Add perftests/gtiff_multi_ds_parallel_write.py to measure TIFF write …

…performance

perftests/gtiff_multi_ds_parallel_write.py

@@ -0,0 +1,92 @@
+# SPDX-License-Identifier: MIT
+# Copyright 2023 Even Rouault
+
+# This benchmark utility helps assessing the efficiency of the bypass
+# of the block cache when writing GeoTIFF datasets.
+
+import array
+import sys
+from threading import Thread
+
+from osgeo import gdal
+
+gdal.UseExceptions()
+
+
+def Usage():
+    print("gtiff_multi_ds_parallel_write.py [--nbands VAL] [--without-optim]")
+    print("                                 [--compress NONE/ZSTD/...]")
+    print("                                 [--buffer-interleave PIXEL/BAND]")
+    sys.exit(1)
+
+
+num_threads = gdal.GetNumCPUs()
+nbands = 1
+with_optim = True
+compression = "ZSTD"
+buffer_pixel_interleaved = True
+width = 2048
+height = 2048
+
+# Parse arguments
+i = 1
+while i < len(sys.argv):
+    if sys.argv[i] == "--nbands":
+        i += 1
+        nbands = int(sys.argv[i])
+    elif sys.argv[i] == "--compress":
+        i += 1
+        compression = sys.argv[i]
+    elif sys.argv[i] == "--buffer-interleave":
+        i += 1
+        buffer_pixel_interleaved = sys.argv[i] == "PIXEL"
+    elif sys.argv[i] == "--without-optim":
+        with_optim = False
+    else:
+        Usage()
+    i += 1
+
+nloops = 1000 // nbands
+data = array.array("B", [i % 255 for i in range(nbands * width * height)])
+
+gdal.SetCacheMax(width * height * nbands * num_threads)
+
+
+def thread_function(num):
+    filename = "/vsimem/tmp%d.tif" % num
+    drv = gdal.GetDriverByName("GTiff")
+    options = ["TILED=YES", "COMPRESS=" + compression]
+    for i in range(nloops):
+        ds = drv.Create(filename, width, height, nbands, options=options)
+        if not with_optim:
+            # Calling ReadRaster() disables the cache bypass write optimization
+            ds.GetRasterBand(1).ReadRaster(0, 0, 1, 1)
+        if nbands > 1:
+            if buffer_pixel_interleaved:
+                # Write pixel-interleaved buffer for maximum efficiency
+                ds.WriteRaster(
+                    0,
+                    0,
+                    width,
+                    height,
+                    data,
+                    buf_pixel_space=nbands,
+                    buf_line_space=width * nbands,
+                    buf_band_space=1,
+                )
+            else:
+                ds.WriteRaster(0, 0, width, height, data)
+        else:
+            ds.GetRasterBand(1).WriteRaster(0, 0, width, height, data)
+
+
+# Spawn num_threads running thread_function
+threads_array = []
+
+for i in range(num_threads):
+    t = Thread(target=thread_function, args=[i])
+    t.start()
+    threads_array.append(t)
+
+for t in threads_array:
+    t.join()