void static
avx2_test (void)
{
union256i_d s1, s2;
union256i_w u;
short e[16];
int i;
s1.x = _mm256_set_epi32 (1, 2, 3, 4, 65000, 20, 30, 90);
s2.x = _mm256_set_epi32 (88, 44, 33, 22, 11, 98, 76, -65000);
u.x = _mm256_packs_epi32 (s1.x, s2.x);
for (i = 0; i < 4; i++)
{
e[i] = int_to_short (s1.a[i]);
e[i + 4] = int_to_short (s2.a[i]);
e[i + 8] = int_to_short (s1.a[i + 4]);
e[i + 12] = int_to_short (s2.a[i + 4]);
}
if (check_union256i_w (u, e))
abort ();
}
int
main(void)
{
//_mm256_permutevar_ps
__m256 da = _mm256_setr_ps(1,2,3,4,5,6,7,8);
__m256i ds = _mm256_set_epi32(0,0,0,0,0,0,0,0x02);
__m256 dc;
printf("da: ");
for(int i=0; i<sizeof(da)/sizeof(da.m256_f32[0]); i++)
printf("%5.1f ", da.m256_f32[i]);
printf("\n");
dc = _mm256_permutevar_ps(da, ds);
printf("dc: ");
for(int i=0; i<sizeof(dc)/sizeof(dc.m256_f32[0]); i++)
printf("%5.1f ", dc.m256_f32[i]);
printf("\n");
ds = _mm256_set_epi32(0,0,0,0,0,0,0,0x01);
dc = _mm256_permutevar_ps(da, ds);
printf("dc: ");
for(int i=0; i<sizeof(dc)/sizeof(dc.m256_f32[0]); i++)
printf("%5.1f ", dc.m256_f32[i]);
printf("\n\n");
//_mm_permutevar_ps
__m128 fa = _mm_setr_ps(1,2,3,4);
__m128i fs = _mm_set_epi32(0,0,0,0x02);
__m128 fc;
printf("fa: ");
for(int i=0; i<sizeof(fa)/sizeof(fa.m128_f32[0]); i++)
printf("%5.1f ", fa.m128_f32[i]);
printf("\n");
fc = _mm_permutevar_ps(fa, fs);
printf("fc: ");
for(int i=0; i<sizeof(fc)/sizeof(fc.m128_f32[0]); i++)
printf("%5.1f ", fc.m128_f32[i]);
printf("\n");
fs = _mm_set_epi32(0,0,0,0x01);
fc = _mm_permutevar_ps(fa, fs);
printf("fc: ");
for(int i=0; i<sizeof(fc)/sizeof(fc.m128_f32[0]); i++)
printf("%5.1f ", fc.m128_f32[i]);
printf("\n");
return 0;
}
__m256i inline Read8(const unsigned char* chunk, int offset) {
__m256i ret = _mm256_set_epi32(
ReadLE32(chunk + 0 + offset),
ReadLE32(chunk + 64 + offset),
ReadLE32(chunk + 128 + offset),
ReadLE32(chunk + 192 + offset),
ReadLE32(chunk + 256 + offset),
ReadLE32(chunk + 320 + offset),
ReadLE32(chunk + 384 + offset),
ReadLE32(chunk + 448 + offset)
);
return _mm256_shuffle_epi8(ret, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
}
// Computes part of matrix.vector v = Wu. Computes N=8 results.
// For details see PartialMatrixDotVector64 with N=8.
static void PartialMatrixDotVector8(const int8_t* wi, const double* scales,
const int8_t* u, int num_in, int num_out,
double* v) {
// Register containing 16-bit ones for horizontal add with 16->32 bit
// conversion.
__m256i ones =
_mm256_set_epi16(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1);
__m256i shift_id = _mm256_set_epi32(0, 7, 6, 5, 4, 3, 2, 1);
// Initialize all the results to 0.
__m256i result0 = _mm256_setzero_si256();
// Iterate over the input (u), one registerful at a time.
for (int j = 0; j < num_in;) {
__m256i inputs =
_mm256_loadu_si256(reinterpret_cast<const __m256i*>(u + j));
// Inputs are processed in groups of kNumInputsPerGroup, replicated
// kNumInputGroups times.
for (int ig = 0; ig < kNumInputGroups && j < num_in;
++ig, j += kNumInputsPerGroup) {
// Replicate the low 32 bits (4 inputs) 8 times.
__m256i rep_input =
_mm256_broadcastd_epi32(_mm256_castsi256_si128(inputs));
// Rotate the inputs in groups of 4, so the next 4 inputs are ready.
inputs = _mm256_permutevar8x32_epi32(inputs, shift_id);
__m256i weights, reps;
// Mul-add, with horizontal add of the 4 inputs to each of the results.
