From 2eb550826dd1d28718fcd2ed5f73f63c57ed26ad Mon Sep 17 00:00:00 2001
From: Shaun Inman <shaun@panic.com>
Date: Thu, 16 Feb 2023 20:31:52 -0500
Subject: [PATCH] added mixed scalers, using them proved difficult

will revisit later
---
 src/common/scaler_neon.c | 81 ++++++++++++++++++++++++++++++++++++++++
 src/common/scaler_neon.h |  6 +++
 src/minarch/main.c       | 17 ++++++++-
 3 files changed, 102 insertions(+), 2 deletions(-)

diff --git a/src/common/scaler_neon.c b/src/common/scaler_neon.c
index 19c3803..91dbde5 100755
--- a/src/common/scaler_neon.c
+++ b/src/common/scaler_neon.c
@@ -1039,3 +1039,84 @@ void scale6x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32
 		memcpy((uint8_t*)dst+dp*5, dst, swl);
 	}
 }
+
+// 
+//	mixed scalers
+//
+
+void scale1x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp) {
+	if (!sw||!sh) { return; }
+	uint32_t swl = sw*sizeof(uint16_t);
+	if (!sp) { sp = swl; } if (!dp) { dp = swl*1; }
+	for (; sh>0; sh--, src=(uint8_t*)src+sp, dst=(uint8_t*)dst+dp*2) {
+		memcpy(dst, src, swl);
+		memcpy((uint8_t*)dst+dp, dst, swl);
+	}
+}
+
+void scale2x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp) {
+	if (!sw||!sh) { return; }
+	uint32_t x, dx, pix, dpix1, dpix2, swl = sw*sizeof(uint16_t);
+	if (!sp) { sp = swl; } swl*=2; if (!dp) { dp = swl; }
+	for (; sh>0; sh--, src=(uint8_t*)src+sp, dst=(uint8_t*)dst+dp*1) {
+		uint32_t *s = (uint32_t* __restrict)src;
+		uint32_t *d = (uint32_t* __restrict)dst;
+		for (x=dx=0; x<(sw/2); x++, dx+=2) {
+			pix = s[x];
+			dpix1=(pix & 0x0000FFFF)|(pix<<16);
+			dpix2=(pix & 0xFFFF0000)|(pix>>16);
+			d[dx] = dpix1; d[dx+1] = dpix2;
+		}
+		if (sw&1) {
+			uint16_t *s16 = (uint16_t*)s;
+			uint16_t pix16 = s16[x*2];
+			d[dx] = pix16|(pix16<<16);
+		}
+	}
+}
+
+void scale2x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp) {
+	if (!sw||!sh) { return; }
+	uint32_t x, dx, pix, dpix1, dpix2, swl = sw*sizeof(uint16_t);
+	if (!sp) { sp = swl; } swl*=2; if (!dp) { dp = swl; }
+	for (; sh>0; sh--, src=(uint8_t*)src+sp, dst=(uint8_t*)dst+dp*3) {
+		uint32_t *s = (uint32_t* __restrict)src;
+		uint32_t *d = (uint32_t* __restrict)dst;
+		for (x=dx=0; x<(sw/2); x++, dx+=2) {
+			pix = s[x];
+			dpix1=(pix & 0x0000FFFF)|(pix<<16);
+			dpix2=(pix & 0xFFFF0000)|(pix>>16);
+			d[dx] = dpix1; d[dx+1] = dpix2;
+		}
+		if (sw&1) {
+			uint16_t *s16 = (uint16_t*)s;
+			uint16_t pix16 = s16[x*2];
+			d[dx] = pix16|(pix16<<16);
+		}
+		memcpy((uint8_t*)dst+dp*1, dst, swl);
+		memcpy((uint8_t*)dst+dp*2, dst, swl);
+	}
+}
+
+void scale4x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp) {
+	if (!sw||!sh) { return; }
+	uint32_t x, dx, pix, dpix1, dpix2, swl = sw*sizeof(uint16_t);
+	if (!sp) { sp = swl; } swl*=4; if (!dp) { dp = swl; }
+	for (; sh>0; sh--, src=(uint8_t*)src+sp, dst=(uint8_t*)dst+dp*2) {
+		uint32_t *s = (uint32_t* __restrict)src;
+		uint32_t *d = (uint32_t* __restrict)dst;
+		for (x=dx=0; x<(sw/2); x++, dx+=4) {
+			pix = s[x];
+			dpix1=(pix & 0x0000FFFF)|(pix<<16);
+			dpix2=(pix & 0xFFFF0000)|(pix>>16);
+			d[dx] = dpix1; d[dx+1] = dpix1; d[dx+2] = dpix2; d[dx+3] = dpix2;
+		}
+		if (sw&1) {
+			uint16_t *s16 = (uint16_t*)s;
+			uint16_t pix16 = s16[x*2];
+			dpix1 = pix16|(pix16<<16);
+			d[dx] = dpix1; d[dx+1] = dpix1;
+		}
+		memcpy((uint8_t*)dst+dp*1, dst, swl);
+	}
+}
\ No newline at end of file
diff --git a/src/common/scaler_neon.h b/src/common/scaler_neon.h
index cdbd9f9..7ca4670 100755
--- a/src/common/scaler_neon.h
+++ b/src/common/scaler_neon.h
@@ -51,4 +51,10 @@ void scale5x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32
 void scale6x_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
 void scale6x_c32(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
 
