admin/263-0007-00L-Advanced-Systems-Lab
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Perf plot

Browse Source
This commit is contained in:
[Quy Anh] «Elliot» Nguyen
2026-06-18 11:59:50 +02:00
committed by Linus Baumberger
parent ca969564a6
commit e7af3103e8
153 changed files with 22042 additions and 24220 deletions
Download Patch File Download Diff File Expand all files Collapse all files
variants/variant_inlining_8_SoA_AVX512_polyexp_moreavx.cpp
+120 -10
View File
@@ -13,6 +13,8 @@
#define SEED 42
#define BOARD_WIDTH 10.0
#define BOARD_LENGTH 50.0
#define BOARD_LENGTH_MINUS_ONE BOARD_LENGTH - 1.0
#define BOARD_LENGTH_MINUS_ONE_POINT_FIVE BOARD_LENGTH - 1.5
#define WALL_EPS 1e-3
#define BOARD_WIDTH_MINUS_WALL_EPS 9.999
@@ -162,23 +164,42 @@ Pedestrian *init_pedestrians(int n_p) {
Board init_board_without_obstacles() { return Board(BOARD_LENGTH, BOARD_WIDTH, 0, nullptr); }
// # 128 MULTs, 112 ADDs
static inline __m512d _avx512_exp_pd(__m512d x) {
const __m512d log2e = _mm512_set1_pd(1.4426950408889634);
// # 8 MULTs
__m512d nf = _mm512_roundscale_pd(_mm512_mul_pd(x, log2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
const __m512d ln2 = _mm512_set1_pd(0.6931471805599453);
// # 8 MULTs, 8 ADDs
__m512d r = _mm512_fnmadd_pd(nf, ln2, x);
const __m512d c7 = _mm512_set1_pd(1.0 / 5040.0);
const __m512d c6 = _mm512_set1_pd(1.0 / 720.0);
const __m512d c5 = _mm512_set1_pd(1.0 / 120.0);
const __m512d c4 = _mm512_set1_pd(1.0 / 24.0);
const __m512d c3 = _mm512_set1_pd(1.0 / 6.0);
const __m512d c2 = _mm512_set1_pd(0.5);
const __m512d c1 = _mm512_set1_pd(1.0);
const __m512d c0 = _mm512_set1_pd(1.0);
const __m512d c13 = _mm512_set1_pd(c13_scalar);
const __m512d c12 = _mm512_set1_pd(c12_scalar);
const __m512d c11 = _mm512_set1_pd(c11_scalar);
const __m512d c10 = _mm512_set1_pd(c10_scalar);
const __m512d c9 = _mm512_set1_pd(c9_scalar);
const __m512d c8 = _mm512_set1_pd(c8_scalar);
const __m512d c7 = _mm512_set1_pd(c7_scalar);
const __m512d c6 = _mm512_set1_pd(c6_scalar);
const __m512d c5 = _mm512_set1_pd(c5_scalar);
const __m512d c4 = _mm512_set1_pd(c4_scalar);
const __m512d c3 = _mm512_set1_pd(c3_scalar);
const __m512d c2 = _mm512_set1_pd(c2_scalar);
const __m512d c1 = _mm512_set1_pd(c1_scalar);
const __m512d c0 = _mm512_set1_pd(c0_scalar);
__m512d p = c7;
__m512d p = c13;
// # 104 MULTs, 104 ADDs
p = _mm512_fmadd_pd(p, r, c12);
p = _mm512_fmadd_pd(p, r, c11);
p = _mm512_fmadd_pd(p, r, c10);
p = _mm512_fmadd_pd(p, r, c9);
p = _mm512_fmadd_pd(p, r, c8);
p = _mm512_fmadd_pd(p, r, c7);
p = _mm512_fmadd_pd(p, r, c6);
p = _mm512_fmadd_pd(p, r, c5);