MultiplyGroup(rep_input, ones, wi, weights, reps, result0);
}
}
ExtractResults(result0, shift_id, wi, scales, num_out, v);
}
static void
avx2_test (void)
{
union256i_d u, s1, s2;
int e[8];
int i;
s1.x = _mm256_set_epi32 (1, 2, 3, 4, 10, 20, 30, 90000);
s2.x = _mm256_set_epi32 (88, 44, 3, 22, 11, 98, 76, -100);
u.x = _mm256_cmpgt_epi32 (s1.x, s2.x);
for (i = 0; i < 8; i++)
e[i] = (s1.a[i] > s2.a[i]) ? -1 : 0;
if (check_union256i_d (u, e))
abort ();
}
void inline Write8(unsigned char* out, int offset, __m256i v) {
v = _mm256_shuffle_epi8(v, _mm256_set_epi32(0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL, 0x0C0D0E0FUL, 0x08090A0BUL, 0x04050607UL, 0x00010203UL));
WriteLE32(out + 0 + offset, _mm256_extract_epi32(v, 7));
WriteLE32(out + 32 + offset, _mm256_extract_epi32(v, 6));
WriteLE32(out + 64 + offset, _mm256_extract_epi32(v, 5));
WriteLE32(out + 96 + offset, _mm256_extract_epi32(v, 4));
WriteLE32(out + 128 + offset, _mm256_extract_epi32(v, 3));
WriteLE32(out + 160 + offset, _mm256_extract_epi32(v, 2));
WriteLE32(out + 192 + offset, _mm256_extract_epi32(v, 1));
WriteLE32(out + 224 + offset, _mm256_extract_epi32(v, 0));
}
static void FlowInterExtra_AVX2(
uint8_t *pdst8, int dst_pitch,
const uint8_t *prefB8, const uint8_t *prefF8, int ref_pitch,
const int16_t *VXFullB, const int16_t *VXFullF,
const int16_t *VYFullB, const int16_t *VYFullF,
const uint8_t *MaskB, const uint8_t *MaskF, int VPitch,
int width, int height,
int time256, int nPel,
const int16_t *VXFullBB, const int16_t *VXFullFF,
const int16_t *VYFullBB, const int16_t *VYFullFF) {
const PixelType *prefB = (const PixelType *)prefB8;
const PixelType *prefF = (const PixelType *)prefF8;
PixelType *pdst = (PixelType *)pdst8;
ref_pitch /= sizeof(PixelType);
dst_pitch /= sizeof(PixelType);
int nPelLog = ilog2(nPel);
const __m256i dwords_time256 = _mm256_set1_epi32(time256);
const __m256i dwords_256_time256 = _mm256_set1_epi32(256 - time256);
const __m256i dwords_ref_pitch = _mm256_set1_epi32(ref_pitch);
const __m256i dwords_hoffsets = _mm256_set_epi32(7 << nPelLog, 6 << nPelLog, 5 << nPelLog, 4 << nPelLog, 3 << nPelLog, 2 << nPelLog, 1 << nPelLog, 0);
const int pixels_per_iteration = 8;
const int width_avx2 = width & ~(pixels_per_iteration - 1);
for (int h = 0; h < height; h++) {
for (int w = 0; w < width_avx2; w += pixels_per_iteration)
FlowInterExtra_8px_AVX2(w, pdst, prefB, prefF, VXFullB, VXFullF, VYFullB, VYFullF, MaskB, MaskF, nPelLog, VXFullBB, VXFullFF, VYFullBB, VYFullFF, dwords_time256, dwords_256_time256, dwords_ref_pitch, dwords_hoffsets);
if (width_avx2 < width)
FlowInterExtra_8px_AVX2(width - pixels_per_iteration, pdst, prefB, prefF, VXFullB, VXFullF, VYFullB, VYFullF, MaskB, MaskF, nPelLog, VXFullBB, VXFullFF, VYFullBB, VYFullFF, dwords_time256, dwords_256_time256, dwords_ref_pitch, dwords_hoffsets);
pdst += dst_pitch;
prefB += ref_pitch << nPelLog;
prefF += ref_pitch << nPelLog;
VXFullB += VPitch;
VYFullB += VPitch;
VXFullF += VPitch;
VYFullF += VPitch;
MaskB += VPitch;
MaskF += VPitch;
VXFullBB += VPitch;
VYFullBB += VPitch;
VXFullFF += VPitch;
VYFullFF += VPitch;
}
}
// Compare rank with all values currently in the queue. Returns -1 if the value already exists
// or is larger than all values.
// Otherwise, returns the index of the register in which the value should be inserted.
// Mask is replicated to both lanes, so it can be used for both value and rank lane.
int PriorityQueue_AVX2::compare(__m256i mrank, int &field, __m256i >mask)
{
static const __m256i eq4mask = _mm256_set_epi32(0, 0, 0, 0, -1, -1, -1, -1);
__m256i eq, eq4;
int reg, mask;
// Because items are sorted in ascending order within each (double) register, the mask after GT
// comparison must be of the form 000...1111, which is one less than a power of two.
