perf: Speed up maketx --envlatl when multithreaded by over 10x. (#4825)

[As reported on Slack](https://academysoftwarefdn.slack.com/archives/C05782U3806/p1751642730171029),
between OIIO 2.5 and 3.0, we had a big slowdown when doing a
multithreaded `maketx --envlatl`, which was traced almost entirely to
the new mutex lock in ImageBuf::IteratorBase::init_ib(), which were all
happening inside the call stack of `resize_block_<float>()`.

The reason was that it was calling ImageBuf::interppixel_NDC for every
pixel, which in turn works by creating (internally) an
ImageBuf::ConstIterator to sample the 4 nearby pixels. But every one of
those ConstIterator constructors called init_ib, which briefly grabbed
the mutex for the IB, which in addition to being wasteful on its own,
caused the threads to block on each other.

The solution is straightforward: there is no need to construct a new
Iterator for every pixel. We can create the iterator once (a single call
to the init_ib for each thread region of the image) and then for each
pixel, just call its `rerange()` method to reset the set of pixels to
loop over for the samples needed for that pixel.

I also opportunistically eliminated a few redundant spec() calls in
various routines in imagebuf.cpp.

On my Mac laptop, when doing a maketx --latlong on a 16k image with 16
threads, it was previously taking 170.5s (including 117.8s for the
initial resize and 42.8s for the MIP computations). With this change,
the same operation takes 12.4s (including 3.7s for the initial resize
and 1.6s for he MIP computations). That's almost a 14x speedup. YMMV,
depending on platform, compiler, image size, and number of threads.

Signed-off-by: Larry Gritz <[email protected]>

Author: lgritz

src/libOpenImageIO/imagebuf.cpp

@@ -2961,7 +2961,8 @@ ImageBuf::xbegin() const
 int
 ImageBuf::xend() const
 {
-    return spec().x + spec().width;
+    const ImageSpec& spec(m_impl->spec());
+    return spec.x + spec.width;
 }
 
 
@@ -2977,7 +2978,8 @@ ImageBuf::ybegin() const
 int
 ImageBuf::yend() const
 {
-    return spec().y + spec().height;
+    const ImageSpec& spec(m_impl->spec());
+    return spec.y + spec.height;
 }
 
 
@@ -2993,7 +2995,8 @@ ImageBuf::zbegin() const
 int
 ImageBuf::zend() const
 {
-    return spec().z + std::max(spec().depth, 1);
+    const ImageSpec& spec(m_impl->spec());
+    return spec.z + std::max(spec.depth, 1);
 }
 
 
@@ -3009,7 +3012,8 @@ ImageBuf::xmin() const
 int
 ImageBuf::xmax() const
 {
-    return spec().x + spec().width - 1;
+    const ImageSpec& spec(m_impl->spec());
+    return spec.x + spec.width - 1;
 }
 
 
@@ -3025,7 +3029,8 @@ ImageBuf::ymin() const
 int
 ImageBuf::ymax() const
 {
-    return spec().y + spec().height - 1;
+    const ImageSpec& spec(m_impl->spec());
+    return spec.y + spec.height - 1;
 }
 
 
@@ -3041,7 +3046,8 @@ ImageBuf::zmin() const
 int
 ImageBuf::zmax() const
 {
-    return spec().z + std::max(spec().depth, 1) - 1;
+    const ImageSpec& spec(m_impl->spec());
+    return spec.z + std::max(spec.depth, 1) - 1;
 }
 
 

src/libOpenImageIO/maketexture.cpp

@@ -138,17 +138,19 @@ datestring(time_t t)
 
 template<class SRCTYPE>
 static void
-interppixel_NDC_clamped(const ImageBuf& buf, float x, float y,
-                        span<float> pixel, bool envlatlmode)
+interppixel_NDC(const ImageBuf& buf, float x, float y, span<float> pixel,
+                bool envlatlmode, ImageBuf::ConstIterator<SRCTYPE>& it,
+                ImageBuf::WrapMode wrapmode)
 {
-    int fx = buf.spec().full_x;
-    int fy = buf.spec().full_y;
-    int fw = buf.spec().full_width;
-    int fh = buf.spec().full_height;
+    const ImageSpec& spec(buf.spec());
+    int fx = spec.full_x;
+    int fy = spec.full_y;
+    int fw = spec.full_width;
+    int fh = spec.full_height;
     x      = static_cast<float>(fx) + x * static_cast<float>(fw);
     y      = static_cast<float>(fy) + y * static_cast<float>(fh);
 
-    int n     = buf.spec().nchannels;
+    int n     = spec.nchannels;
     float *p0 = OIIO_ALLOCA(float, 4 * n), *p1 = p0 + n, *p2 = p1 + n,
           *p3 = p2 + n;
 
