2 = _mm_or_si128(x2, y2);
y3 = _mm_or_si128(x3, y3);
y4 = _mm_or_si128(x4, y4);
y5 = _mm_or_si128(x5, y5);
y6 = _mm_or_si128(x6, y6);
y7 = _mm_or_si128(x7, y7);
_assembler_store_full(dst, s, y0, y1, y2, y3, y4, y5, y6, y7);
}
int
exponent_sum_square_error_sse2(uint8_t *exp0, uint8_t *exp1, int ncoefs)
{
int i, err;
int exp_error = 0;
union {
__m128i v;
int32_t res[4];
} ures;
__m128i vzero = _mm_setzero_si128();
__m128i vres = vzero;
for (i = 0; i < (ncoefs & ~15); i+=16) {
__m128i vexp = _mm_loadu_si128((__m128i*)&exp0[i]);
__m128i vexp2 = _mm_loadu_si128((__m128i*)&exp1[i]);
#if 0
//safer but needed?
__m128i vexphi = _mm_unpackhi_epi8(vexp, vzero);
__m128i vexp2hi = _mm_unpackhi_epi8(vexp2, vzero);
__m128i vexplo = _mm_unpacklo_epi8(vexp, vzero);
__m128i vexp2lo = _mm_unpacklo_epi8(vexp2, vzero);
__m128i verrhi = _mm_sub_epi16(vexphi, vexp2hi);
__m128i verrlo = _mm_sub_epi16(vexplo, vexp2lo);
#else
__m128i verr = _mm_sub_epi8(vexp, vexp2);
__m128i vsign = _mm_cmplt_epi8(verr, vzero);
__m128i verrhi = _mm_unpackhi_epi8(verr, vsign);
__m128i verrlo = _mm_unpacklo_epi8(verr, vsign);
#endif
verrhi = _mm_madd_epi16(verrhi, verrhi);
verrlo = _mm_madd_epi16(verrlo, verrlo);
verrhi = _mm_add_epi32(verrhi, verrlo);
vres = _mm_add_epi32(vres, verrhi);
}
_mm_store_si128(&ures.v, vres);
ures.res[0]+=ures.res[1];
ures.res[2]+=ures.res[3];
exp_error += ures.res[0]+ures.res[2];
for (; i < ncoefs; ++i) {
err = exp0[i] - exp1[i];
exp_error += (err * err);
}
return exp_error;
}
void
filterScanlinesSSE( unsigned char* filtered,
unsigned char* image,
unsigned int WIDTH,
unsigned int HEIGHT )
{
int blocks = 3*WIDTH/16;
// Create move-mask for last block of each scanline
__m128i mask = _mm_cmplt_epi8( _mm_set_epi8( 15, 14, 13, 12, 11, 10, 9, 8,
7, 6, 5, 4, 3, 2, 1, 0 ),
_mm_set1_epi8( 3*WIDTH-16*blocks ) );
{
const unsigned char* in = image;
unsigned char* out = filtered;
*out++ = 0;
for(int b=0; b<blocks; b++ ) {
_mm_storeu_si128( (__m128i*)out, _mm_lddqu_si128( (__m128i const*)in ) );
in += 16;
out += 16;
}
_mm_maskmoveu_si128( _mm_lddqu_si128( (__m128i const*)in ), mask, (char*)out );
}
for( unsigned int j=1; j<HEIGHT; j++ ) {
const unsigned char* in = image + 3*WIDTH*(j-1);
unsigned char* out = filtered + (3*WIDTH+1)*j;
*out++ = 2;
for(int b=0; b<blocks; b++ ) {
__m128i _t0 = _mm_lddqu_si128( (__m128i const*)in );
__m128i _t1 = _mm_lddqu_si128( (__m128i const*)(in + 3*WIDTH ) );
_mm_storeu_si128( (__m128i*)out,
_mm_sub_epi8( _t1, _t0 ) );
in += 16;
out += 16;
}
_mm_maskmoveu_si128( _mm_lddqu_si128( (__m128i const*)in ),
mask,
(char*)out );
}
}
static void add_child16(art_node16 *n, art_node **ref, unsigned char c, void *child) {
if (n->n.num_children < 16) {
__m128i cmp;
// Compare the key to all 16 stored keys
cmp = _mm_cmplt_epi8(_mm_set1_epi8(c),
_mm_loadu_si128((__m128i*)n->keys));
// Use a mask to ignore children that don't exist
unsigned mask = (1 << n->n.num_children) - 1;
unsigned bitfield = _mm_movemask_epi8(cmp) & mask;
// Check if less than any
