perf: IBA::unsharp_mask() speed and memory optimization (#4513)

Replacing 3x IBA + Helper function that generate 4 fulls size image
buffers with single unsharp_mask_impl() that use parallel_image() to
compute unsharp:
src + contr * (((src - blur) < threshold) ? 0.0 : (src - blur))

Added two pass 1D convolution for a kernels higher than 3x3

## Tests

```
	ImageBuf sharped(input.spec());
	const int repeats = 50;

	std::cout << "Start sharpening\n";
	auto start = std::chrono::high_resolution_clock::now();

	for (int i = 0; i < repeats; i++)
	{
		//ok = ImageBufAlgo::unsharp_mask(sharped, input, "gaussian", 15.0f, 10.0f, 0.01f);
		ok = ImageBufAlgo::unsharp_mask(sharped, input, "gaussian", 5.0f, 2.0f, 0.05f);
		std::cout << ".";
	}

	std::cout << "\n";

	auto part1 = std::chrono::high_resolution_clock::now();
	std::chrono::duration<double> elapsed_part1 = part1 - start;
	std::cout << "Elapsed time: " << elapsed_part1.count() << " s\n";
```

both single threaded (one IB at time) and multithreaded (multiply IB at
time) show pretty good speedup:
~30-40% with less memory use.

for 5x5 gaussian kernels two pass mode should add at least 20% speedup.

(if someone can do independent benchmark, will be great. As soon as I
had a big differences on them depend on real or synthetic use)

---------

Signed-off-by: Vlad (Kuzmin) Erium <[email protected]>

Author: ssh4net

src/libOpenImageIO/imagebufalgo.cpp

@@ -947,17 +947,28 @@ ImageBufAlgo::make_kernel(string_view name, float width, float height,
 
 
 
-// Helper function for unsharp mask to perform the thresholding
+template<class Rtype>
 static bool
-threshold_to_zero(ImageBuf& dst, float threshold, ROI roi, int nthreads)
+unsharp_impl(ImageBuf& dst, const ImageBuf& blr, const ImageBuf& src,
+             const float contrast, const float threshold, ROI roi, int nthreads)
 {
-    OIIO_DASSERT(dst.spec().format.basetype == TypeDesc::FLOAT);
+    OIIO_DASSERT(dst.spec().nchannels == src.spec().nchannels
+                 && dst.spec().nchannels == blr.spec().nchannels);
 
     ImageBufAlgo::parallel_image(roi, nthreads, [&](ROI roi) {
-        for (ImageBuf::Iterator<float> p(dst, roi); !p.done(); ++p)
-            for (int c = roi.chbegin; c < roi.chend; ++c)
-                if (fabsf(p[c]) < threshold)
-                    p[c] = 0.0f;
+        ImageBuf::ConstIterator<Rtype> s(src, roi);
+        ImageBuf::ConstIterator<float> b(blr, roi);
+        for (ImageBuf::Iterator<Rtype> d(dst, roi); !d.done(); ++s, ++d, ++b) {
+            for (int c = roi.chbegin; c < roi.chend; ++c) {
+                const float diff     = s[c] - b[c];
+                const float abs_diff = fabsf(diff);
+                if (abs_diff > threshold) {
+                    d[c] = s[c] + contrast * diff;
+                } else {
+                    d[c] = s[c];
+                }
+            }
+        }
     });
     return true;
 }
@@ -977,10 +988,26 @@ ImageBufAlgo::unsharp_mask(ImageBuf& dst, const ImageBuf& src,
     // Blur the source image, store in Blurry
     ImageSpec BlurrySpec = src.spec();
     BlurrySpec.set_format(TypeDesc::FLOAT);  // force float
+    ImageBuf fst_pass(BlurrySpec);
     ImageBuf Blurry(BlurrySpec);
 
     if (kernel == "median") {
         median_filter(Blurry, src, ceilf(width), 0, roi, nthreads);
+    } else if (width > 3.0) {
+        ImageBuf K  = make_kernel(kernel, 1, width);
+        ImageBuf Kt = ImageBufAlgo::transpose(K);
+        if (K.has_error()) {
+            dst.errorfmt("{}", K.geterror());
+            return false;
+        }
+        if (!convolve(fst_pass, src, K, true, roi, nthreads)) {
+            dst.errorfmt("{}", fst_pass.geterror());
+            return false;
+        }
+        if (!convolve(Blurry, fst_pass, Kt, true, roi, nthreads)) {
+            dst.errorfmt("{}", Blurry.geterror());
+            return false;
+        }
     } else {
         ImageBuf K = make_kernel(kernel, width, width);
         if (K.has_error()) {
@@ -993,25 +1020,10 @@ ImageBufAlgo::unsharp_mask(ImageBuf& dst, const ImageBuf& src,
         }
     }
 