p = _mm512_fmadd_pd(p, r, c4);
@@ -204,6 +225,7 @@ static inline __m512d _avx512_exp_pd(__m512d x) {
__m512d pow2n = _mm512_castsi512_pd(scaled);
// # 8 MULTs
return _mm512_mul_pd(p, pow2n);
}
@@ -240,6 +262,26 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
const __m512d v_np = _mm512_set1_pd(n_p);
// # ADDs: n_iter * n_p * 20 +
// # n_iter * n_p * (n_p_pad / 8) * 232 +
// # n_iter * n_p * board.n_obstacles * 10 +
// # n_iter * (n_p_pad / 8) * 56
// # MULTs: n_iter * n_p * 4 +
// # n_iter * n_p * (n_p_pad / 8) * 320 +
// # n_iter * n_p * board.n_obstacles * 13 +
// # n_iter * (n_p_pad / 8) * 120
// # DIVs: n_iter * n_p * (n_p_pad / 8) * 8 +
// # n_iter * n_p * board.n_obstacles +
// # n_iter * (n_p_pad / 8) * 16
// # SQRTs: n_iter * n_p * (n_p_pad / 8) * 24 +
// # n_iter * n_p * board.n_obstacles +
// # n_iter * (n_p_pad / 8) * 24
// # EXPs: n_iter * n_p * board.n_obstacles
for (int t = 0; t < n_iter; t++) {
for (int i = 0; i < n_p; i++) {
@@ -251,6 +293,7 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
double curr_vel_x_i = curr_vel_x[i];
double curr_vel_y_i = curr_vel_y[i];
// # 4 MULTs, 2 ADDs
double f_alpha_dt_x = (e_x_i * desired_v0_i - curr_vel_x_i) * dt_times_tau_alpha_inv;
double f_alpha_dt_y = (e_y_i * desired_v0_i - curr_vel_y_i) * dt_times_tau_alpha_inv;
@@ -264,6 +307,8 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512d acc_y = _mm512_setzero_pd();
for (int j = 0; j < n_p_pad; j += 8) {
// # 8 ADDs
__m512d j_idx = _mm512_add_pd(_mm512_set1_pd((double)j), lane_off);
__mmask8 m_neq_i = _mm512_cmp_pd_mask(j_idx, v_i, _CMP_NEQ_OQ);
@@ -276,35 +321,63 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512d sbe_jy = _mm512_loadu_pd(&sb_e_y[j]);
__m512d sb_jsq = _mm512_loadu_pd(&sb_sq[j]);
// # 16 ADDs
__m512d r_ab_x = _mm512_sub_pd(pix, pjx);
__m512d r_ab_y = _mm512_sub_pd(piy, pjy);
// # 16 MULTs, 8 ADDs
__m512d r_ab_n2 = _mm512_fmadd_pd(r_ab_x, r_ab_x, _mm512_mul_pd(r_ab_y, r_ab_y));
// # 8 SQRTs
__m512d r_ab_n = _mm512_sqrt_pd(r_ab_n2);
// # 16 ADDs
__m512d r_ab2_x = _mm512_sub_pd(r_ab_x, sbe_jx);
__m512d r_ab2_y = _mm512_sub_pd(r_ab_y, sbe_jy);
// # 16 MULTs, 8 ADDs
__m512d r_ab2_n2 = _mm512_fmadd_pd(r_ab2_x, r_ab2_x, _mm512_mul_pd(r_ab2_y, r_ab2_y));
// # 8 SQRTs
__m512d r_ab2_n = _mm512_sqrt_pd(r_ab2_n2);
// # 8 ADDs
__m512d r_sum = _mm512_add_pd(r_ab_n, r_ab2_n);
// # 8 MULTs, 8 ADDs
__m512d two_b_sq = _mm512_fmsub_pd(r_sum, r_sum, sb_jsq);
// # 8 SQRTs
__m512d two_b = _mm512_sqrt_pd(two_b_sq);
// # 8 MULTs
__m512d e_arg = _mm512_mul_pd(two_b, v_neg_2s_inv);
// # 128 MULTs, 112 ADDs
__m512d e_val = _avx512_exp_pd(e_arg);
// # 32 MULTs
