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_float_inlining_SoA_AVX512_polyexp_moreavx.cpp
+142 -13
View File
@@ -13,6 +13,8 @@
#define SEED 42
#define BOARD_WIDTH 10.0f
#define BOARD_LENGTH 50.0f
#define BOARD_LENGTH_MINUS_ONE BOARD_LENGTH - 1.0f
#define BOARD_LENGTH_MINUS_ONE_POINT_FIVE BOARD_LENGTH - 1.5f
#define WALL_EPS 1e-3f
#define BOARD_WIDTH_MINUS_WALL_EPS 9.999f
@@ -161,21 +163,46 @@ Pedestrian *init_pedestrians(int n_p) {
Board init_board_without_obstacles() { return Board(BOARD_LENGTH, BOARD_WIDTH, 0, nullptr); }
// # 256 MULTs, 224 ADDs
static inline __m512 _avx512_exp_ps(__m512 x) {
const __m512 log2e = _mm512_set1_ps(1.4426950408889634f);
// # 16 MULTs
__m512 nf = _mm512_roundscale_ps(_mm512_mul_ps(x, log2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
const __m512 ln2 = _mm512_set1_ps(0.6931471805599453f);
// # 16 MULTs, 16 ADDs
__m512 r = _mm512_fnmadd_ps(nf, ln2, x);
const __m512 c5 = _mm512_set1_ps(1.0f / 120.0f);
const __m512 c4 = _mm512_set1_ps(1.0f / 24.0f);
const __m512 c3 = _mm512_set1_ps(1.0f / 6.0f);
const __m512 c2 = _mm512_set1_ps(0.5f);
const __m512 c1 = _mm512_set1_ps(1.0f);
const __m512 c0 = _mm512_set1_ps(1.0f);
const __m512 c13 = _mm512_set1_ps(c13_scalar);
const __m512 c12 = _mm512_set1_ps(c12_scalar);
const __m512 c11 = _mm512_set1_ps(c11_scalar);
const __m512 c10 = _mm512_set1_ps(c10_scalar);
const __m512 c9 = _mm512_set1_ps(c9_scalar);
const __m512 c8 = _mm512_set1_ps(c8_scalar);
const __m512 c7 = _mm512_set1_ps(c7_scalar);
const __m512 c6 = _mm512_set1_ps(c6_scalar);
const __m512 c5 = _mm512_set1_ps(c5_scalar);
const __m512 c4 = _mm512_set1_ps(c4_scalar);
const __m512 c3 = _mm512_set1_ps(c3_scalar);
const __m512 c2 = _mm512_set1_ps(c2_scalar);
const __m512 c1 = _mm512_set1_ps(c1_scalar);
const __m512 c0 = _mm512_set1_ps(c0_scalar);
__m512 p = c5;
__m512 p = c13;
// # 13 * (16 MULTs, 16 ADDs)
// # = 208 MULTs, 208 ADDs
p = _mm512_fmadd_ps(p, r, c12);
p = _mm512_fmadd_ps(p, r, c11);
p = _mm512_fmadd_ps(p, r, c10);
p = _mm512_fmadd_ps(p, r, c9);
p = _mm512_fmadd_ps(p, r, c8);
p = _mm512_fmadd_ps(p, r, c7);
p = _mm512_fmadd_ps(p, r, c6);
p = _mm512_fmadd_ps(p, r, c5);
p = _mm512_fmadd_ps(p, r, c4);
p = _mm512_fmadd_ps(p, r, c3);
p = _mm512_fmadd_ps(p, r, c2);
@@ -187,6 +214,7 @@ static inline __m512 _avx512_exp_ps(__m512 x) {
__m512i scaled = _mm512_slli_epi32(biased, 23);
__m512 pow2n = _mm512_castsi512_ps(scaled);
// # 16 MULTs
return _mm512_mul_ps(p, pow2n);
}
@@ -225,11 +253,42 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
const __m512 v_np = _mm512_set1_ps((float)n_p);
const __m512 v_exp_min = _mm512_set1_ps(-87.0f);
// # ADDs: n_iter * (
// # n_p * 36 +
// # n_p * n_p_pad * 29 +
// # n_p * board.n_obstacles * 10 +
// # n_p_pad * 7
// # )
// # MULTs: n_iter * (
// # n_p * 4 +
// # n_p * n_p_pad * 40 +
// # n_p * board.n_obstacles * 13 +
// # n_p_pad * 15
// # )