{
__m256i r0_7 = _mm256_permute2x128_si256(_rv[1], _rv[0], 0x20); // [0 .. 7]
gtmask = _mm256_cmpgt_epi32(r0_7, mrank);
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask));
eq = _mm256_cmpeq_epi32(r0_7, mrank);
_ASSERTE(((mask + 1) & mask) == 0);
reg = 1;
}
if (!mask) {
__m256i r8_15 = _mm256_permute2x128_si256(_rv[3], _rv[2], 0x20); // [8 .. 15]
gtmask = _mm256_cmpgt_epi32(r8_15, mrank);
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask));
eq = _mm256_or_si256(eq, _mm256_cmpeq_epi32(r8_15, mrank));
_ASSERTE(((mask + 1) & mask) == 0);
reg = 3;
}
if (!mask) {
gtmask = _mm256_cmpgt_epi32(_rv[4], mrank); // [16 .. 19]; don't care about value
eq4 = _mm256_and_si256(eq4mask, _mm256_cmpeq_epi32(mrank, _rv[4])); // .. ditto
mask = _mm256_movemask_ps(_mm256_castsi256_ps(gtmask)) & 0xF; // ignore comparison with values
eq = _mm256_or_si256(eq, eq4);
_ASSERTE(((mask + 1) & mask) == 0);
reg = 4;
}
if (_mm256_movemask_ps(_mm256_castsi256_ps(eq)) != 0)
mask = 0;
if (!mask)
return -1;
// Adjust register according to mask (higher 128-bits i double register: one register lower)
// There is no "previous" register to test against for equality if we need to insert in the
// very first register. Also duplicate the same mask to both lanes.
if (mask > 0xF) {
mask >>= 4;
--reg;
gtmask = _mm256_permute2x128_si256(gtmask, gtmask, 0x11); // replicate high lane to both
}
/* Routine optimized for shuffling a buffer for a type size of 4 bytes. */
static void
shuffle4_avx2(uint8_t* const dest, const uint8_t* const src,
const size_t vectorizable_elements, const size_t total_elements)
{
static const size_t bytesoftype = 4;
size_t i;
int j;
__m256i ymm0[4], ymm1[4];
/* Create the shuffle mask.
NOTE: The XMM/YMM 'set' intrinsics require the arguments to be ordered from
most to least significant (i.e., their order is reversed when compared to
loading the mask from an array). */
const __m256i mask = _mm256_set_epi32(
0x07, 0x03, 0x06, 0x02, 0x05, 0x01, 0x04, 0x00);
for (i = 0; i < vectorizable_elements; i += sizeof(__m256i)) {
/* Fetch 32 elements (128 bytes) then transpose bytes and words. */
for (j = 0; j < 4; j++) {
ymm0[j] = _mm256_loadu_si256((__m256i*)(src + (i * bytesoftype) + (j * sizeof(__m256i))));
ymm1[j] = _mm256_shuffle_epi32(ymm0[j], 0xd8);
ymm0[j] = _mm256_shuffle_epi32(ymm0[j], 0x8d);
ymm0[j] = _mm256_unpacklo_epi8(ymm1[j], ymm0[j]);
ymm1[j] = _mm256_shuffle_epi32(ymm0[j], 0x04e);
ymm0[j] = _mm256_unpacklo_epi16(ymm0[j], ymm1[j]);
}
/* Transpose double words */
for (j = 0; j < 2; j++) {
ymm1[j*2] = _mm256_unpacklo_epi32(ymm0[j*2], ymm0[j*2+1]);
ymm1[j*2+1] = _mm256_unpackhi_epi32(ymm0[j*2], ymm0[j*2+1]);
}
/* Transpose quad words */
for (j = 0; j < 2; j++) {
ymm0[j*2] = _mm256_unpacklo_epi64(ymm1[j], ymm1[j+2]);
ymm0[j*2+1] = _mm256_unpackhi_epi64(ymm1[j], ymm1[j+2]);
}
for (j = 0; j < 4; j++) {
ymm0[j] = _mm256_permutevar8x32_epi32(ymm0[j], mask);
}
/* Store the result vectors */
uint8_t* const dest_for_ith_element = dest + i;
for (j = 0; j < 4; j++) {
_mm256_storeu_si256((__m256i*)(dest_for_ith_element + (j * total_elements)), ymm0[j]);
}
}
}
static inline void do_encode_12bytes(const char (*alphabet)[2], char *out, __m256i chunk)
{
const __m256i shufflemask = _mm256_set_epi8(
-1, 9, 10, 11,
-1, 9, 10, 11,
-1, 6, 7, 8,
-1, 6, 7, 8,
-1, 3, 4, 5,
-1, 3, 4, 5,
-1, 0, 1, 2,
-1, 0, 1, 2
);
const __m256i shifts = _mm256_set_epi32(0, 12, 0, 12, 0, 12, 0, 12);
const __m256i masks = _mm256_set1_epi32(4095);
// convert from big endian and rearrange the bytes
chunk = _mm256_shuffle_epi8(chunk, shufflemask);
chunk = _mm256_srlv_epi32(chunk, shifts);
chunk = _mm256_and_si256(chunk, masks);
// write the two halves to memory
do_encode_6bytes(alphabet, out + 0, _mm256_extracti128_si256(chunk, 0));