unsigned idx;
if (bitfield) {
idx = __builtin_ctz(bitfield);
memmove(n->keys+idx+1,n->keys+idx,n->n.num_children-idx);
memmove(n->children+idx+1,n->children+idx,
(n->n.num_children-idx)*sizeof(void*));
} else
idx = n->n.num_children;
// Set the child
n->keys[idx] = c;
n->children[idx] = (art_node*)child;
n->n.num_children++;
} else {
art_node48 *new_node = (art_node48*)alloc_node(NODE48);
// Copy the child pointers and populate the key map
memcpy(new_node->children, n->children,
sizeof(void*)*n->n.num_children);
for (int i=0;i<n->n.num_children;i++) {
new_node->keys[n->keys[i]] = i + 1;
}
copy_header((art_node*)new_node, (art_node*)n);
*ref = (art_node*)new_node;
free(n);
add_child48(new_node, ref, c, child);
}
}
static inline void calc_lbp_16_strip(IplImage * src, IplImage * dst, unsigned base)
{
const signed char* src_data = (signed char*)(src->imageData + base);
unsigned char * dst_data = (unsigned char*)(dst->imageData + base);
const signed char* const src_end = (signed char*)src->imageData + (src->height-1) * src->widthStep;
__m128i pixels[3];
// Load first two rows
//pixels[0] = *(__m128i*)src_data;//_mm_set_epi64(*(__m64*)(src_data+8), *(__m64*)(src_data));
pixels[0] = _mm_set_epi64(*(__m64*)(src_data+8), *(__m64*)(src_data));
//pixels[0] = _mm_xor_si128(pixels[0], sign_bit.q); // conversion from unsigned to signed - invert sign bit
src_data += src->widthStep;
//pixels[1] = *(__m128i*)src_data;//_mm_set_epi64(*(__m64*)(src_data+8), *(__m64*)(src_data));
pixels[1] = _mm_set_epi64(*(__m64*)(src_data+8), *(__m64*)(src_data));
//pixels[1] = _mm_xor_si128(pixels[1], sign_bit.q);
src_data += src->widthStep;
int phase = 2;
__m128i * phase_map[3][3] = {
{pixels+1, pixels+2, pixels},
{pixels+2, pixels, pixels+1},
{pixels, pixels+1, pixels+2},
};
while (src_data < src_end)
{
register __m128i weight = ones.q;
register __m128i code = _mm_setzero_si128();
//pixels[phase] = _mm_set_epi64(*(__m64*)(src_data+8), *(__m64*)(src_data));
//pixels[phase] = _mm_xor_si128(pixels[phase], sign_bit.q);
//pixels[phase] = _mm_xor_si128(_mm_lddqu_si128((__m128i*)src_data), sign_bit.q);
pixels[phase] = _mm_lddqu_si128((__m128i*)src_data);
src_data += src->widthStep;
dst_data += dst->widthStep;
_mm_prefetch(src_data, _MM_HINT_T0);
register __m128i a = *(phase_map[phase][0]);
register __m128i b = *(phase_map[phase][1]);
register __m128i c = *(phase_map[phase][2]);
phase++;
phase = (phase == 3) ? 0 : phase;
// X . . A
// . o . B
// . . . C
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_slli_si128(a, 1)), weight));
weight = _mm_slli_epi64(weight, 1);
// . X .
// . .
// . . .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, a), weight));
weight = _mm_slli_epi64(weight, 1);
// . . X
// . .
// . . .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_srli_si128(a, 1)), weight));
weight = _mm_slli_epi64(weight, 1);
// . . .
// . X
// . . .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_srli_si128(b, 1)), weight));
weight = _mm_slli_epi64(weight, 1);
// . . .