// # DIVs: n_iter * (
// # n_p * n_p_pad +
// # n_p * board.n_obstacles +
// # n_p_pad * 2
// # )
// # SQRTs: n_iter * (
// # n_p * n_p_pad * 3 +
// # n_p * board.n_obstacles +
// # n_p_pad * 3
// # )
// # EXPs: n_iter * n_p * board.n_obstacles
for (int t = 0; t < n_iter; t++) {
for (int i = 0; i < n_p; i++) {
// # 2 MULTs
float desired_velocity_x = e_x[i] * desired_v0[i];
float desired_velocity_y = e_y[i] * desired_v0[i];
// # 2 ADDs, 2 MULTs
float f_alpha_dt_x = (desired_velocity_x - curr_vel_x[i]) * dt_times_tau_alpha_inv;
float f_alpha_dt_y = (desired_velocity_y - curr_vel_y[i]) * dt_times_tau_alpha_inv;
@@ -243,6 +302,8 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512 acc_y = _mm512_setzero_ps();
for (int j = 0; j < n_p_pad; j += 16) {
// # 16 ADDs
__m512 j_idx = _mm512_add_ps(_mm512_set1_ps((float)j), lane_off);
__mmask16 m_neq_i = _mm512_cmp_ps_mask(j_idx, v_i, _CMP_NEQ_OQ);
@@ -255,42 +316,75 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512 sbe_jy = _mm512_loadu_ps(&sb_e_y[j]);
__m512 sb_jsq = _mm512_loadu_ps(&sb_sq[j]);
// # 32 ADDs
__m512 r_ab_x = _mm512_sub_ps(pix, pjx);
__m512 r_ab_y = _mm512_sub_ps(piy, pjy);
// # 32 MULTs, 16 ADDs
__m512 r_ab_n2 = _mm512_fmadd_ps(r_ab_x, r_ab_x, _mm512_mul_ps(r_ab_y, r_ab_y));
// # 16 SQRTs
__m512 r_ab_n = _mm512_sqrt_ps(r_ab_n2);
r_ab_n = _mm512_max_ps(r_ab_n, v_eps);
// # 32 ADDs
__m512 r_ab2_x = _mm512_sub_ps(r_ab_x, sbe_jx);
__m512 r_ab2_y = _mm512_sub_ps(r_ab_y, sbe_jy);
// # 32 MULTs, 16 ADDs
__m512 r_ab2_n2 = _mm512_fmadd_ps(r_ab2_x, r_ab2_x, _mm512_mul_ps(r_ab2_y, r_ab2_y));
// # 16 SQRTs
__m512 r_ab2_n = _mm512_sqrt_ps(r_ab2_n2);
r_ab2_n = _mm512_max_ps(r_ab2_n, v_eps);
// # 16 ADDs
__m512 r_sum = _mm512_add_ps(r_ab_n, r_ab2_n);
// # 16 MULTs, 16 ADDs
__m512 two_b_sq = _mm512_fmsub_ps(r_sum, r_sum, sb_jsq);
two_b_sq = _mm512_max_ps(two_b_sq, v_eps_sq);
// # 16 SQRTs
__m512 two_b = _mm512_sqrt_ps(two_b_sq);
// # 16 MULTs
__m512 e_arg = _mm512_mul_ps(two_b, v_neg_2s_inv);
e_arg = _mm512_max_ps(e_arg, v_exp_min);
// # 256 MULTs, 224 ADDs
__m512 e_val = _avx512_exp_ps(e_arg);
// # 64 MULTs
__m512 numer = _mm512_mul_ps(_mm512_mul_ps(r_sum, v_neg_v0_div_2s), e_val);
__m512 denom = _mm512_mul_ps(_mm512_mul_ps(two_b, r_ab_n), r_ab2_n);
__m512 denom_safe = _mm512_mask_blend_ps(active, v_one, denom);
// # 16 DIVs
__m512 smc = _mm512_div_ps(numer, denom_safe);
// # 32 MULTs
__m512 sm_r_ab = _mm512_mul_ps(smc, r_ab2_n);
__m512 sm_r_ab2 = _mm512_mul_ps(smc, r_ab_n);
// # 64 MULTs, 32 ADDs
__m512 f_x = _mm512_fmadd_ps(r_ab2_x, sm_r_ab2, _mm512_mul_ps(r_ab_x, sm_r_ab));
__m512 f_y = _mm512_fmadd_ps(r_ab2_y, sm_r_ab2, _mm512_mul_ps(r_ab_y, sm_r_ab));
// # 32 MULTs, 16 ADDs
__m512 f_norm_sq = _mm512_fmadd_ps(f_x, f_x, _mm512_mul_ps(f_y, f_y));
// # 16 MULTs
__m512 rhs = _mm512_mul_ps(f_norm_sq, v_cos_phi_sq);
// # 32 MULTs, 16 ADDs
__m512 e_dot_f = _mm512_fmadd_ps(eix, f_x, _mm512_mul_ps(eiy, f_y));
// # 16 MULTs
