From 6ca25e44aa55b8930d14ad313fa1ad5c94c4f8bf Mon Sep 17 00:00:00 2001
From: Jeffrey Pfau <jeffrey@endrift.com>
Date: Sat, 26 Apr 2014 03:31:53 -0700
Subject: [PATCH] Add some NEON resizing code

---
 src/platform/sdl/sw-main.c | 49 ++++++++++++++++++++++++-----
 src/util/arm-algo.S        | 63 +++++++++++++++++++++++++++++++++++++-
 2 files changed, 103 insertions(+), 9 deletions(-)

diff --git a/src/platform/sdl/sw-main.c b/src/platform/sdl/sw-main.c
index bda2e25cf..d0b9239e4 100644
--- a/src/platform/sdl/sw-main.c
+++ b/src/platform/sdl/sw-main.c
@@ -18,6 +18,11 @@
 #include <signal.h>
 #include <sys/time.h>
 
+#ifdef __ARM_NEON
+void _neon2x(void* dest, void* src, int width, int height);
+void _neon4x(void* dest, void* src, int width, int height);
+#endif
+
 struct SoftwareRenderer {
 	struct GBAVideoSoftwareRenderer d;
 	struct GBASDLAudio audio;
@@ -26,9 +31,11 @@ struct SoftwareRenderer {
 	SDL_Window* window;
 	SDL_Texture* tex;
 	SDL_Renderer* sdlRenderer;
+#else
+	int ratio;
+#endif
 	int viewportWidth;
 	int viewportHeight;
-#endif
 };
 
 static int _GBASDLInit(struct SoftwareRenderer* renderer);
@@ -47,6 +54,9 @@ int main(int argc, char** argv) {
 		return 1;
 	}
 
+	renderer.viewportWidth = opts.width;
+	renderer.viewportHeight = opts.height;
+
 	if (!_GBASDLInit(&renderer)) {
 		return 1;
 	}
@@ -65,8 +75,6 @@ int main(int argc, char** argv) {
 	GBAMapOptionsToContext(&opts, &context);
 
 #if SDL_VERSION_ATLEAST(2, 0, 0)
-	renderer.viewportWidth = opts.width;
-	renderer.viewportHeight = opts.height;
 	renderer.events.fullscreen = opts.fullscreen;
 	renderer.window = SDL_CreateWindow("GBAc", SDL_WINDOWPOS_UNDEFINED, SDL_WINDOWPOS_UNDEFINED, renderer.viewportWidth, renderer.viewportHeight, SDL_WINDOW_OPENGL | (SDL_WINDOW_FULLSCREEN_DESKTOP * renderer.events.fullscreen));
 	SDL_GetWindowSize(renderer.window, &renderer.viewportWidth, &renderer.viewportHeight);
@@ -91,12 +99,23 @@ int main(int argc, char** argv) {
 #else
 	SDL_Surface* surface = SDL_GetVideoSurface();
 	SDL_LockSurface(surface);
-	renderer.d.outputBuffer = surface->pixels;
+
+	renderer.ratio = renderer.viewportWidth / VIDEO_HORIZONTAL_PIXELS;
+	if (renderer.ratio == 1) {
+		renderer.d.outputBuffer = surface->pixels;
 #ifdef COLOR_16_BIT
-	renderer.d.outputBufferStride = surface->pitch / 2;
+		renderer.d.outputBufferStride = surface->pitch / 2;
 #else
-	renderer.d.outputBufferStride = surface->pitch / 4;
+		renderer.d.outputBufferStride = surface->pitch / 4;
 #endif
+	} else {
+#ifdef COLOR_16_BIT
+		renderer.d.outputBuffer = malloc(240 * 160 * 2);
+#else
+		renderer.d.outputBuffer = malloc(240 * 160 * 4);
+#endif
+		renderer.d.outputBufferStride = 240;
+	}
 #endif
 
