.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "_as_gen/benchmarks/run_sph_homogeneous_results.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note :ref:`Go to the end ` to download the full example code. .. rst-class:: sphx-glr-example-title .. _sphx_glr__as_gen_benchmarks_run_sph_homogeneous_results.py: Sph homogeneous benchmarks results ================================== Show the results on various devices .. GENERATED FROM PYTHON SOURCE LINES 7-177 .. image-sg:: /_as_gen/benchmarks/images/sphx_glr_run_sph_homogeneous_results_001.png :alt: SPH homogeneous - rate by device :srcset: /_as_gen/benchmarks/images/sphx_glr_run_sph_homogeneous_results_001.png :class: sphx-glr-single-img .. rst-class:: sphx-glr-script-out .. code-block:: none results from /work/examples/benchmarks/sph_homogeneous_bench_result.json Apple M4 Max: - 1 ranks, 851313.6209866619 rate, 8484840 cnt, 9.966761709 step time NVIDIA GeForce RTX 3070: - 1 ranks, 1732146.8265657455 rate, 4254912 cnt, 2.4564384120000002 step time Intel(R) Core(TM) Ultra 9 285K: - 1 ranks, 1612764.908074928 rate, 8484840 cnt, 5.261051972000001 step time NVIDIA H100: - 1 ranks, 25502055.876729604 rate, 33848064 cnt, 1.3272680510000001 step time AMD EPYC 9654 96-Core Processor : - 1 ranks, 2353244.022271519 rate, 8464638 cnt, 3.597008181 step time Intel(R) Data Center GPU Max 1550: - 1 ranks, 14612536.848131763 rate, 33848064 cnt, 2.316371507 step time 12 x Intel(R) Data Center GPU Max 1550: - 12 ranks, 124464069.34448674 rate, 404289600 cnt, 3.2482434660000004 step time | .. code-block:: Python :lineno-start: 9 import json import os import textwrap import matplotlib.pyplot as plt import numpy as np from matplotlib.ticker import MaxNLocator try: base_path = os.path.dirname(os.path.abspath(__file__)) except NameError: base_path = os.getcwd() json_file = os.path.join(base_path, "sph_homogeneous_bench_result.json") results = json.load(open(json_file)) json.dump(results, open(json_file, "w"), indent=4) print(f"results from {json_file}") results_per_model = {} def key_name(name, world_size): if world_size == 1: return name else: return f"{world_size} x {name}" for result in results: name = key_name(result["device_properties"]["name"], result["world_size"]) if name not in results_per_model: results_per_model[name] = result else: if result["rate"] > results_per_model[name]["rate"]: results_per_model[name] = result for name, result in results_per_model.items(): print(f"{name}:") print( f" - {result['world_size']} ranks, {result['rate']} rate, {result['cnt']} cnt, {result['step_time']} step time" ) def _rate_bar_color(device_name: str) -> str: """Color for the rate bar from device name (case-insensitive).""" lower = device_name.lower() if "nvidia" in lower: return "#2ca02c" # green if "amd" in lower or "radeon" in lower: return "#d62728" # red if "intel" in lower: return "#1f77b4" # blue if "apple" in lower: return "#7f7f7f" # grey return "steelblue" def _micro_bw_and_fma(result): """saxpy f64 -> GB/s; fma_chains f32/f64 -> Gflops (MicroBenchmark raw flop/s, /1e9).""" m = result.get("microbench_results") or {} bw_bs = m.get("saxpy_f64") f64 = m.get("fma_chains_f64") f32 = m.get("fma_chains_f32") bw_gbps = (bw_bs / 1e9) if bw_bs is not None else float("nan") flops_f64 = (f64) if f64 is not None else float("nan") flops_f32 = (f32) if f32 is not None else float("nan") return bw_gbps, flops_f64, flops_f32 # Stable sort by rate descending