do_encode_6bytes(alphabet, out + 8, _mm256_extracti128_si256(chunk, 1));
}
double bst_compute_129_m256_maskstore_root_aligned( void*_bst_obj, double* p, double* q, size_t nn ) {
segments_t* mem = (segments_t*) _bst_obj;
int n, i, r, l_end, j, l_end_pre;
double t, e_tmp;
double* e = mem->e, *w = mem->w;
int* root = mem->r;
__m256d v_tmp;
__m256d v00, v01, v02, v03;
__m256d v10, v11, v12, v13;
__m256d v20, v21, v22, v23;
__m256d v30, v31, v32, v33;
__m256i v_cur_roots;
__m256 v_rootmask0, v_rootmask1;
// initialization
// mem->n = nn;
n = nn; // subtractions with n potentially negative. say hello to all the bugs
int idx1, idx1_root;
int idx2;
int idx3, idx3_root;
int pad_root, pad, pad_r;
idx1 = ((int) mem->e_sz) - 1;
idx1_root = ((int) mem->r_sz);
// the conventio is that iteration i, idx1 points to the first element of line i+1
e[idx1++] = q[n];
// pad contains the padding for row i+1
// for row n it's always 3
pad = 3;
pad_root = 7;
for (i = n-1; i >= 0; --i) {
idx1 -= 2*(n-i)+1 + pad;
idx1_root -= 2*(n-i)+1 + pad_root;
idx2 = idx1 + 1;
e[idx1] = q[i];
w[idx1] = q[i];
for (j = i+1; j < n+1; ++j,++idx2) {
e[idx2] = INFINITY;
w[idx2] = w[idx2-1] + p[j-1] + q[j];
}
idx2 += pad; // padding of line i+1
// idx2 now points to the first element of the next line
idx3 = idx1;
idx3_root = idx1_root;
pad_r = pad;
for (r = i; r < n; ++r) {
pad_r = (pad_r+1)&3; // padding of line r+1
idx1 = idx3;
idx1_root = idx3_root;
l_end = idx2 + (n-r);
// l_end points to the first entry after the current row
e_tmp = e[idx1++];
idx1_root++;
// calculate until a multiple of 8 doubles is left
// 8 = 4 * 2 128-bit vectors
l_end_pre = idx2 + ((n-r)&15);
for( ; (idx2 < l_end_pre) && (idx2 < l_end); ++idx2 ) {
t = e_tmp + e[idx2] + w[idx1];
if (t < e[idx1]) {
e[idx1] = t;
root[idx1_root] = r;
}
idx1++;
idx1_root++;
}
v_tmp = _mm256_set_pd( e_tmp, e_tmp, e_tmp, e_tmp );
// execute the shit for 4 vectors of size 2
v_cur_roots = _mm256_set_epi32(r, r, r, r, r, r, r, r);
for( ; idx2 < l_end; idx2 += 16 ) {
v01 = _mm256_load_pd( &w[idx1 ] );
v11 = _mm256_load_pd( &w[idx1+ 4] );
v21 = _mm256_load_pd( &w[idx1+ 8] );
v31 = _mm256_load_pd( &w[idx1+12] );
v00 = _mm256_load_pd( &e[idx2 ] );
v01 = _mm256_add_pd( v01, v_tmp );
v10 = _mm256_load_pd( &e[idx2+ 4] );
v11 = _mm256_add_pd( v11, v_tmp );
v20 = _mm256_load_pd( &e[idx2+ 8] );
v21 = _mm256_add_pd( v21, v_tmp );
v30 = _mm256_load_pd( &e[idx2+12] );
v31 = _mm256_add_pd( v31, v_tmp );
v01 = _mm256_add_pd( v01, v00 );
v03 = _mm256_load_pd( &e[idx1 ] );
v11 = _mm256_add_pd( v11, v10 );
v13 = _mm256_load_pd( &e[idx1+ 4] );
v21 = _mm256_add_pd( v21, v20 );
v23 = _mm256_load_pd( &e[idx1+ 8] );
v31 = _mm256_add_pd( v31, v30 );
v33 = _mm256_load_pd( &e[idx1+12] );
v02 = _mm256_cmp_pd( v01, v03, _CMP_LT_OQ );
v12 = _mm256_cmp_pd( v11, v13, _CMP_LT_OQ );
v22 = _mm256_cmp_pd( v21, v23, _CMP_LT_OQ );
v32 = _mm256_cmp_pd( v31, v33, _CMP_LT_OQ );
_mm256_maskstore_pd( &e[idx1 ],
_mm256_castpd_si256( v02 ), v01 );
_mm256_maskstore_pd( &e[idx1+ 4],
_mm256_castpd_si256( v12 ), v11 );
v_rootmask0 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v02)),
_mm256_cvtpd_ps(v12) , 1
);
_mm256_maskstore_pd( &e[idx1+ 8],
_mm256_castpd_si256( v22 ), v21 );
_mm256_maskstore_pd( &e[idx1+12],
_mm256_castpd_si256( v32 ), v31 );
v_rootmask1 = _mm256_insertf128_ps(
_mm256_castps128_ps256(
_mm256_cvtpd_ps(v22)),
_mm256_cvtpd_ps(v32) , 1
);
_mm256_maskstore_ps( &root[idx1_root ],
_mm256_castps_si256( v_rootmask0 ),
_mm256_castsi256_ps( v_cur_roots ) );
_mm256_maskstore_ps( &root[idx1_root + 8],
_mm256_castps_si256( v_rootmask1 ),
_mm256_castsi256_ps( v_cur_roots ) );
idx1 += 16;
idx1_root += 16;
}
idx2 += pad_r;
idx3++;
idx3_root++;
}
pad = (pad -1)&3;
pad_root = (pad_root-1)&7;
}
