From 1ba907209fe89a8a58b4c85adc7354d37e9030a1 Mon Sep 17 00:00:00 2001
From: Behdad Esfahbod <behdad@behdad.org>
Date: Sun, 23 Feb 2025 21:34:54 -0700
Subject: [PATCH] [VARC] Speed up MultiVarData::get_delta

---
 src/hb-open-type.hh        | 121 +++++++++++++++++++++++++++++++++++++
 src/hb-ot-layout-common.hh |  13 +---
 2 files changed, 123 insertions(+), 11 deletions(-)

diff --git a/src/hb-open-type.hh b/src/hb-open-type.hh
index 3083e0d96..7dab842c2 100644
--- a/src/hb-open-type.hh
+++ b/src/hb-open-type.hh
@@ -1854,6 +1854,121 @@ struct TupleValues
       return it;
     }
   };
+
+  struct fetcher_t
+  {
+    fetcher_t (const unsigned char *p_, unsigned len_)
+	      : p (p_), end (p_ + len_) {}
+
+    private:
+    const unsigned char *p;
+    const unsigned char * const end;
+    signed run_count = 0;
+    unsigned width = 0;
+
+    bool ensure_run ()
+    {
+      if (likely (run_count > 0)) return true;
+
+      if (unlikely (p >= end))
+      {
+        run_count = 0;
+	return false;
+      }
+
+      unsigned control = *p++;
+      run_count = (control & VALUE_RUN_COUNT_MASK) + 1;
+      width = control & VALUES_SIZE_MASK;
+      switch (width)
+      {
+        case VALUES_ARE_ZEROS: width = 0; break;
+	case VALUES_ARE_BYTES: width = HBINT8::static_size;  break;
+	case VALUES_ARE_WORDS: width = HBINT16::static_size; break;
+	case VALUES_ARE_LONGS: width = HBINT32::static_size; break;
+	default: assert (false);
+      }
+
+      if (unlikely (p + run_count * width > end))
+      {
+	run_count = 0;
+	return false;
+      }
+
+      return true;
+    }
+
+    void skip (unsigned n)
+    {
+      while (n)
+      {
+	if (unlikely (!ensure_run ()))
+	  return;
+	unsigned i = hb_min (n, (unsigned) run_count);
+	run_count -= i;
+	n -= i;
+	p += i * width;
+      }
+    }
+
+    template <bool scaled>
+    void _add_to (hb_array_t<float> out, float scale = 1.0f)
+    {
+      unsigned n = out.length;
+
+      for (unsigned i = 0; i < n;)
+      {
+	if (unlikely (!ensure_run ()))
+	  break;
+	unsigned count = hb_min (n - i, (unsigned) run_count);
+	switch (width)
+	{
+	  case 1:
+	  {
+	    const auto *pp = (const HBINT8 *) p;
+	    for (unsigned j = 0; j < count; j++)
+	      out.arrayZ[i + j] += scaled ? *pp++ * scale : *pp++;
+	  }
+	  break;
+	  case 2:
+	  {
+	    const auto *pp = (const HBINT16 *) p;
+	    for (unsigned j = 0; j < count; j++)
+	      out.arrayZ[i + j] += scaled ? *pp++ * scale : *pp++;
+	  }
+	  break;
+	  case 4:
+	  {
+	    const auto *pp = (const HBINT32 *) p;
+	    for (unsigned j = 0; j < count; j++)
+	      out.arrayZ[i + j] += scaled ? *pp++ * scale : *pp++;
+	  }
+	  break;
+	}
+	p += count * width;
+	run_count -= count;
+	i += count;
+      }
+    }
+
+    public:
+    void add_to (hb_array_t<float> out, float scale = 1.0f)
+    {
+      unsigned n = out.length;
+
+      if (scale == 0.0f)
+      {
+        skip (n);
+	return;
+      }
+
+#ifndef HB_OPTIMIZE_SIZE
+      if (scale == 1.0f)
+        _add_to<false> (out);
+      else
+#endif
+        _add_to<true> (out, scale);
+    }
+  };
 };
 
 struct TupleList : CFF2Index
@@ -1863,6 +1978,12 @@ struct TupleList : CFF2Index
     auto bytes = CFF2Index::operator [] (i);
     return TupleValues::iter_t (bytes.arrayZ, bytes.length);
   }
+
+  TupleValues::fetcher_t fetcher (unsigned i) const
+  {
+    auto bytes = CFF2Index::operator [] (i);
+    return TupleValues::fetcher_t (bytes.arrayZ, bytes.length);
+  }
 };
 
 
diff --git a/src/hb-ot-layout-common.hh b/src/hb-ot-layout-common.hh
index 87e40bc1d..1dfda05e3 100644
--- a/src/hb-ot-layout-common.hh
+++ b/src/hb-ot-layout-common.hh
@@ -3147,23 +3147,14 @@ struct MultiVarData
   {
     auto &deltaSets = StructAfter<decltype (deltaSetsX)> (regionIndices);
 
-    auto values_iter = deltaSets[inner];
-
+    auto values_iter = deltaSets.fetcher (inner);
     unsigned regionCount = regionIndices.len;
-    unsigned count = out.length;
     for (unsigned regionIndex = 0; regionIndex < regionCount; regionIndex++)
     {
       float scalar = regions.evaluate (regionIndices.arrayZ[regionIndex],
 				       coords, coord_count,
 				       cache);
-      if (scalar == 1.f)
-	for (unsigned i = 0; i < count; i++)
-	  out.arrayZ[i] += *values_iter++;
-      else if (scalar)
-	for (unsigned i = 0; i < count; i++)
-	  out.arrayZ[i] += *values_iter++ * scalar;
-      else
-        values_iter += count;
+      values_iter.add_to (out, scalar);
     }
   }