// . .
// . . X
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_srli_si128(c, 1)), weight));
weight = _mm_slli_epi64(weight, 1);
// . . .
// . .
// . X .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, c), weight));
weight = _mm_slli_epi64(weight, 1);
// . . .
// . .
// X . .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_slli_si128(c, 1)), weight));
weight = _mm_slli_epi64(weight, 1);
// . . .
// X .
// . . .
code = _mm_or_si128(code, _mm_and_si128(_mm_cmplt_epi8(b, _mm_slli_si128(b, 1)), weight));
_mm_maskmoveu_si128(code, lbp_valid_mask.q, (char*)dst_data); // store the results - unaligned write
}
}
}bool validate_utf8_sse(const char *src, size_t len) {
const char *end = src + len;
while (src + 16 < end) {
__m128i chunk = _mm_loadu_si128((const __m128i *)(src));
int asciiMask = _mm_movemask_epi8(chunk);
if (!asciiMask) {
src += 16;
continue;
}
__m128i chunk_signed = _mm_add_epi8(chunk, _mm_set1_epi8(0x80));
__m128i cond2 =
_mm_cmplt_epi8(_mm_set1_epi8(0xc2 - 1 - 0x80), chunk_signed);
__m128i state = _mm_set1_epi8((char)(0x0 | 0x80));
state = _mm_blendv_epi8(state, _mm_set1_epi8((char)(0x2 | 0xc0)), cond2);
__m128i cond3 =
_mm_cmplt_epi8(_mm_set1_epi8(0xe0 - 1 - 0x80), chunk_signed);
state = _mm_blendv_epi8(state, _mm_set1_epi8((char)(0x3 | 0xe0)), cond3);
__m128i mask3 = _mm_slli_si128(cond3, 1);
__m128i cond4 =
_mm_cmplt_epi8(_mm_set1_epi8(0xf0 - 1 - 0x80), chunk_signed);
// Fall back to the scalar processing
if (_mm_movemask_epi8(cond4)) {
break;
}
__m128i count = _mm_and_si128(state, _mm_set1_epi8(0x7));
__m128i count_sub1 = _mm_subs_epu8(count, _mm_set1_epi8(0x1));
__m128i counts = _mm_add_epi8(count, _mm_slli_si128(count_sub1, 1));
__m128i shifts = count_sub1;
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 1));
counts = _mm_add_epi8(
counts, _mm_slli_si128(_mm_subs_epu8(counts, _mm_set1_epi8(0x2)), 2));
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 2));
if (asciiMask ^ _mm_movemask_epi8(_mm_cmpgt_epi8(counts, _mm_set1_epi8(0))))
return false; // error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 4));
if (_mm_movemask_epi8(_mm_cmpgt_epi8(
_mm_sub_epi8(_mm_slli_si128(counts, 1), counts), _mm_set1_epi8(1))))
return false; // error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 8));
__m128i mask = _mm_and_si128(state, _mm_set1_epi8(0xf8));
shifts =
_mm_and_si128(shifts, _mm_cmplt_epi8(counts, _mm_set1_epi8(2))); // <=1
chunk =
_mm_andnot_si128(mask, chunk); // from now on, we only have usefull bits
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 1),
_mm_srli_si128(_mm_slli_epi16(shifts, 7), 1));
__m128i chunk_right = _mm_slli_si128(chunk, 1);
__m128i chunk_low = _mm_blendv_epi8(
chunk,
_mm_or_si128(chunk, _mm_and_si128(_mm_slli_epi16(chunk_right, 6),
_mm_set1_epi8(0xc0))),
_mm_cmpeq_epi8(counts, _mm_set1_epi8(1)));
__m128i chunk_high =
_mm_and_si128(chunk, _mm_cmpeq_epi8(counts, _mm_set1_epi8(2)));
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 2),
_mm_srli_si128(_mm_slli_epi16(shifts, 6), 2));
chunk_high = _mm_srli_epi32(chunk_high, 2);
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 4),
_mm_srli_si128(_mm_slli_epi16(shifts, 5), 4));
chunk_high = _mm_or_si128(
chunk_high, _mm_and_si128(_mm_and_si128(_mm_slli_epi32(chunk_right, 4),