// the index of the last item of the first row is ((n/4)+1)*4-1, due to the padding
// if n is even, the total number of entries in the first
// row of the table is odd, so we need padding
return e[ ((n/4)+1)*4 - 1 ];
}
parasail_result_t *result = parasail_result_new();
#endif
#endif
int32_t i = 0;
int32_t j = 0;
int32_t end_query = 0;
int32_t end_ref = 0;
int32_t score = NEG_INF;
__m256i vNegInf = _mm256_set1_epi32(NEG_INF);
__m256i vOpen = _mm256_set1_epi32(open);
__m256i vGap = _mm256_set1_epi32(gap);
__m256i vOne = _mm256_set1_epi32(1);
__m256i vN = _mm256_set1_epi32(N);
__m256i vGapN = _mm256_set1_epi32(gap*N);
__m256i vNegOne = _mm256_set1_epi32(-1);
__m256i vI = _mm256_set_epi32(0,1,2,3,4,5,6,7);
__m256i vJreset = _mm256_set_epi32(0,-1,-2,-3,-4,-5,-6,-7);
__m256i vMax = vNegInf;
__m256i vILimit = _mm256_set1_epi32(s1Len);
__m256i vILimit1 = _mm256_sub_epi32(vILimit, vOne);
__m256i vJLimit = _mm256_set1_epi32(s2Len);
__m256i vJLimit1 = _mm256_sub_epi32(vJLimit, vOne);
__m256i vIBoundary = _mm256_set_epi32(
-open-0*gap,
-open-1*gap,
-open-2*gap,
-open-3*gap,
-open-4*gap,
-open-5*gap,
-open-6*gap,
-open-7*gap
foo (int x1, int x2, int x3, int x4,
int x5, int x6, int x7, int x8)
{
return _mm256_set_epi32 (x1, x2, x3, x4, x5, x6, x7, x8);
}
INLINE avxi( StepTy ) : m256(_mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0)) {}
#endif
#endif
int32_t i = 0;
int32_t j = 0;
int32_t end_query = 0;
int32_t end_ref = 0;
int32_t score = NEG_INF;
__m256i vNegInf = _mm256_set1_epi32(NEG_INF);
__m256i vNegInf0 = _mm256_srli_si256_rpl(vNegInf, 4); /* shift in a 0 */
__m256i vOpen = _mm256_set1_epi32(open);
__m256i vGap = _mm256_set1_epi32(gap);
__m256i vZero = _mm256_set1_epi32(0);
__m256i vOne = _mm256_set1_epi32(1);
__m256i vN = _mm256_set1_epi32(N);
__m256i vNegOne = _mm256_set1_epi32(-1);
__m256i vI = _mm256_set_epi32(0,1,2,3,4,5,6,7);
__m256i vJreset = _mm256_set_epi32(0,-1,-2,-3,-4,-5,-6,-7);
__m256i vMax = vNegInf;
__m256i vEndI = vNegInf;
__m256i vEndJ = vNegInf;
__m256i vILimit = _mm256_set1_epi32(s1Len);
__m256i vJLimit = _mm256_set1_epi32(s2Len);
/* convert _s1 from char to int in range 0-23 */
for (i=0; i<s1Len; ++i) {
s1[i] = matrix->mapper[(unsigned char)_s1[i]];
}
/* pad back of s1 with dummy values */
for (i=s1Len; i<s1Len_PAD; ++i) {
s1[i] = 0; /* point to first matrix row because we don't care */
void Decoder::ADMMDecoder_deg_6_7_2_3_6()
{
int maxIter = maxIteration;
float mu = 5.5f;
float tableau[12] = { 0.0f };
if ((mBlocklength == 576) && (mNChecks == 288))
{
mu = 3.37309f;//penalty
tableau[2] = 0.00001f;
tableau[3] = 2.00928f;
tableau[6] = 4.69438f;
}
else if((mBlocklength == 2304) && (mNChecks == 1152) )
{
mu = 3.81398683f;//penalty
tableau[2] = 0.29669288f;
tableau[3] = 0.46964023f;
tableau[6] = 3.19548154f;
}
else
{
mu = 5.5;//penalty
tableau[2] = 0.8f;
tableau[3] = 0.8f;
tableau[6] = 0.8f;
}
const float rho = 1.9f; //over relaxation parameter;
const float un_m_rho = 1.0 - rho;
const auto _rho = _mm256_set1_ps( rho );
const auto _un_m_rho = _mm256_set1_ps( un_m_rho );
float tableaX[12];
//
// ON PRECALCULE LES CONSTANTES
//
#pragma unroll
for (int i = 0; i < 7; i++)
{
tableaX[i] = tableau[ i ] / mu;
}
const auto t_mu = _mm256_set1_ps ( mu );
const auto t2_amu = _mm256_set1_ps ( tableau[ 2 ] / mu );
const auto t3_amu = _mm256_set1_ps ( tableau[ 3 ] / mu );
const auto t6_amu = _mm256_set1_ps ( tableau[ 6 ] / mu );
const auto t2_2amu = _mm256_set1_ps ( 2.0f * tableau[ 2 ] / mu );
const auto t3_2amu = _mm256_set1_ps ( 2.0f * tableau[ 3 ] / mu );
const auto t6_2amu = _mm256_set1_ps ( 2.0f * tableau[ 6 ] / mu );
const auto t2_deg = _mm256_set1_ps ( 2.0f );
const auto t3_deg = _mm256_set1_ps ( 3.0f );
const auto t6_deg = _mm256_set1_ps ( 6.0f );
const auto zero = _mm256_set1_ps ( 0.0f );
