From 7cedf7a80b2858d8a210234e86984013fdda6505 Mon Sep 17 00:00:00 2001
From: Behdad Esfahbod <behdad@behdad.org>
Date: Thu, 26 Sep 2024 13:52:12 -0600
Subject: [PATCH] [MultiVarStore] Add NEON codepath

Slows down in my testing.
---
 src/hb-ot-layout-common.hh | 33 +++++++++++++++++++++++++++++++--
 1 file changed, 31 insertions(+), 2 deletions(-)

diff --git a/src/hb-ot-layout-common.hh b/src/hb-ot-layout-common.hh
index eb9276150..accd977d2 100644
--- a/src/hb-ot-layout-common.hh
+++ b/src/hb-ot-layout-common.hh
@@ -40,6 +40,8 @@
 
 #ifdef __SSE__
 #include <xmmintrin.h>
+#elif defined(__ARM_NEON)
+#include <arm_neon.h>
 #endif
 
 
@@ -3154,7 +3156,6 @@ struct MultiVarData
     {
       if (scalar == 1.f)
       {
-	// SSE version
 	unsigned i = 0;
 	for (; i + 4 <= count; i += 4)
 	{
@@ -3167,7 +3168,6 @@ struct MultiVarData
       }
       else
       {
-	// SSE version
 	unsigned i = 0;
 	__m128 s = _mm_set1_ps (scalar);
 	for (; i + 4 <= count; i += 4)
@@ -3181,6 +3181,35 @@ struct MultiVarData
       }
       return;
     }
+#elif defined(__ARM_NEON)
+    {
+      if (scalar == 1.f)
+      {
+	unsigned i = 0;
+	for (; i + 4 <= count; i += 4)
+	{
+	  float32x4_t a = vld1q_f32 (in + i);
+	  float32x4_t b = vld1q_f32 (out + i);
+	  vst1q_f32 (out + i, vaddq_f32 (a, b));
+	}
+	for (; i < count; i++)
+	  out[i] += in[i];
+      }
+      else
+      {
+	unsigned i = 0;
+	float32x4_t s = vdupq_n_f32 (scalar);
+	for (; i + 4 <= count; i += 4)
+	{
+	  float32x4_t a = vld1q_f32 (in + i);
+	  float32x4_t b = vld1q_f32 (out + i);
+	  vst1q_f32 (out + i, vmlaq_f32 (b, a, s));
+	}
+	for (; i < count; i++)
+	  out[i] += in[i] * scalar;
+      }
+      return;
+    }
 #endif
 
     // Fallback