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));
}
__m128i test_mm_abs_epi8(__m128i a) {
// CHECK-LABEL: test_mm_abs_epi8
// CHECK: [[SUB:%.+]] = sub <16 x i8> zeroinitializer, [[A:%.+]]
// CHECK: [[CMP:%.+]] = icmp sgt <16 x i8> [[A]], zeroinitializer
// CHECK: %{{.*}} = select <16 x i1> [[CMP]], <16 x i8> [[A]], <16 x i8> [[SUB]]
return _mm_abs_epi8(a);
}
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));
}
int main(int, char**)
{
volatile __m128i a = _mm_set1_epi32(42);
_mm_abs_epi8(a);
volatile __m128i result = _mm_sign_epi16(a, _mm_set1_epi32(64));
(void)result;
return 0;
}
void demod_16qam_lte_b_sse(const cf_t *symbols, int8_t *llr, int nsymbols) {
float *symbolsPtr = (float*) symbols;
__m128i *resultPtr = (__m128i*) llr;
__m128 symbol1, symbol2, symbol3, symbol4;
__m128i symbol_i1, symbol_i2, symbol_i3, symbol_i4, symbol_i, symbol_abs, symbol_12, symbol_34;
__m128i offset = _mm_set1_epi8(2*SCALE_BYTE_CONV_QAM16/sqrt(10));
__m128i result1n, result1a, result2n, result2a;
__m128 scale_v = _mm_set1_ps(-SCALE_BYTE_CONV_QAM16);
__m128i shuffle_negated_1 = _mm_set_epi8(0xff,0xff,7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0);
__m128i shuffle_abs_1 = _mm_set_epi8(7,6,0xff,0xff,5,4,0xff,0xff,3,2,0xff,0xff,1,0,0xff,0xff);
__m128i shuffle_negated_2 = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8);
__m128i shuffle_abs_2 = _mm_set_epi8(15,14,0xff,0xff,13,12,0xff,0xff,11,10,0xff,0xff,9,8,0xff,0xff);
for (int i=0;i<nsymbols/8;i++) {
symbol1 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol2 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol3 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol4 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol_i1 = _mm_cvtps_epi32(_mm_mul_ps(symbol1, scale_v));
symbol_i2 = _mm_cvtps_epi32(_mm_mul_ps(symbol2, scale_v));
symbol_i3 = _mm_cvtps_epi32(_mm_mul_ps(symbol3, scale_v));
symbol_i4 = _mm_cvtps_epi32(_mm_mul_ps(symbol4, scale_v));
symbol_12 = _mm_packs_epi32(symbol_i1, symbol_i2);
symbol_34 = _mm_packs_epi32(symbol_i3, symbol_i4);
symbol_i = _mm_packs_epi16(symbol_12, symbol_34);
symbol_abs = _mm_abs_epi8(symbol_i);
symbol_abs = _mm_sub_epi8(symbol_abs, offset);
result1n = _mm_shuffle_epi8(symbol_i, shuffle_negated_1);
result1a = _mm_shuffle_epi8(symbol_abs, shuffle_abs_1);
result2n = _mm_shuffle_epi8(symbol_i, shuffle_negated_2);
result2a = _mm_shuffle_epi8(symbol_abs, shuffle_abs_2);
_mm_store_si128(resultPtr, _mm_or_si128(result1n, result1a)); resultPtr++;
_mm_store_si128(resultPtr, _mm_or_si128(result2n, result2a)); resultPtr++;
}
// Demodulate last symbols
for (int i=8*(nsymbols/8);i<nsymbols;i++) {
short yre = (int8_t) (SCALE_BYTE_CONV_QAM16*crealf(symbols[i]));
short yim = (int8_t) (SCALE_BYTE_CONV_QAM16*cimagf(symbols[i]));
llr[4*i+0] = -yre;
llr[4*i+1] = -yim;
llr[4*i+2] = abs(yre)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
llr[4*i+3] = abs(yim)-2*SCALE_BYTE_CONV_QAM16/sqrt(10);
}
}
void demod_64qam_lte_b_sse(const cf_t *symbols, int8_t *llr, int nsymbols)
{
float *symbolsPtr = (float*) symbols;
