Merge pull request #3 from neurolabusc/main

Lookup table to improve float16 performance

README.md

@@ -23,7 +23,7 @@ This code should be supported by all modern web browsers. If you want to use thi
 
 The module `streamlineIO.mjs` provides JavaScript reading for streamlines in the TCX, TRK, TRX, and VTK formats. The included file `bench.mjs` demonstrates these capabilities by loading a streamline 11 times and reporting details regarding file size, loading time, and streamline properties.
 
-Assuming you have [node 14.11.2 or later]([web site](https://nodejs.org/en/download/) you can run the demo:
+Assuming you have [node 14.11.2 or later](https://nodejs.org/en/download/) you can run the demo:
 
 ```
 $ git clone https://github.com/tee-ar-ex/trx-javascript
@@ -68,7 +68,7 @@ There are several important considerations regarding supporting the TRX format w
 
 The included JavaScript `bench` provides a method to evaluate performance. This benchmark is likely specific to JavaScript and so caution should be excercised in evaluating relative performance. The script will report the time to load a TRK, TCK, VTK or TRX file 10 times (it loads the tracts 11 times, and ignores the first run).
 
-The graph below shows the time load the [left inferior fronto-occipital fasciculus (IFOF) ](https://brain.labsolver.org/hcp_trk_atlas.html) with 30856 streamlines and 11437105 vertices. The benchmark was run on a passively-cooled 15w M1-based MacBook Air. The ideal format would be both fast to load (to the left on the horizontal axis) and have a small file size (toward the bottom in the right axis). However, compression typically trades load time for file size. Here all data is loaded from a local solid state drive, whereas smaller files would benefit if data was loaded using a slow internet connection. The following file formats are illustrated (except where noted, both positions and indices are stored with 32-bit precision):
+The graph below shows the time load the [left inferior fronto-occipital fasciculus (IFOF) ](https://brain.labsolver.org/hcp_trk_atlas.html) from the HCP1065 Population-Averaged Tractography Atlas (Yeh, 2022). This has with 21209 streamlines and 4915166 vertices. The different formats are generated with the [tff_convert_tractogram.py](https://github.com/tee-ar-ex/trx-python) script.The benchmark was run on a MacBook laptop with a M2 Pro CPU. The ideal format would be both fast to load (to the left on the horizontal axis) and have a small file size (toward the bottom in the right axis). However, compression typically trades load time for file size. Here all data is loaded from a local solid state drive, whereas smaller files would benefit if data was loaded using a slow internet connection. The following file formats are illustrated (except where noted, both positions and indices are stored with 32-bit precision):
 
  - vtk: streamlines saved by an [undocumented](https://discourse.vtk.org/t/upcoming-changes-to-vtkcellarray/2066) extension to the [VTK legacy file format](https://vtk.org/wp-content/uploads/2015/04/file-formats.pdf).
  - tck: [MRtrix format](https://mrtrix.readthedocs.io/en/latest/getting_started/image_data.html#tracks-file-format-tck).
@@ -80,4 +80,4 @@ The graph below shows the time load the [left inferior fronto-occipital fascicul
  - 16.trx: uncompressed TRX format using 16-bit floats for positions (which are not native to JavaScript).
  - 16z.trx: zip-compressed TRX format using 16-bit floats for positions (which are not native to JavaScript).
 
-![M1 Performance](M1.png)
+![M2 Performance](M2.png)

graphM2.py

@@ -0,0 +1,32 @@
+import pandas as pd
+import matplotlib.pyplot as plt
+import seaborn as sns
+
+# Create a dataframe
+data = {
+    'format': ['vtk', 'tck', 'trk', 'trk.gz', 'trk.zst', 'trx', 'z.trx', '16.trx', '16z.trx'],
+    'time': [459, 866, 924, 3881, 4012, 159, 2763, 373, 2708],
+    'bytes': [113218694, 59236654, 59067828, 28746432, 28291562, 59152204, 28502799, 29661208, 24245538]
+}
+
+df = pd.DataFrame(data)
+
+# Define a color palette with as many colors as there are formats
+colors = sns.color_palette('husl', n_colors=len(df['format']))
+
+# Set a larger figure size
+plt.figure(figsize=(10, 5))
+
+# Create a scatter plot with different colors for each format and double the diameter of each point
+ax = sns.scatterplot(x='time', y='bytes', hue='format', palette=colors, s=200, data=df)
+
+# Set the legend outside the plot for better visibility
+ax.legend(title='Format', bbox_to_anchor=(1.05, 1), loc='upper left')
+
+plt.title('Relative Time vs Size for Each Format')
+plt.xlabel('Relative Time')
+plt.ylabel('Size (Bytes)')
+
+plt.tight_layout()  # Adjust layout to prevent clipping
+#plt.show()
+plt.savefig('M2.png', dpi=300)

streamlineIO.mjs

@@ -591,7 +591,9 @@ async function readTRX(url, urlIsLocalFile = false) {
       nval = data.length / 2; //2 bytes per 16bit input
       vals = new Float32Array(nval);
       var u16 = new Uint16Array(data.buffer);
-      for (let i = 0; i < nval; i++) vals[i] = decodeFloat16(u16[i]);
+      const lut = new Float32Array(65536)
+      for (let i = 0; i < 65536; i++) lut[i] = decodeFloat16(i)
+      for (let i = 0; i < nval; i++) vals[i] = lut[u16[i]]
     } else continue; //not a data array
     nval = vals.length;
     //next: read data_per_group