From d874010d09c985faf6b22a585f9fe13b232fc90a Mon Sep 17 00:00:00 2001
From: Kleis Auke Wolthuizen <github@kleisauke.nl>
Date: Sat, 17 Sep 2022 13:26:00 +0200
Subject: [PATCH] convi/reducev: use convsuswb in Orc path (#3053)

* reducev: use convsuswb in Orc path

Saves a few instructions.

* convi: use convsuswb in Orc path

Saves a few instructions.
---
 libvips/convolution/convi.c  | 13 +------------
 libvips/resample/reducev.cpp | 13 +------------
 2 files changed, 2 insertions(+), 24 deletions(-)

diff --git a/libvips/convolution/convi.c b/libvips/convolution/convi.c
index f4eb9c55..eef38f48 100644
--- a/libvips/convolution/convi.c
+++ b/libvips/convolution/convi.c
@@ -420,8 +420,6 @@ vips_convi_compile_clip( VipsConvi *convi )
 	VipsVector *v;
 	char rnd[256];
 	char exp[256];
-	char c0[256];
-	char c255[256];
 	char off[256];
 
 	convi->vector = v = vips_vector_new( "convi", 1 );
@@ -442,16 +440,7 @@ vips_convi_compile_clip( VipsConvi *convi )
 	CONST( off, offset, 2 ); 
 	ASM3( "addw", "value", "value", off );
 
-	/* You'd think "convsuswb" (convert signed 16-bit to unsigned
-	 * 8-bit with saturation) would be quicker, but it's a lot
-	 * slower.
-	 */
-	CONST( c0, 0, 2 );
-	ASM3( "maxsw", "value", c0, "value" ); 
-	CONST( c255, 255, 2 );
-	ASM3( "minsw", "value", c255, "value" ); 
-
-	ASM2( "convwb", "d1", "value" );
+	ASM2( "convsuswb", "d1", "value" );
 
 	if( !vips_vector_compile( v ) ) 
 		return( -1 );
diff --git a/libvips/resample/reducev.cpp b/libvips/resample/reducev.cpp
index f201f6fd..2b43c7a8 100644
--- a/libvips/resample/reducev.cpp
+++ b/libvips/resample/reducev.cpp
@@ -268,24 +268,13 @@ vips_reducev_compile_section( VipsReducev *reducev, Pass *pass, gboolean first )
 	if( pass->last >= reducev->n_point - 1 ) {
 		char c32[256];
 		char c6[256];
-		char c0[256];
-		char c255[256];
 
 		CONST( c32, 32, 2 );
 		ASM3( "addw", "sum", "sum", c32 );
 		CONST( c6, 6, 2 );
 		ASM3( "shrsw", "sum", "sum", c6 );
 
-		/* You'd think "convsuswb", convert signed 16-bit to unsigned
-		 * 8-bit with saturation, would be quicker, but it's a lot
-		 * slower.
-		 */
-		CONST( c0, 0, 2 );
-		ASM3( "maxsw", "sum", c0, "sum" ); 
-		CONST( c255, 255, 2 );
-		ASM3( "minsw", "sum", c255, "sum" ); 
-
-		ASM2( "convwb", "d1", "sum" );
+		ASM2( "convsuswb", "d1", "sum" );
 	}
 	else 
 		ASM2( "copyw", "d2", "sum" );