__m128i *resultPtr = (__m128i*) llr;
__m128 symbol1, symbol2, symbol3, symbol4;
__m128i symbol_i1, symbol_i2, symbol_i3, symbol_i4, symbol_i, symbol_abs, symbol_abs2,symbol_12, symbol_34;
__m128i offset1 = _mm_set1_epi8(4*SCALE_BYTE_CONV_QAM64/sqrt(42));
__m128i offset2 = _mm_set1_epi8(2*SCALE_BYTE_CONV_QAM64/sqrt(42));
__m128 scale_v = _mm_set1_ps(-SCALE_BYTE_CONV_QAM64);
__m128i result11, result12, result13, result22, result21,result23, result31, result32, result33;
__m128i shuffle_negated_1 = _mm_set_epi8(0xff,0xff,5,4,0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0);
__m128i shuffle_negated_2 = _mm_set_epi8(11,10,0xff,0xff,0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff);
__m128i shuffle_negated_3 = _mm_set_epi8(0xff,0xff,0xff,0xff,15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff);
__m128i shuffle_abs_1 = _mm_set_epi8(5,4,0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0,0xff,0xff);
__m128i shuffle_abs_2 = _mm_set_epi8(0xff,0xff,0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff,0xff,0xff);
__m128i shuffle_abs_3 = _mm_set_epi8(0xff,0xff,15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff,11,10);
__m128i shuffle_abs2_1 = _mm_set_epi8(0xff,0xff,0xff,0xff,3,2,0xff,0xff,0xff,0xff,1,0,0xff,0xff,0xff,0xff);
__m128i shuffle_abs2_2 = _mm_set_epi8(0xff,0xff,9,8,0xff,0xff,0xff,0xff,7,6,0xff,0xff,0xff,0xff,5,4);
__m128i shuffle_abs2_3 = _mm_set_epi8(15,14,0xff,0xff,0xff,0xff,13,12,0xff,0xff,0xff,0xff,11,10,0xff,0xff);
for (int i=0;i<nsymbols/8;i++) {
symbol1 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol2 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol3 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol4 = _mm_load_ps(symbolsPtr); symbolsPtr+=4;
symbol_i1 = _mm_cvtps_epi32(_mm_mul_ps(symbol1, scale_v));
symbol_i2 = _mm_cvtps_epi32(_mm_mul_ps(symbol2, scale_v));
symbol_i3 = _mm_cvtps_epi32(_mm_mul_ps(symbol3, scale_v));
symbol_i4 = _mm_cvtps_epi32(_mm_mul_ps(symbol4, scale_v));
symbol_12 = _mm_packs_epi32(symbol_i1, symbol_i2);
symbol_34 = _mm_packs_epi32(symbol_i3, symbol_i4);
symbol_i = _mm_packs_epi16(symbol_12, symbol_34);
symbol_abs = _mm_abs_epi8(symbol_i);
symbol_abs = _mm_sub_epi8(symbol_abs, offset1);
symbol_abs2 = _mm_sub_epi8(_mm_abs_epi8(symbol_abs), offset2);
result11 = _mm_shuffle_epi8(symbol_i, shuffle_negated_1);
result12 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_1);
result13 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_1);
result21 = _mm_shuffle_epi8(symbol_i, shuffle_negated_2);
result22 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_2);
result23 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_2);
result31 = _mm_shuffle_epi8(symbol_i, shuffle_negated_3);
result32 = _mm_shuffle_epi8(symbol_abs, shuffle_abs_3);
result33 = _mm_shuffle_epi8(symbol_abs2, shuffle_abs2_3);
_mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result11, result12),result13)); resultPtr++;
_mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result21, result22),result23)); resultPtr++;
_mm_store_si128(resultPtr, _mm_or_si128(_mm_or_si128(result31, result32),result33)); resultPtr++;
}
for (int i=8*(nsymbols/8);i<nsymbols;i++) {
float yre = (int8_t) (SCALE_BYTE_CONV_QAM64*crealf(symbols[i]));
float yim = (int8_t) (SCALE_BYTE_CONV_QAM64*cimagf(symbols[i]));
llr[6*i+0] = -yre;
llr[6*i+1] = -yim;
llr[6*i+2] = abs(yre)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
llr[6*i+3] = abs(yim)-4*SCALE_BYTE_CONV_QAM64/sqrt(42);
llr[6*i+4] = abs(llr[6*i+2])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
llr[6*i+5] = abs(llr[6*i+3])-2*SCALE_BYTE_CONV_QAM64/sqrt(42);
}
}
__m128i test_mm_abs_epi8(__m128i a) {
// CHECK-LABEL: test_mm_abs_epi8
// CHECK: call <16 x i8> @llvm.x86.ssse3.pabs.b.128
return _mm_abs_epi8(a);
}
