Improve performance of color distance calculations by kernel fusion (#…

…809)

This MR is purely to fuse many separate kernels into a single elementwise kernel for each of the "color distance" functions. This is expected to result in substantial performance improvement by reduced kernel launch overhead and because a single pass through the image is much more memory efficient than many separate kernel calls.

There is not expected to be any change in behavior (existing tests must continue to pass)

## Benchmark Results

I added benchmarks for these functions and compared the results before and after this change.

The acceleration values are the relative speedup as compared to the scikit-image implementation

### For a pair of 32-bit LAB images of shape: (512, 512, 3)

function         | acceleration (old) | acceleration (new)
-----------------|--------------------|-------------------
deltaE_cie76     |   5.49             |   21.85
deltaE_ciede94   |   37.60            |   65.09
deltaE_ciede2000 |   12.54            |  109.71
deltaE_cmc       |   27.18            |   89.23

### For a pair of 32-bit LAB images of shape: (3840, 2160, 3)

function         | acceleration (old) | acceleration (new)
-----------------|--------------------|-------------------
deltaE_cie76     |   31.12            |  250.74
deltaE_ciede94   |   79.15            |  117.50
deltaE_ciede2000 |   71.36            |  132.73
deltaE_cmc       |   70.65            |  154.83

Authors:
  - Gregory Lee (https://github.com/grlee77)

Approvers:
  - Gigon Bae (https://github.com/gigony)

URL: #809

benchmarks/skimage/cucim_color_bench.py

@@ -1,4 +1,5 @@
 import argparse
+import math
 import os
 import pickle
 
@@ -25,6 +26,10 @@
     "lab2xyz",
     "rgba2rgb",
     "label2rgb",
+    "deltaE_cie76",
+    "deltaE_ciede94",
+    "deltaE_ciede2000",
+    "deltaE_cmc",
 ]
 
 
@@ -40,6 +45,32 @@ def set_args(self, dtype):
         self.args_gpu = (imaged,)
 
 
+class DeltaEBench(ImageBench):
+    def set_args(self, dtype):
+        from skimage import color, data
+
+        # create synthetic lab image pair
+        rgb1 = data.astronaut()
+        lab1 = color.rgb2lab(rgb1)
+        lab2 = color.rgb2lab(np.roll(rgb1, (1, 1), axis=(0, 1)))
+
+        # change to desired dtype
+        lab1 = lab1.astype(dtype, copy=False)
+        lab2 = lab2.astype(dtype, copy=False)
+
+        # tile then crop as needed to get the expected size
+        n_tile0 = math.ceil(self.shape[0] / lab1.shape[0])
+        n_tile1 = math.ceil(self.shape[1] / lab1.shape[1])
+        lab1 = np.tile(lab1, (n_tile0, n_tile1, 1))
+        lab1 = lab1[: self.shape[0], : self.shape[1], :]
+        lab2 = np.tile(lab2, (n_tile0, n_tile1, 1))
+        lab2 = lab2[: self.shape[0], : self.shape[1], :]
+
+        print(f"{lab1.shape=}")
+        self.args_cpu = (lab1, lab2)
+        self.args_gpu = (cp.asarray(lab1), cp.asarray(lab2))
+
+
 class RGBABench(ImageBench):
     def set_args(self, dtype):
         if self.shape[-1] != 4:
@@ -162,6 +193,23 @@ def main(args):
                     results = B.run_benchmark(duration=args.duration)
                     all_results = pd.concat([all_results, results["full"]])
 
+        elif function_name.startswith("deltaE"):
+            # only run these functions for floating point data types
+            float_dtypes = [t for t in dtypes if np.dtype(t).kind == "f"]
+
+            B = DeltaEBench(
+                function_name=function_name,
+                shape=shape + (3,),
+                dtypes=float_dtypes,
+                fixed_kwargs={},
+                var_kwargs={},
+                # index_str=f"{fromspace.lower()}2{tospace.lower()}",
+                module_cpu=skimage.color,
+                module_gpu=cucim.skimage.color,
+                run_cpu=run_cpu,
+            )
+            results = B.run_benchmark(duration=args.duration)
+            all_results = pd.concat([all_results, results["full"]])
         elif function_name == "rgba2rgb":
             B = RGBABench(
                 function_name="rgba2rgb",

benchmarks/skimage/run-nv-bench-color.sh

@@ -9,3 +9,14 @@ for shape in "${param_shape[@]}"; do
         done
     done
 done
+
+param_shape=(512,512 3840,2160)
+param_filt=(deltaE_cie76 deltaE_ciede94 deltaE_ciede2000 deltaE_cmc)
+param_dt=(float32, float64)
+for shape in "${param_shape[@]}"; do
+    for filt in "${param_filt[@]}"; do
+        for dt in "${param_dt[@]}"; do
+            python cucim_color_bench.py -f $filt -i $shape -d $dt -t 10
+        done
+    done
+done