__m512 e_dot_f_sq = _mm512_mul_ps(e_dot_f, e_dot_f);
__mmask16 c1 = _mm512_cmp_ps_mask(e_dot_f, v_zero, _CMP_LT_OQ);
@@ -302,19 +396,24 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512 f_x_safe = _mm512_maskz_mov_ps(active, f_x);
__m512 f_y_safe = _mm512_maskz_mov_ps(active, f_y);
// # 32 MULTs, 32 ADDs
acc_x = _mm512_fmadd_ps(f_x_safe, w_val, acc_x);
acc_y = _mm512_fmadd_ps(f_y_safe, w_val, acc_y);
}
// # 32 ADDs
f_alpha_dt_x += _mm512_reduce_add_ps(acc_x);
f_alpha_dt_y += _mm512_reduce_add_ps(acc_y);
for (int k = 0; k < board.n_obstacles; k++) {
// # 2 ADDs
float v1_x = pos_x[i] - board.obstacles[k].start_x;
float v1_y = pos_y[i] - board.obstacles[k].start_y;
float v2_x = board.obstacles[k].obstacle_v_x;
float v2_y = board.obstacles[k].obstacle_v_y;
// # 3 MULTs, 1 ADD
float projection = (v1_x * v2_x + v1_y * v2_y) * board.obstacles[k].length_sq_inv;
if (projection < 0.0f)
@@ -322,12 +421,15 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
else if (projection > 1.0f)
projection = 1.0f;
// # 4 ADDs, 2 MULTs
float r_aB_x = pos_x[i] - (board.obstacles[k].start_x + board.obstacles[k].obstacle_v_x * projection);
float r_aB_y = pos_y[i] - (board.obstacles[k].start_y + board.obstacles[k].obstacle_v_y * projection);
// # 1 SQRT, 2 MULTs, 1 ADD
float r_aB_norm = sqrt(r_aB_x * r_aB_x + r_aB_y * r_aB_y);
r_aB_norm = r_aB_norm > EPS ? r_aB_norm : EPS;
// # 4 MULTs, 1 EXP, 1 DIV
float scalar_mult = (NEG_U0_AB_div_R * exp(r_aB_norm * NEG_R_INV)) / r_aB_norm;
float f_alpha_B_neg_x = r_aB_x * scalar_mult;
float f_alpha_B_neg_y = r_aB_y * scalar_mult;
@@ -338,19 +440,22 @@ 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
float f_alpha_B_dt_x = f_alpha_B_neg_x * w_val;
float f_alpha_B_dt_y = f_alpha_B_neg_y * w_val;
// # 2 ADDs
f_alpha_dt_x += f_alpha_B_dt_x;
f_alpha_dt_y += f_alpha_B_dt_y;
}
// # 2 ADDs
w_x[i] += f_alpha_dt_x;
w_y[i] += f_alpha_dt_y;
}
for (int i = 0; i < n_p_pad; i += 16) {
__m512 w_x_i_vec = _mm512_loadu_ps(&w_x[i]);
__m512 w_y_i_vec = _mm512_loadu_ps(&w_y[i]);
__m512 v_max_i_vec = _mm512_loadu_ps(&v_max[i]);
@@ -359,23 +464,31 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__m512 destination_y_i_vec = _mm512_loadu_ps(&destination_y[i]);
__m512 destination_x_i_vec = _mm512_loadu_ps(&destination_x[i]);
__m512 desired_v0_i_vec = _mm512_loadu_ps(&desired_v0[i]);
__m512i color_i_vec = _mm512_loadu_si512((__m512i*)&color[i]);
__m512i color_i_vec = _mm512_loadu_si512((__m512i *)&color[i]);
// # 32 MULTs, 16 ADDs
__m512 w_norm_sq = _mm512_fmadd_ps(w_x_i_vec, w_x_i_vec, _mm512_mul_ps(w_y_i_vec, w_y_i_vec));
// # 16 SQRTs
__m512 w_norm = _mm512_sqrt_ps(w_norm_sq);
__m512 w_norm_safe = _mm512_max_ps(w_norm, _mm512_set1_ps(EPS));
// # 16 DIVs
__m512 correction_factor = _mm512_min_ps(_mm512_set1_ps(1.0f), _mm512_div_ps(v_max_i_vec, w_norm_safe));
// # 32 MULTs