__m512d numer = _mm512_mul_pd(_mm512_mul_pd(r_sum, v_neg_v0_div_2s), e_val);
__m512d denom = _mm512_mul_pd(_mm512_mul_pd(two_b, r_ab_n), r_ab2_n);
// # 8 DIVs
__m512d smc = _mm512_div_pd(numer, denom);
// # 16 MULTs
__m512d sm_r_ab = _mm512_mul_pd(smc, r_ab2_n);
__m512d sm_r_ab2 = _mm512_mul_pd(smc, r_ab_n);
// # 32 MULTs, 16 ADDs
__m512d f_x = _mm512_fmadd_pd(r_ab2_x, sm_r_ab2, _mm512_mul_pd(r_ab_x, sm_r_ab));
__m512d f_y = _mm512_fmadd_pd(r_ab2_y, sm_r_ab2, _mm512_mul_pd(r_ab_y, sm_r_ab));
// # 24 MULTs, 8 ADDs
__m512d a = _mm512_mul_pd(_mm512_fmadd_pd(eix, f_x, _mm512_mul_pd(eiy, f_y)), v_cos_phi_inv);
// # 8 MULTs
__m512d a_sq = _mm512_mul_pd(a, a);
// # 16 MULTs, 8 ADDs
__m512d b_sq = _mm512_fmadd_pd(f_x, f_x, _mm512_mul_pd(f_y, f_y));
__mmask8 k_ge = _mm512_cmp_pd_mask(a, v_zero, _CMP_GE_OQ);
@@ -314,18 +387,23 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512d f_x_safe = _mm512_maskz_mov_pd(active, f_x);
__m512d f_y_safe = _mm512_maskz_mov_pd(active, f_y);
// # 16 MULTs, 16 ADDs
acc_x = _mm512_fmadd_pd(f_x_safe, w_val, acc_x);
acc_y = _mm512_fmadd_pd(f_y_safe, w_val, acc_y);
}
// # 16 ADDS
f_alpha_dt_x += _mm512_reduce_add_pd(acc_x);
f_alpha_dt_y += _mm512_reduce_add_pd(acc_y);
for (int k = 0; k < board.n_obstacles; k++) {
// # 2 ADDs
double v1_x = pos_x_i - board.obstacles[k].start_x;
double v1_y = pos_y_i - board.obstacles[k].start_y;
// # 3 MULTs, 1 ADD
double projection = (v1_x * board.obstacles[k].obstacle_v_x + v1_y * board.obstacles[k].obstacle_v_y) * board.obstacles[k].length_sq_inv;
if (projection < 0.0)
@@ -333,11 +411,17 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
else if (projection > 1.0)
projection = 1.0;
// # 4 ADDs, 2 MULTs
double r_aB_x = pos_x_i - (board.obstacles[k].start_x + board.obstacles[k].obstacle_v_x * projection);
double r_aB_y = pos_y_i - (board.obstacles[k].start_y + board.obstacles[k].obstacle_v_y * projection);
// # 1 SQRT, 2 MULTs, 1 ADD
double r_aB_norm = sqrt(r_aB_x * r_aB_x + r_aB_y * r_aB_y);
// # 2 MULTs, 1 EXP, 1 DIV
double scalar_mult = (NEG_U0_AB_div_R * exp(r_aB_norm * NEG_R_INV)) / r_aB_norm;
// # 2 MULTs
double f_alpha_B_neg_x = r_aB_x * scalar_mult;
double f_alpha_B_neg_y = r_aB_y * scalar_mult;
@@ -346,10 +430,12 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
if (e_dot_f < cos_phi)
w_val = c_neg_times_dt;
// # 2 MULTs, 2 ADDs
f_alpha_dt_x += f_alpha_B_neg_x * w_val;
f_alpha_dt_y += f_alpha_B_neg_y * w_val;
}
// # 2 ADDs
w_x[i] += f_alpha_dt_x;
w_y[i] += f_alpha_dt_y;
}
@@ -367,19 +453,27 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m256i color_i_vec = _mm256_loadu_si256((__m256i *)&color[i]);