_mm_set1_epi8(0xf0)),
mask3));
int c = _mm_extract_epi16(counts, 7);
int source_advance = !(c & 0x0200) ? 16 : !(c & 0x02) ? 15 : 14;
__m128i high_bits = _mm_and_si128(chunk_high, _mm_set1_epi8(0xf8));
if (!_mm_testz_si128(
mask3,
_mm_or_si128(_mm_cmpeq_epi8(high_bits, _mm_set1_epi8(0x00)),
_mm_cmpeq_epi8(high_bits, _mm_set1_epi8(0xd8)))))
return false;
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 8),
_mm_srli_si128(_mm_slli_epi16(shifts, 4), 8));
chunk_high = _mm_slli_si128(chunk_high, 1);
__m128i shuf =
_mm_add_epi8(shifts, _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5,
4, 3, 2, 1, 0));
chunk_low = _mm_shuffle_epi8(chunk_low, shuf);
chunk_high = _mm_shuffle_epi8(chunk_high, shuf);
__m128i utf16_low = _mm_unpacklo_epi8(chunk_low, chunk_high);
__m128i utf16_high = _mm_unpackhi_epi8(chunk_low, chunk_high);
if (_mm_cmpestrc(_mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_high, 8,
_SIDD_UWORD_OPS | _SIDD_CMP_RANGES) |
_mm_cmpestrc(_mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_low, 8,
_SIDD_UWORD_OPS | _SIDD_CMP_RANGES)) {
return false;
}
src += source_advance;
}
return validate_utf8(src, end - src);
}
FORCE_INLINE
static inline void fast_scan_1(const std::uint8_t* partition, const unsigned* labels,
const float* dists, const __m128i (&min4)[4], __m128i (&ft4)[4][16],
const float qmin, const float qmax, binheap* bh, unsigned scan_pqcode_count) {
const unsigned simd_pqcode_count = 16;
const int comp_block_size = 16;
const unsigned simd_block_size = simd_pqcode_count * (4 * 1 + 4 * 0.5);
const group_header* hdr;
float bh_bound = qmax;
__m128i bh_bound_quant = _mm_set1_epi8(Q127(bh_bound, qmin, bh_bound)); // CHK. Is 127
for (;;) {
// Parse group header
hdr = reinterpret_cast<const group_header*>(partition);
// Check if last group (All bits of size set to 1)
if (hdr->size == std::numeric_limits<decltype(hdr->size)>::max()) {
return;
}
partition += sizeof(*hdr);
unsigned simd_block_count = (static_cast<unsigned>(hdr->size)
+ simd_pqcode_count - 1) / simd_pqcode_count;
// Load tables
__m128i ft4_group[4];
ft4_group[0] = ft4[0][hdr->values[0] >> 4];
ft4_group[1] = ft4[1][hdr->values[1] >> 4];
ft4_group[2] = ft4[2][hdr->values[2] >> 4];
ft4_group[3] = ft4[3][hdr->values[3] >> 4];
// Scan SIMD Blocks
while (simd_block_count--) {
const __m128i low_bits_mask = _mm_set_epi64x(0x0f0f0f0f0f0f0f0f,
0x0f0f0f0f0f0f0f0f);
// Component 0
const __m128i comps_0 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(partition));
const __m128i masked_comps_0 = _mm_and_si128(comps_0, low_bits_mask);
__m128i candidates = _mm_shuffle_epi8(min4[0], masked_comps_0);
// Components 1..3
for (int comp_i = 1; comp_i < 4; ++comp_i) {
const __m128i comps = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(partition
+ comp_i * comp_block_size));
const __m128i masked_comps = _mm_and_si128(comps, low_bits_mask);
const __m128i partial = _mm_shuffle_epi8(min4[comp_i], masked_comps);
candidates = _mm_adds_epi8(candidates, partial);
}
// Components 4-5
__m128i comps_45 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(partition
+ 4 * comp_block_size));
const __m128i masked_comps_4 = _mm_and_si128(comps_45, low_bits_mask);
const __m128i partial_4 = _mm_shuffle_epi8(ft4_group[0], masked_comps_4);