 	GBAThreadStart(&context);
@@ -128,9 +147,9 @@ static int _GBASDLInit(struct SoftwareRenderer* renderer) {
 
 #if !SDL_VERSION_ATLEAST(2, 0, 0)
 #ifdef COLOR_16_BIT
-	SDL_SetVideoMode(240, 160, 16, SDL_DOUBLEBUF | SDL_HWSURFACE);
+	SDL_SetVideoMode(renderer->viewportWidth, renderer->viewportHeight, 16, SDL_DOUBLEBUF | SDL_HWSURFACE);
 #else
-	SDL_SetVideoMode(240, 160, 32, SDL_DOUBLEBUF | SDL_HWSURFACE);
+	SDL_SetVideoMode(renderer->viewportWidth, renderer->viewportHeight, 32, SDL_DOUBLEBUF | SDL_HWSURFACE);
 #endif
 #endif
 
@@ -156,6 +175,20 @@ static void _GBASDLRunloop(struct GBAThread* context, struct SoftwareRenderer* r
 			renderer->d.outputBufferStride /= 4;
 #endif
 #else
+			switch (renderer->ratio) {
+#if defined(__ARM_NEON) && COLOR_16_BIT
+			case 2:
+				_neon2x(surface->pixels, renderer->d.outputBuffer, 240, 160);
+				break;
+			case 4:
+				_neon4x(surface->pixels, renderer->d.outputBuffer, 240, 160);
+				break;
+#endif
+			case 1:
+				break;
+			default:
+				abort();
+			}
 			SDL_UnlockSurface(surface);
 			SDL_Flip(surface);
 			SDL_LockSurface(surface);
diff --git a/src/util/arm-algo.S b/src/util/arm-algo.S
index 48c3903b1..eb07ae384 100644
--- a/src/util/arm-algo.S
+++ b/src/util/arm-algo.S
@@ -26,6 +26,67 @@ strh r6, [r8], #2
 strh r7, [r8], #2
 subs r10, #8
 bne .L1
-.L9:
 pop {r4-r10}
 bx lr
+
+# r0: Destination
+# r1: Source
+# r2: Width
+# r3: Height
+.global _neon2x
+_neon2x:
+push {r4-r5}
+lsl r4, r2, #2
+.n20:
+mov r2, r4, lsr #4
+add r5, r0, r4
+.n21:
+vld2.32 {d0[], d1[]}, [r1]!
+vmov d2, d0
+vmov d3, d1
+vzip.16 d0, d2
+vzip.16 d1, d3
+vst1.32 {q0}, [r0]!
+vst1.32 {q0}, [r5]!
+subs r2, #1
+bne .n21
+subs r3, #1
+mov r0, r5
+bne .n20
+pop {r4-r5}
+bx lr
+
+.global _neon4x
+_neon4x:
+push {r4-r7}
+lsl r4, r2, #3
+.n40:
+mov r2, r4, lsr #5
+add r5, r0, r4
+add r6, r5, r4
+add r7, r6, r4
+.n41:
+vld4.16 {d0[], d1[], d2[], d3[]}, [r1]!
+vst1.16 {d0}, [r0]!
+vst1.16 {d0}, [r5]!
+vst1.16 {d0}, [r6]!
+vst1.16 {d0}, [r7]!
+vst1.16 {d1}, [r0]!
+vst1.16 {d1}, [r5]!
+vst1.16 {d1}, [r6]!
+vst1.16 {d1}, [r7]!
+vst1.16 {d2}, [r0]!
+vst1.16 {d2}, [r5]!
+vst1.16 {d2}, [r6]!
+vst1.16 {d2}, [r7]!
+vst1.16 {d3}, [r0]!
+vst1.16 {d3}, [r5]!
+vst1.16 {d3}, [r6]!
+vst1.16 {d3}, [r7]!
+subs r2, #1
+bne .n41
+subs r3, #1
+mov r0, r7
+bne .n40
+pop {r4-r7}
+bx lr