// # 16 MULTs, 8 ADDs
__m512d w_norm_sq = _mm512_fmadd_pd(w_x_i_vec, w_x_i_vec, _mm512_mul_pd(w_y_i_vec, w_y_i_vec));
// # 8 SQRTs
__m512d w_norm = _mm512_sqrt_pd(w_norm_sq);
// # 8 DIVs
__m512d correction_factor = _mm512_min_pd(_mm512_set1_pd(1.0), _mm512_div_pd(v_max_i_vec, w_norm));
// # 16 MULTs
__m512d curr_vel_x_i_vec = _mm512_mul_pd(w_x_i_vec, correction_factor);
__m512d curr_vel_y_i_vec = _mm512_mul_pd(w_y_i_vec, correction_factor);
__m512d dt_vec = _mm512_set1_pd(dt);
// # 16 MULTs, 16 ADDs
pos_x_i_vec = _mm512_fmadd_pd(curr_vel_x_i_vec, dt_vec, pos_x_i_vec);
pos_y_i_vec = _mm512_fmadd_pd(curr_vel_y_i_vec, dt_vec, pos_y_i_vec);
__mmask8 cond_1_left = _mm256_cmpeq_epi32_mask(color_i_vec, _mm256_setzero_si256());
__mmask8 cond_1_right = _mm512_cmp_pd_mask(pos_x_i_vec, _mm512_set1_pd(BOARD_LENGTH - 1.5), _CMP_GT_OQ);
__mmask8 cond_1_right = _mm512_cmp_pd_mask(pos_x_i_vec, _mm512_set1_pd(BOARD_LENGTH_MINUS_ONE_POINT_FIVE), _CMP_GT_OQ);
__mmask8 cond_1 = cond_1_left & cond_1_right;
pos_x_i_vec = _mm512_mask_blend_pd(cond_1, pos_x_i_vec, _mm512_set1_pd(1.0));
@@ -404,16 +498,30 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
pos_y_i_vec = _mm512_max_pd(pos_y_i_vec, _mm512_set1_pd(WALL_EPS));
pos_y_i_vec = _mm512_min_pd(pos_y_i_vec, _mm512_set1_pd(BOARD_WIDTH_MINUS_WALL_EPS));
// # 16 ADDs
__m512d ex_vec = _mm512_sub_pd(destination_x_i_vec, pos_x_i_vec);
__m512d ey_vec = _mm512_sub_pd(destination_y_i_vec, pos_y_i_vec);
// # 16 MULTs, 8 ADDs
__m512d e_norm_sq = _mm512_fmadd_pd(ex_vec, ex_vec, _mm512_mul_pd(ey_vec, ey_vec));
// # 8 SQRTs
__m512d e_norm = _mm512_sqrt_pd(e_norm_sq);
// # 8 DIVs
__m512d e_norm_inv = _mm512_div_pd(_mm512_set1_pd(1.0), e_norm);
// # 16 MULT
__m512d e_x_i_vec = _mm512_mul_pd(ex_vec, e_norm_inv);
__m512d e_y_i_vec = _mm512_mul_pd(ey_vec, e_norm_inv);
// # 16 MULTs, 8 ADDs
__m512d vel_norm_sq = _mm512_fmadd_pd(curr_vel_x_i_vec, curr_vel_x_i_vec, _mm512_mul_pd(curr_vel_y_i_vec, curr_vel_y_i_vec));
// # 8 MULTs
__m512d sb_sq_i_vec = _mm512_mul_pd(vel_norm_sq, _mm512_set1_pd(delta_t_sq));
// # 8 SQRTs
__m512d vel_norm_dt = _mm512_sqrt_pd(sb_sq_i_vec);
_mm512_storeu_pd(&curr_vel_x[i], curr_vel_x_i_vec);
@@ -426,6 +534,8 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
_mm512_storeu_pd(&e_y[i], e_y_i_vec);
_mm512_storeu_pd(&destination_y[i], destination_y_i_vec);
_mm512_storeu_pd(&sb_sq[i], sb_sq_i_vec);
// # 16 MULTs
_mm512_storeu_pd(&sb_e_x[i], _mm512_mul_pd(e_x_i_vec, vel_norm_dt));
_mm512_storeu_pd(&sb_e_y[i], _mm512_mul_pd(e_y_i_vec, vel_norm_dt));
}