-    // Compute the difference between the source image and the blurry
-    // version.  (We store it in the same buffer we used for the difference
-    // image.)
-    ImageBuf& Diff(Blurry);
-    bool ok = sub(Diff, src, Blurry, roi, nthreads);
-
-    if (ok && threshold > 0.0f)
-        ok = threshold_to_zero(Diff, threshold, roi, nthreads);
-
-    // Scale the difference image by the contrast
-    if (ok)
-        ok = mul(Diff, Diff, contrast, roi, nthreads);
-    if (!ok) {
-        dst.errorfmt("{}", Diff.geterror());
-        return false;
-    }
-
-    // Add the scaled difference to the original, to get the final answer
-    ok = add(dst, src, Diff, roi, nthreads);
+    bool ok;
+    OIIO_DISPATCH_COMMON_TYPES(ok, "unsharp_mask", unsharp_impl,
+                               dst.spec().format, dst, Blurry, src, contrast,
+                               threshold, roi, nthreads);
 
     return ok;
 }

testsuite/docs-examples-cpp/ref/out-arm.txt

@@ -137,7 +137,7 @@ checker_with_alpha_filled.exr :  256 x  256, 4 channel, half openexr
 tahoe_median_filter.tif :  512 x  384, 3 channel, uint8 tiff
     SHA-1: A0B2E3A10A16EA8CC905F144C5F91B6A0964A177
 tahoe_unsharp_mask.tif :  512 x  384, 3 channel, uint8 tiff
-    SHA-1: CDE3FAC8053381C59B7BEB3B47991F357E14D9D2
+    SHA-1: 5842D16483BC74700DE9FD27967B2FFBD54DFCD2
 Comparing "simple.tif" and "ref/simple.tif"
 PASS
 Comparing "scanlines.tif" and "ref/scanlines.tif"

testsuite/docs-examples-cpp/ref/out.txt

@@ -137,7 +137,7 @@ checker_with_alpha_filled.exr :  256 x  256, 4 channel, half openexr
 tahoe_median_filter.tif :  512 x  384, 3 channel, uint8 tiff
     SHA-1: A0B2E3A10A16EA8CC905F144C5F91B6A0964A177
 tahoe_unsharp_mask.tif :  512 x  384, 3 channel, uint8 tiff
-    SHA-1: D3B56074F48EC5D3ADDA4BDE1F487192ABE9BA76
+    SHA-1: C1C9C843D45D90B7C0BBD7BCDB7A11814668FC6D
 Comparing "simple.tif" and "ref/simple.tif"
 PASS
 Comparing "scanlines.tif" and "ref/scanlines.tif"

testsuite/docs-examples-python/ref/out-arm.txt

@@ -137,7 +137,7 @@ checker_with_alpha_filled.exr :  256 x  256, 4 channel, half openexr
 tahoe_median_filter.tif :  512 x  384, 3 channel, uint8 tiff
     SHA-1: A0B2E3A10A16EA8CC905F144C5F91B6A0964A177
 tahoe_unsharp_mask.tif :  512 x  384, 3 channel, uint8 tiff
-    SHA-1: CDE3FAC8053381C59B7BEB3B47991F357E14D9D2
+    SHA-1: 5842D16483BC74700DE9FD27967B2FFBD54DFCD2
 Comparing "simple.tif" and "../docs-examples-cpp/ref/simple.tif"
 PASS
 Comparing "scanlines.tif" and "../docs-examples-cpp/ref/scanlines.tif"

testsuite/docs-examples-python/ref/out.txt

@@ -137,7 +137,7 @@ checker_with_alpha_filled.exr :  256 x  256, 4 channel, half openexr
 tahoe_median_filter.tif :  512 x  384, 3 channel, uint8 tiff
     SHA-1: A0B2E3A10A16EA8CC905F144C5F91B6A0964A177
 tahoe_unsharp_mask.tif :  512 x  384, 3 channel, uint8 tiff
-    SHA-1: D3B56074F48EC5D3ADDA4BDE1F487192ABE9BA76
+    SHA-1: C1C9C843D45D90B7C0BBD7BCDB7A11814668FC6D
 Comparing "simple.tif" and "../docs-examples-cpp/ref/simple.tif"
 PASS
 Comparing "scanlines.tif" and "../docs-examples-cpp/ref/scanlines.tif"