+// mixed scalers
+void scale1x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
+void scale2x1_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
+void scale2x3_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
+void scale4x2_c16(void* __restrict src, void* __restrict dst, uint32_t sw, uint32_t sh, uint32_t sp, uint32_t dp);
+
 #endif
diff --git a/src/minarch/main.c b/src/minarch/main.c
index 011790a..bb002ad 100644
--- a/src/minarch/main.c
+++ b/src/minarch/main.c
@@ -2242,7 +2242,6 @@ static void selectScaler_PAR(int width, int height, int pitch) {
 			}
 		}
 	}
-	
 	//////////////////////////////
 	
 	// DEBUG HUD
@@ -2266,8 +2265,8 @@ static void selectScaler_AR(int width, int height, int pitch) {
 	if (scale>2) scale = 4; // TODO: pillar/letterboxing at 3x produces vertical banding (some kind of alignment issue?)
 
 	// reduce scale if we don't have enough memory to accomodate it
-	// TODO: some resolutions are getting through here unadjusted? oh maybe because of aspect ratio adjustments below? revisit
 	// scaled width and height can't be greater than our fixed page width or height
+	// TODO: some resolutions are getting through here unadjusted? oh maybe because of aspect ratio adjustments below? revisit
 	while (src_w * scale * FIXED_BPP * src_h * scale > PAGE_SIZE || src_w * scale > PAGE_WIDTH || src_h * scale > PAGE_HEIGHT) scale -= 1; 
 	
 	int dst_w = src_w * scale;
@@ -2317,6 +2316,20 @@ static void selectScaler_AR(int width, int height, int pitch) {
 		case 3: renderer.scaler = scale3x_n16; break;
 		case 2: renderer.scaler = scale2x_n16; break;
 		default: renderer.scaler = scale1x_n16; break;
+		// case 4:
+		// 	if (scale_y==2) renderer.scaler = scale4x2_c16;
+		// 	else renderer.scaler = scale4x_c16;
+		// 	break;
+		// case 3:
+		// 	if (scale_x==2) renderer.scaler = scale2x3_c16;
+		// 	else renderer.scaler = scale3x_c16;
+		// 	break;
+		// case 2:
+		// 	if (scale_x==1) renderer.scaler = scale1x2_c16;
+		// 	else if (scale_y==1) renderer.scaler = scale2x1_c16;
+		// 	else renderer.scaler = scale2x_c16;
+		// break;
+		// default: renderer.scaler = scale1x_n16; break;
 	}
 	
 	if (scaler_surface) SDL_FreeSurface(scaler_surface);