Inlining + Reorganization + Roofline

2026-05-19 23:17:44 +02:00
parent 43f11a661b
commit b86c856e08
34 changed files with 2303 additions and 221 deletions
@@ -0,0 +1,159 @@
+import matplotlib.pyplot as plt
+from matplotlib.ticker import FuncFormatter
+import numpy as np
+
+ram_beta = 1.91
+l1d_l2_beta = 64
+l3_beta = 32
+
+pi_scalar = 6.0
+pi_vector = 48.0
+
+I_vals = np.logspace(-50, 8, 500, base=2)
+
+plt.figure(figsize=(10, 7))
+
+plt.gca().set_facecolor('lightgrey')
+
+plt.plot(I_vals, ram_beta * I_vals, 'k-', label='_nolegend_')
+plt.plot(I_vals, l1d_l2_beta * I_vals, 'k-', label='_nolegend_')
+plt.plot(I_vals, l3_beta * I_vals, 'k-', label='_nolegend_')
+
+ram_bound_x_pos = 0.25
+ram_bound_y_pos = ram_beta * ram_bound_x_pos + 0.175
+
+plt.text(
+  ram_bound_x_pos,
+  ram_bound_y_pos,
+  f'Read β mem. ({ram_beta} bytes/cycle)',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+  rotation=20.5,
+)
+
+l1d_l2_bound_x_pos = 0.25
+l1d_l2_bound_y_pos = l1d_l2_beta * l1d_l2_bound_x_pos + 3.5
+
+plt.text(
+  l1d_l2_bound_x_pos,
+  l1d_l2_bound_y_pos,
+  f'Read β L1D/L2 ({l1d_l2_beta} bytes/cycle)',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+  rotation=20.5,
+)
+
+l3_bound_x_pos = 0.375
+l3_bound_y_pos = l3_beta * l3_bound_x_pos + 3.5
+
+plt.text(
+  l3_bound_x_pos,
+  l3_bound_y_pos,
+  f'Read β L3 ({l3_beta} bytes/cycle)',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+  rotation=20.5,
+)
+
+plt.plot(I_vals, [pi_scalar] * len(I_vals), 'k-', label='_nolegend_')
+plt.text(
+  85,
+  pi_scalar + 1.5,
+  f'Peak π seq. ({pi_scalar} flops/cycle)',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+)
+
+plt.plot(I_vals, [pi_vector] * len(I_vals), 'k-', label='_nolegend_')
+plt.text(
+  85,
+  pi_vector + 12.5,
+  f'Peak π par. ({pi_vector} flops/cycle)',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+)
+
+
+def fraction_formatter(x, _pos):
+  if x == 0:
+    return '0'
+  log2_x = np.log2(x)
+  if abs(log2_x - round(log2_x)) < 1e-6:
+    exponent = int(round(log2_x))
+    if exponent < 0:
+      denominator = 2**(-exponent)
+      return f'{1/denominator:.2f}'
+    elif exponent == 0:
+      return '1'
+    else:
+      return str(int(x))
+  return f'{x:.2f}'
+
+
+formatter = FuncFormatter(fraction_formatter)
+
+plt.xscale('log', base=2)
+plt.yscale('log', base=2)
+plt.gca().xaxis.set_major_formatter(formatter)
+plt.gca().yaxis.set_major_formatter(formatter)
+
+plt.xlabel('Operational Intensity [flops/byte]', fontsize=10)
+plt.title("""Roofline Plot for the Social Force Model
+CPU: AMD Ryzen AI 7 350 w/ Radeon 860M @ 2 GHz
+Compiler: g++ (Debian 15.2.0-16) 15.2.0
+Command: ./main 100 10000
+
+Performance [flops/cycle]""",
+          loc='left',
+          size=10)
+plt.scatter([53], [0.0936126],
+            c='black',
+            marker='o',
+            edgecolors='black',
+            label='baseline (-O0)')
+plt.text(
+  53,
+  0.0936126 - 0.03,
+  f'0.09 flops/cycle',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+)
+plt.scatter([53], [0.627801],
+            c='black',
+            marker='s',
+            edgecolors='black',
+            label='baseline (-O3 -march=native)')
+plt.text(
+  53,
+  0.627801 - 0.2,
+  f'0.63 flops/cycle',
+  color='black',
+  va='center',
+  ha='center',
+  fontsize=8,
+)
+
+plt.legend(loc='lower right',
+           fontsize=11,
+           frameon=True,
+           facecolor='white',
+           edgecolor='black')
+plt.grid(True, which="both", ls="-", color='white', alpha=0.8, linewidth=1)
+
+plt.xlim(0.05, 200)
+plt.ylim(0.0005, 100)
+
+plt.tight_layout()
+plt.show()