for a readable chart items = sorted(results_per_model.items(), key=lambda kv: kv[1]["rate"], reverse=True) names = [kv[0] for kv in items] rates = [kv[1]["rate"] for kv in items] bw_gbps = [] flops_f64 = [] flops_f32 = [] for _, r in items: bw, f64, f32 = _micro_bw_and_fma(r) bw_gbps.append(bw) flops_f64.append(f64) flops_f32.append(f32) h_in = max(3.0, 0.45 * len(names) + 5) y = np.arange(len(names)) fig, (ax_rate, ax_micro) = plt.subplots( 1, 2, sharey=True, figsize=(15, h_in), gridspec_kw={"width_ratios": [75, 25], "wspace": 0.025}, ) # Wrap long device names so they stay inside the figure margin _name_labels = ["\n".join(textwrap.wrap(n, 34)) for n in names] _rate_colors = [_rate_bar_color(n) for n in names] bars = ax_rate.barh(y, rates, color=_rate_colors, edgecolor="white", linewidth=0.5) ax_rate.set_yticks(y) ax_rate.set_yticklabels(_name_labels) ax_rate.set_xlabel("rate (solver objects / s)") ax_rate.set_xscale("log") ax_rate.set_title("SPH homogeneous - rate by device") ax_rate.bar_label(bars, fmt="%.3g", padding=3) ax_rate.grid(axis="x", linestyle=":", alpha=0.6) ax_rate.invert_yaxis() # Extra room for bar-end labels; drop rightmost x tick (avoids clash with right panel) _xmin, _xmax = ax_rate.get_xlim() ax_rate.set_xlim(_xmin, _xmax + 0.5 * (_xmax - _xmin)) # ax_rate.xaxis.set_major_locator(MaxNLocator(prune="upper")) # Three equal-height rows per device, evenly spaced around the tick (name at y) _bar_h = 0.22 _spacing = 0.26 # distance between bar centers; middle bar (f32) on the tick _y_saxpy = y - _spacing _y_f32 = y _y_f64 = y + _spacing ax_micro.barh( _y_saxpy, bw_gbps, height=_bar_h, color="coral", label="saxpy f64 (GB/s)", edgecolor="white", linewidth=0.5, ) ax_micro.set_xlabel("Memory bandwidth saxpy f64 (GB/s)") ax_micro.grid(axis="x", linestyle=":", alpha=0.6) ax_micro.tick_params(axis="y", labelleft=False) # f32 / f64 FMA can differ a lot in scale -> log-scaled Gflops axis (same y layout as saxpy) ax_micro_top = ax_micro.twiny() ax_micro_top.barh( _y_f32, flops_f32, height=_bar_h, color="mediumpurple", label="fma_chains f32 (flops)", edgecolor="white", linewidth=0.5, ) ax_micro_top.barh( _y_f64, flops_f64, height=_bar_h, color="seagreen", label="fma_chains f64 (flops)", edgecolor="white", linewidth=0.5, ) ax_micro_top.set_xlabel("Peak FMA f32 / f64 (flops, log scale)") ax_micro_top.set_xscale("log") ax_micro.set_xscale("log") h0, l0 = ax_micro.get_legend_handles_labels() h1, l1 = ax_micro_top.get_legend_handles_labels() ax_micro.legend(h0 + h1, l0 + l1, loc="lower right", fontsize=8) # Flush panels: constrained_layout always leaves a gap; manual wspace=0 truly abuts axes ax_rate.spines["right"].set_visible(True) ax_micro.spines["left"].set_visible(False) fig.subplots_adjust(left=0.22, right=0.99, top=0.90, bottom=0.12, wspace=0) plt.show() .. rst-class:: sphx-glr-timing **Total running time of the script:** (0 minutes 0.511 seconds) **Estimated memory usage:** 157 MB .. _sphx_glr_download__as_gen_benchmarks_run_sph_homogeneous_results.py: .. only:: html .. container:: sphx-glr-footer sphx-glr-footer-example .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: run_sph_homogeneous_results.ipynb ` .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: run_sph_homogeneous_results.py ` .. container:: sphx-glr-download sphx-glr-download-zip :download:`Download zipped: run_sph_homogeneous_results.zip ` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_