From a607f257a6d4af90a534163c7dfa851007ad0e78 Mon Sep 17 00:00:00 2001
From: Kleis Auke Wolthuizen <github@kleisauke.nl>
Date: Thu, 21 Nov 2024 18:31:20 +0100
Subject: [PATCH] shrink{h,v}: add fixed-point arithmetic path for uchar images

About 9% faster.
---
 libvips/resample/shrinkh.c | 33 +++++++++++++++++++++++++++------
 libvips/resample/shrinkv.c | 36 ++++++++++++++++++++++++++++++------
 2 files changed, 57 insertions(+), 12 deletions(-)

diff --git a/libvips/resample/shrinkh.c b/libvips/resample/shrinkh.c
index 95bb619d8..574a44cf5 100644
--- a/libvips/resample/shrinkh.c
+++ b/libvips/resample/shrinkh.c
@@ -72,6 +72,25 @@ typedef VipsResampleClass VipsShrinkhClass;
 
 G_DEFINE_TYPE(VipsShrinkh, vips_shrinkh, VIPS_TYPE_RESAMPLE);
 
+/* Fixed-point arithmetic path for uchar images.
+ */
+#define UCHAR_SHRINK(BANDS) \
+	{ \
+		unsigned char *restrict p = (unsigned char *) in; \
+		unsigned char *restrict q = (unsigned char *) out; \
+\
+		for (x = 0; x < width; x++) { \
+			for (b = 0; b < BANDS; b++) { \
+				int sum = amend; \
+				for (x1 = b; x1 < ne; x1 += BANDS) \
+					sum += p[x1]; \
+				q[b] = (sum * multiplier) >> 24; \
+			} \
+			p += ne; \
+			q += BANDS; \
+		} \
+	}
+
 /* Integer shrink.
  */
 #define ISHRINK(ACC_TYPE, TYPE, BANDS) \
@@ -129,7 +148,9 @@ vips_shrinkh_gen2(VipsShrinkh *shrink, VipsRegion *out_region, VipsRegion *ir,
 	int x1, b;
 
 	switch (resample->in->BandFmt) {
-	case VIPS_FORMAT_UCHAR:
+	case VIPS_FORMAT_UCHAR: {
+		unsigned int multiplier = (1LL << 32) / ((1 << 8) * shrink->hshrink);
+
 		/* Generate a special path for 1, 3 and 4 band uchar data. The
 		 * compiler will be able to vectorise these.
 		 *
@@ -138,20 +159,20 @@ vips_shrinkh_gen2(VipsShrinkh *shrink, VipsRegion *out_region, VipsRegion *ir,
 		 */
 		switch (bands) {
 		case 1:
-			ISHRINK(int, unsigned char, 1);
+			UCHAR_SHRINK(1);
 			break;
 		case 3:
-			ISHRINK(int, unsigned char, 3);
+			UCHAR_SHRINK(3);
 			break;
 		case 4:
-			ISHRINK(int, unsigned char, 4);
+			UCHAR_SHRINK(4);
 			break;
 		default:
-			ISHRINK(int, unsigned char, bands);
+			UCHAR_SHRINK(bands);
 			break;
 		}
 		break;
-
+	}
 	case VIPS_FORMAT_CHAR:
 		ISHRINK(int, char, bands);
 		break;
diff --git a/libvips/resample/shrinkv.c b/libvips/resample/shrinkv.c
index e38f24eaf..c6c310964 100644
--- a/libvips/resample/shrinkv.c
+++ b/libvips/resample/shrinkv.c
@@ -111,6 +111,28 @@ typedef VipsResampleClass VipsShrinkvClass;
 
 G_DEFINE_TYPE(VipsShrinkv, vips_shrinkv, VIPS_TYPE_RESAMPLE);
 
+/* Fixed-point arithmetic path for uchar images.
+ */
+#define UCHAR_SHRINK(BANDS) \
+	{ \
+		unsigned char *restrict p = (unsigned char *) in; \
+		unsigned char *restrict q = (unsigned char *) out; \
+\
+		for (x = 0; x < width; x++) { \
+			for (b = 0; b < BANDS; b++) { \
+				int sum = amend; \
+				unsigned char *restrict row_ptr = p + b; \
+				for (yy = 0; yy < shrink->vshrink; yy++) { \
+					sum += *row_ptr; \
+					row_ptr += sz; \
+				} \
+				q[b] = (sum * multiplier) >> 24; \
+			} \
+			p += BANDS; \
+			q += BANDS; \
+		} \
+	}
+
 /* Integer shrink.
  */
 #define ISHRINK(ACC_TYPE, TYPE, BANDS) \
@@ -174,7 +196,9 @@ vips_shrinkv_gen2(VipsShrinkv *shrink, VipsRegion *out_region, VipsRegion *ir,
 	int yy, b;
 
 	switch (resample->in->BandFmt) {
-	case VIPS_FORMAT_UCHAR:
+	case VIPS_FORMAT_UCHAR: {
+		unsigned int multiplier = (1LL << 32) / ((1 << 8) * shrink->vshrink);
+
 		/* Generate a special path for 1, 3 and 4 band uchar data. The
 		 * compiler will be able to vectorise these.
 		 *
@@ -183,20 +207,20 @@ vips_shrinkv_gen2(VipsShrinkv *shrink, VipsRegion *out_region, VipsRegion *ir,
 		 */
 		switch (bands) {
 		case 1:
-			ISHRINK(int, unsigned char, 1);
+			UCHAR_SHRINK(1);
 			break;
 		case 3:
-			ISHRINK(int, unsigned char, 3);
+			UCHAR_SHRINK(3);
 			break;
 		case 4:
-			ISHRINK(int, unsigned char, 4);
+			UCHAR_SHRINK(4);
 			break;
 		default:
-			ISHRINK(int, unsigned char, bands);
+			UCHAR_SHRINK(bands);
 			break;
 		}
 		break;
-
+	}
 	case VIPS_FORMAT_CHAR:
 		ISHRINK(int, char, bands);
 		break;