@@ -160,8 +162,7 @@ interppixel_NDC_clamped(const ImageBuf& buf, float x, float y,
     yfrac = floorfrac(y, &ytexel);
 
     // Get the four texels
-    ImageBuf::ConstIterator<SRCTYPE> it(
-        buf, ROI(xtexel, xtexel + 2, ytexel, ytexel + 2), ImageBuf::WrapClamp);
+    it.rerange(xtexel, xtexel + 2, ytexel, ytexel + 2, 0, 1, wrapmode);
     for (int c = 0; c < n; ++c)
         p0[c] = it[c];
     ++it;
@@ -181,8 +182,10 @@ interppixel_NDC_clamped(const ImageBuf& buf, float x, float y,
         // wrong will tend to over-represent the high latitudes in
         // low-res MIP levels.  We fold the area weighting into our
         // linear interpolation by adjusting yfrac.
-        int ynext = OIIO::clamp(ytexel + 1, buf.ymin(), buf.ymax());
-        ytexel    = OIIO::clamp(ytexel, buf.ymin(), buf.ymax());
+        int ymin  = spec.y;
+        int ymax  = ymin + spec.height - 1;
+        int ynext = OIIO::clamp(ytexel + 1, ymin, ymax);
+        ytexel    = OIIO::clamp(ytexel, ymin, ymax);
         float w0  = (1.0f - yfrac)
                    * sinf((float)M_PI * (ytexel + 0.5f) / (float)fh);
         float w1 = yfrac * sinf((float)M_PI * (ynext + 0.5f) / (float)fh);
@@ -218,15 +221,16 @@ resize_block_(ImageBuf& dst, const ImageBuf& src, ROI roi, bool envlatlmode)
     float yscale    = 1.0f / (float)dstspec.full_height;
     int nchannels   = dst.nchannels();
     OIIO_DASSERT(dst.spec().format == TypeFloat);
+    ImageBuf::WrapMode wrapmode = src_is_crop ? ImageBuf::WrapBlack
+                                              : ImageBuf::WrapClamp;
+    ImageBuf::ConstIterator<SRCTYPE> srcit(src);
     ImageBuf::Iterator<float> d(dst, roi);
     for (int y = y0; y < y1; ++y) {
         float t = (y + 0.5f) * yscale + yoffset;
         for (int x = x0; x < x1; ++x, ++d) {
             float s = (x + 0.5f) * xscale + xoffset;
-            if (src_is_crop)
-                src.interppixel_NDC(s, t, pel);
-            else
-                interppixel_NDC_clamped<SRCTYPE>(src, s, t, pel, envlatlmode);
+            interppixel_NDC<SRCTYPE>(src, s, t, pel, envlatlmode, srcit,
+                                     wrapmode);
             for (int c = 0; c < nchannels; ++c)
                 d[c] = pel[c];
         }
@@ -311,7 +315,7 @@ resize_block(ImageBuf& dst, const ImageBuf& src, ROI roi, bool envlatlmode,
         !envlatlmode &&                          // not latlong wrap mode
         roi.xbegin == 0 &&                       // Region x at origin
         dstspec.width == roi.width() &&          // Full width ROI
-        dstspec.width == (srcspec.width / 2) &&  // Src is 2x resize
+        (2 * dstspec.width) == srcspec.width &&  // Src is 2x resize
         dstspec.format == srcspec.format &&      // Same formats
         dstspec.x == 0 && dstspec.y == 0 &&      // Not a crop or overscan
         srcspec.x == 0 && srcspec.y == 0) {
@@ -387,15 +391,16 @@ lightprobe_to_envlatl(ImageBuf& dst, const ImageBuf& src, bool y_is_up,
 
         span<float> pixel = OIIO_ALLOCA_SPAN(float, nchannels);
         float dw = dstspec.width, dh = dstspec.height;
+        ImageBuf::ConstIterator<SRCTYPE> srcit(src);
         for (ImageBuf::Iterator<float> d(dst, roi); !d.done(); ++d) {
             Imath::V3f V = latlong_to_dir((d.x() + 0.5f) / dw,
                                           (dh - 1.0f - d.y() + 0.5f) / dh,
                                           y_is_up);
             float r      = M_1_PI * acosf(V[2]) / hypotf(V[0], V[1]);
             float u      = (V[0] * r + 1.0f) * 0.5f;
             float v      = (V[1] * r + 1.0f) * 0.5f;
-            interppixel_NDC_clamped<SRCTYPE>(src, float(u), float(v), pixel,
-                                             false);
+            interppixel_NDC<SRCTYPE>(src, float(u), float(v), pixel, false,
+                                     srcit, ImageBuf::WrapClamp);
             for (int c = roi.chbegin; c < roi.chend; ++c)
                 d[c] = pixel[c];
         }