__m512 curr_vel_x_i_vec = _mm512_mul_ps(w_x_i_vec, correction_factor);
__m512 curr_vel_y_i_vec = _mm512_mul_ps(w_y_i_vec, correction_factor);
__m512 dt_vec = _mm512_set1_ps(dt);
// # 32 MULTs, 32 ADDs
pos_x_i_vec = _mm512_fmadd_ps(curr_vel_x_i_vec, dt_vec, pos_x_i_vec);
pos_y_i_vec = _mm512_fmadd_ps(curr_vel_y_i_vec, dt_vec, pos_y_i_vec);
__mmask16 cond_1_left = _mm512_cmpeq_epi32_mask(color_i_vec, _mm512_setzero_si512());
__mmask16 cond_1_right = _mm512_cmp_ps_mask(pos_x_i_vec, _mm512_set1_ps(BOARD_LENGTH - 1.5f), _CMP_GT_OQ);
__mmask16 cond_1_right = _mm512_cmp_ps_mask(pos_x_i_vec, _mm512_set1_ps(BOARD_LENGTH_MINUS_ONE_POINT_FIVE), _CMP_GT_OQ);
__mmask16 cond_1 = cond_1_left & cond_1_right;
pos_x_i_vec = _mm512_mask_blend_ps(cond_1, pos_x_i_vec, _mm512_set1_ps(1.0f));
@@ -389,7 +502,7 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
__mmask16 cond_2_right = _mm512_cmp_ps_mask(pos_x_i_vec, _mm512_set1_ps(1.5f), _CMP_LT_OQ);
__mmask16 cond_2 = cond_2_left & cond_2_right;
pos_x_i_vec = _mm512_mask_blend_ps(cond_2, pos_x_i_vec, _mm512_set1_ps(BOARD_LENGTH - 1.0f));
pos_x_i_vec = _mm512_mask_blend_ps(cond_2, pos_x_i_vec, _mm512_set1_ps(BOARD_LENGTH_MINUS_ONE));
destination_y_i_vec = _mm512_mask_blend_ps(cond_2, destination_y_i_vec, pos_y_i_vec);
__m512 neg_desired_v0 = _mm512_xor_ps(desired_v0_i_vec, _mm512_set1_ps(-0.0f));
w_x_i_vec = _mm512_mask_blend_ps(cond_2, w_x_i_vec, neg_desired_v0);
@@ -400,16 +513,30 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
pos_y_i_vec = _mm512_max_ps(pos_y_i_vec, _mm512_set1_ps(WALL_EPS));
pos_y_i_vec = _mm512_min_ps(pos_y_i_vec, _mm512_set1_ps(BOARD_WIDTH_MINUS_WALL_EPS));
// # 32 ADDs
__m512 ex_vec = _mm512_sub_ps(destination_x_i_vec, pos_x_i_vec);
__m512 ey_vec = _mm512_sub_ps(destination_y_i_vec, pos_y_i_vec);
// # 32 MULTs, 16 ADDs
__m512 e_norm_sq = _mm512_fmadd_ps(ex_vec, ex_vec, _mm512_mul_ps(ey_vec, ey_vec));
// # 16 SQRTs
__m512 e_norm = _mm512_sqrt_ps(e_norm_sq);
// # 16 DIVs
__m512 e_norm_inv = _mm512_div_ps(_mm512_set1_ps(1.0f), e_norm);
// # 32 MULTs
__m512 e_x_i_vec = _mm512_mul_ps(ex_vec, e_norm_inv);
__m512 e_y_i_vec = _mm512_mul_ps(ey_vec, e_norm_inv);
// # 32 MULTs, 16 ADDs
__m512 vel_norm_sq = _mm512_fmadd_ps(curr_vel_x_i_vec, curr_vel_x_i_vec, _mm512_mul_ps(curr_vel_y_i_vec, curr_vel_y_i_vec));
// # 16 MULTs
__m512 sb_sq_i_vec = _mm512_mul_ps(vel_norm_sq, _mm512_set1_ps(delta_t_sq));
// # 16 SQRTs
__m512 vel_norm_dt = _mm512_sqrt_ps(sb_sq_i_vec);
_mm512_storeu_ps(&curr_vel_x[i], curr_vel_x_i_vec);
@@ -422,6 +549,8 @@ int run_simulation(Pedestrian *pedestrians, Board &board, int n_p, int n_iter, c
_mm512_storeu_ps(&e_y[i], e_y_i_vec);
_mm512_storeu_ps(&destination_y[i], destination_y_i_vec);
_mm512_storeu_ps(&sb_sq[i], sb_sq_i_vec);
// # 32 MULTs
_mm512_storeu_ps(&sb_e_x[i], _mm512_mul_ps(e_x_i_vec, vel_norm_dt));
_mm512_storeu_ps(&sb_e_y[i], _mm512_mul_ps(e_y_i_vec, vel_norm_dt));
}