const auto un = _mm256_set1_ps ( 1.0f );
const __m256 a = _mm256_set1_ps ( 0.0f );
const __m256 b = _mm256_set1_ps ( 0.5f );
//////////////////////////////////////////////////////////////////////////////////////
#pragma unroll
for( int j = 0; j < _mPCheckMapSize; j+=8 )
{
_mm256_store_ps(&Lambda [j], a);
_mm256_store_ps(&zReplica[j], b);
_mm256_store_ps(&latestProjVector[j], b);
}
//////////////////////////////////////////////////////////////////////////////////////
for(int i = 0; i < maxIter; i++)
{
int ptr = 0;
mIteration = i + 1;
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto start = timer();
#endif
//
// VN processing kernel
//
#pragma unroll
for (int j = 0; j < _mBlocklength; j++)
{
const int degVn = VariableDegree[j];
float M[8] __attribute__((aligned(64)));
if( degVn == 2 ){
#if 1
const int dVN = 2;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t2_amu), _mm256_sub_ps(t2_deg, t2_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 2;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 3 ){
#if 1
const int dVN = 3;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t3_amu), _mm256_sub_ps(t3_deg, t3_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 3;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 6 ){
#if 1
const int dVN = 6;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t6_amu), _mm256_sub_ps(t6_deg, t6_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 6;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}
}
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
t_vn += (timer() - start);
#endif
//
// CN processing kernel
//
int CumSumCheckDegree = 0; // cumulative position of currect edge in factor graph
int allVerified = 0;
float vector_before_proj[8] __attribute__((aligned(64)));
const auto zero = _mm256_set1_ps ( 0.0f );
const auto mask_6 = _mm256_set_epi32(0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto mask_7 = _mm256_set_epi32(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto dot5 = _mm256_set1_ps( 0.5f );
//
// MEASURE OF THE CN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto starT = timer();
#endif
const auto seuilProj = _mm256_set1_ps( 1e-5f );
for(int j = 0; j < _mNChecks; j++)
{
if( CheckDegree[j] == 6 ){
const int cDeg6 = 0x3F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_6), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
int test = (_mm256_movemask_ps( synd ) & cDeg6); // deg 6
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps (xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg6) == 0x00; // degree 6
if( skip == false )
{
const auto _ztemp = mp.projection_deg6( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda[CumSumCheckDegree], mask_6, mLambda);
_mm256_maskstore_ps(&zReplica[CumSumCheckDegree], mask_6, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_6, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 6;
}else if( CheckDegree[j] == 7 )
{
const int cDeg7 = 0x7F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_7), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
const int test = (_mm256_movemask_ps( synd ) & cDeg7); // deg 7
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps ( xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg7) == 0x00; // degree 7
if( skip == false )
{
const auto _ztemp = mp.projection_deg7( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda [CumSumCheckDegree], mask_7, mLambda);
_mm256_maskstore_ps(&zReplica [CumSumCheckDegree], mask_7, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_7, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 7;
}else{
exit( 0 );
}
}
//
// MEASURE OF THE CN LOOP EXECUTION TIME
//
#ifdef PROFILE_ON
t_cn += (timer() - starT);
#endif
#ifdef PROFILE_ON
t_ex += 1;