candidates = _mm_adds_epi8(candidates, partial_4);
comps_45 = _mm_srli_epi64(comps_45, 4);
const __m128i masked_comps_5 = _mm_and_si128(comps_45, low_bits_mask);
const __m128i partial_5 = _mm_shuffle_epi8(ft4_group[1], masked_comps_5);
candidates = _mm_adds_epi8(candidates, partial_5);
// Components 6-7
__m128i comps_67 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(partition
+ 5 * comp_block_size));
const __m128i masked_comps_6 = _mm_and_si128(comps_67, low_bits_mask);
const __m128i partial_6 = _mm_shuffle_epi8(ft4_group[2], masked_comps_6);
candidates = _mm_adds_epi8(candidates, partial_6);
const __m128i comps_7 = _mm_srli_epi64(comps_67, 4);
const __m128i masked_comp_7 = _mm_and_si128(comps_7, low_bits_mask);
const __m128i partial_7 = _mm_shuffle_epi8(ft4_group[3], masked_comp_7);
candidates = _mm_adds_epi8(candidates, partial_7);
// Compare
const __m128i compare = _mm_cmplt_epi8(candidates, bh_bound_quant);
int cmp = _mm_movemask_epi8(compare);
//std::uint64_t cmp_low = (_mm_cvtsi128_si64(compare));
//std::uint64_t cmp_high = (_mm_extract_epi64(compare, 1));
// Compute current block size
int current_block_actual_size = 0;
if(simd_block_count == 0) {
current_block_actual_size = hdr->size % simd_pqcode_count;
if(current_block_actual_size == 0) {
current_block_actual_size = simd_pqcode_count;
} else {
/*__m128i mask;
compute_simd_mask(current_block_actual_size, mask);
compare = _mm_and_si128(compare, mask);*/
/*
std::uint64_t low_mask;
std::uint64_t high_mask;
compute_high_low_mask(current_block_actual_size, low_mask, high_mask);
cmp_low = cmp_low & low_mask;
cmp_high = cmp_high & high_mask;
*/
cmp = cmp & BITMASK(current_block_actual_size);
}
} else {
current_block_actual_size = simd_pqcode_count;
}
if(cmp) {
// Check low quadword
const std::uint8_t cmp_low = cmp & 0xff;
if (cmp_low) {
/*const std::uint64_t low_possible_positions = 0x0706050403020100;
const std::uint64_t match_positions = _pext_u64(
low_possible_positions, cmp_low);*/
const int match_count = _popcnt32(cmp_low);
std::uint64_t match_pos = masktable[cmp_low];
for (int i = 0; i < match_count; ++i) {
const std::uint8_t pos = match_pos & 0xff;
match_pos >>= 8;
const float candidate = scan_pqcode_in_simd_block_1(pos,
partition, hdr->values, dists);
if (candidate < bh_bound) {
bh->push(labels[scan_pqcode_count + pos],
candidate);
bh_bound = bh->max();
bh_bound_quant = _mm_set1_epi8(
Q127(bh_bound, qmin, qmax));
}
}
}
// Check high quadword
const std::uint8_t cmp_high = (cmp >> 8);
if (cmp_high) {
/*const std::uint64_t high_possible_positions = 0x0f0e0d0c0b0a0908;
const std::uint64_t match_positions = _pext_u64(
high_possible_positions, cmp_high);*/
const int match_count = _popcnt32(cmp_high);
std::uint64_t match_pos = masktable[cmp_high] + 0x0808080808080808;
for (int i = 0; i < match_count; ++i) {
const std::uint8_t pos = match_pos & 0xff;
match_pos >>= 8;
const float candidate = scan_pqcode_in_simd_block_1(pos,
partition, hdr->values, dists);
if (candidate < bh_bound) {
bh->push(labels[scan_pqcode_count + pos],
candidate);
bh_bound = bh->max();
bh_bound_quant = _mm_set1_epi8(
Q127(bh_bound, qmin, qmax));
}
}
}
}
partition += simd_block_size;
scan_pqcode_count += current_block_actual_size;
}
}
}