//FILE *ft=fopen("time.txt","a");
//fprintf(ft,"%d \n", t_cn/t_ex);
//fprintf(ft,"%d %d %d \n", t_cn, t_vn, t_pj);
//fclose(ft);
#endif
if(allVerified == 0)
{
mAlgorithmConverge = true;
mValidCodeword = true;
break;
}
}
//
// MEASURE OF THE NUMBER OF EXECUTION
//
// #ifdef PROFILE_ON
// t_ex += 1;
// #endif
}
_mm_set_epi32(0x00200000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x00400000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x00800000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x01000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x02000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x04000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x08000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x10000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x20000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x40000000, 0x00000000, 0x00000000, 0x00000000),
_mm_set_epi32(0x80000000, 0x00000000, 0x00000000, 0x00000000)
};
#ifdef AVX2
__m256i sseMasks256[256] = {
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000001),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000002),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000004),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000008),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000010),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000020),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000040),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000080),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000100),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000200),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000400),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000800),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00001000),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00002000),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00004000),
_mm256_set_epi32(0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00008000),
void run_softmax_int32_float_work_item_latency(nn_workload_item *const work_item)
{
nn_workload_data_t *input_view = work_item->input[0]->output;
const auto &arguments = work_item->arguments.forward_softmax_fixedpoint;
const auto input_width = input_view->parent->lengths.t[NN_DATA_COORD_z] * input_view->parent->lengths.t[NN_DATA_COORD_p];
const auto output_width = work_item->output->view_end.t[NN_DATA_COORD_x] - work_item->output->view_begin.t[NN_DATA_COORD_x] + 1;
const auto num_full_blocks = output_width / C_data_stride;
const auto partial_block_size = (output_width / C_simd_width) % C_max_acc;
const auto subsimd_block_size = output_width % C_simd_width;
const auto output_view_start = work_item->output->view_begin.t[NN_DATA_COORD_x];
const auto input_view_start = input_view->view_begin.t[NN_DATA_COORD_z] * input_view->parent->lengths.t[NN_DATA_COORD_p];
const auto out_fraction = arguments.input_fraction;
float * input_f = (float*)_mm_malloc(input_width * sizeof(float), 64);
auto input_buffer = &static_cast<int32_t*>(input_view->parent->data_buffer)[input_view_start];
auto shift = out_fraction;
if (shift > 0)
{
for (uint32_t i = 0; i < input_width; i++)
input_f[i] = (float)(input_buffer[i]) / (1 << shift);
}
else if (shift < 0)
{
for (uint32_t i = 0; i < input_width; i++)
input_f[i] = (float)(input_buffer[i]) * (1 << -shift);
}
else
{
for (uint32_t i = 0; i < input_width; i++)
input_f[i] = (float)(input_buffer[i]);
}
__m256 acc_sum = _mm256_setzero_ps();
float subsimd_sum = 0.0f;
{
auto input_buffer = input_f;
auto output_buffer = &static_cast<float*>(work_item->output->parent->data_buffer)[output_view_start];
for (auto block = 0u; block < num_full_blocks; ++block)
{
// Run computation.
softmax_compute_block<C_max_acc>(input_buffer, output_buffer, acc_sum);
}
switch (partial_block_size)
{
case 0: break;
case 1: softmax_compute_block< 1>(input_buffer, output_buffer, acc_sum); break;
case 2: softmax_compute_block< 2>(input_buffer, output_buffer, acc_sum); break;
case 3: softmax_compute_block< 3>(input_buffer, output_buffer, acc_sum); break;
case 4: softmax_compute_block< 4>(input_buffer, output_buffer, acc_sum); break;
case 5: softmax_compute_block< 5>(input_buffer, output_buffer, acc_sum); break;
case 6: softmax_compute_block< 6>(input_buffer, output_buffer, acc_sum); break;
case 7: softmax_compute_block< 7>(input_buffer, output_buffer, acc_sum); break;
case 8: softmax_compute_block< 8>(input_buffer, output_buffer, acc_sum); break;
case 9: softmax_compute_block< 9>(input_buffer, output_buffer, acc_sum); break;
case 10: softmax_compute_block<10>(input_buffer, output_buffer, acc_sum); break;
case 11: softmax_compute_block<11>(input_buffer, output_buffer, acc_sum); break;
case 12: softmax_compute_block<12>(input_buffer, output_buffer, acc_sum); break;
case 13: softmax_compute_block<13>(input_buffer, output_buffer, acc_sum); break;
case 14: softmax_compute_block<14>(input_buffer, output_buffer, acc_sum); break;
default: NN_UNREACHABLE_CODE;
}
switch (subsimd_block_size)
{
case 0: break;
case 1: softmax_compute_subsimd<1>(input_buffer, output_buffer, subsimd_sum); break;
case 2: softmax_compute_subsimd<2>(input_buffer, output_buffer, subsimd_sum); break;
case 3: softmax_compute_subsimd<3>(input_buffer, output_buffer, subsimd_sum); break;
case 4: softmax_compute_subsimd<4>(input_buffer, output_buffer, subsimd_sum); break;
case 5: softmax_compute_subsimd<5>(input_buffer, output_buffer, subsimd_sum); break;
case 6: softmax_compute_subsimd<6>(input_buffer, output_buffer, subsimd_sum); break;
case 7: softmax_compute_subsimd<7>(input_buffer, output_buffer, subsimd_sum); break;
default: NN_UNREACHABLE_CODE;
}
}
{
__m256 intermediate_sum = _mm256_hadd_ps(acc_sum, acc_sum);
intermediate_sum = _mm256_permutevar8x32_ps(intermediate_sum, _mm256_set_epi32(0, 1, 4, 5, 2, 3, 6, 7));
intermediate_sum = _mm256_hadd_ps(intermediate_sum, intermediate_sum);
intermediate_sum = _mm256_hadd_ps(intermediate_sum, intermediate_sum);
acc_sum = _mm256_add_ps(intermediate_sum, _mm256_set1_ps(subsimd_sum));
subsimd_sum = _mm_cvtss_f32(_mm256_extractf128_ps(acc_sum, 0));
acc_sum = _mm256_div_ps(_mm256_set1_ps(1.0f), acc_sum);
subsimd_sum = 1.0f / subsimd_sum;
}
{
auto output_buffer = &static_cast<float*>(work_item->output->parent->data_buffer)[output_view_start];
for (auto block = 0u; block < num_full_blocks; ++block)
{
// Run computation.
softmax_finalize_block<C_max_acc>(output_buffer, acc_sum);
}
switch (partial_block_size)
{
case 0: break;
case 1: softmax_finalize_block< 1>(output_buffer, acc_sum); break;
case 2: softmax_finalize_block< 2>(output_buffer, acc_sum); break;
case 3: softmax_finalize_block< 3>(output_buffer, acc_sum); break;
case 4: softmax_finalize_block< 4>(output_buffer, acc_sum); break;
case 5: softmax_finalize_block< 5>(output_buffer, acc_sum); break;
case 6: softmax_finalize_block< 6>(output_buffer, acc_sum); break;
case 7: softmax_finalize_block< 7>(output_buffer, acc_sum); break;
case 8: softmax_finalize_block< 8>(output_buffer, acc_sum); break;
case 9: softmax_finalize_block< 9>(output_buffer, acc_sum); break;
case 10: softmax_finalize_block<10>(output_buffer, acc_sum); break;
case 11: softmax_finalize_block<11>(output_buffer, acc_sum); break;
case 12: softmax_finalize_block<12>(output_buffer, acc_sum); break;
case 13: softmax_finalize_block<13>(output_buffer, acc_sum); break;
case 14: softmax_finalize_block<14>(output_buffer, acc_sum); break;
default: NN_UNREACHABLE_CODE;
}
switch (subsimd_block_size)
{
case 0: break;
case 1: softmax_finalize_subsimd<1>(output_buffer, subsimd_sum); break;
case 2: softmax_finalize_subsimd<2>(output_buffer, subsimd_sum); break;
case 3: softmax_finalize_subsimd<3>(output_buffer, subsimd_sum); break;
case 4: softmax_finalize_subsimd<4>(output_buffer, subsimd_sum); break;
case 5: softmax_finalize_subsimd<5>(output_buffer, subsimd_sum); break;
case 6: softmax_finalize_subsimd<6>(output_buffer, subsimd_sum); break;
case 7: softmax_finalize_subsimd<7>(output_buffer, subsimd_sum); break;
default: NN_UNREACHABLE_CODE;
}
}
_mm_free(input_f);
}
INLINE avxi( int a, int b, int c, int d, int e, int f, int g, int h ) : m256(_mm256_set_epi32(h, g, f, e, d, c, b, a)) {}