From 6338f8d9a2043672ad58bcf9eb0ea83cf13c8d82 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigrid=20Solveig=20Hafl=C3=ADnud=C3=B3ttir?=
 <ftrvxmtrx@gmail.com>
Date: Sun, 26 Dec 2021 07:32:35 +0100
Subject: [PATCH] ppu_aarch64: initial version of ppu_redraw using Neon

---
 src/devices/ppu_aarch64.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)
 create mode 100644 src/devices/ppu_aarch64.c

diff --git a/src/devices/ppu_aarch64.c b/src/devices/ppu_aarch64.c
new file mode 100644
index 0000000..b6cf37a
--- /dev/null
+++ b/src/devices/ppu_aarch64.c
@@ -0,0 +1,28 @@
+#include <arm_neon.h>
+#include "ppu.h"
+
+void
+ppu_redraw(Ppu *p, Uint32 *screen)
+{
+	/* FIXME(sigrid): do this better */
+	Uint32 *rgba = __builtin_assume_aligned(screen, 16);
+	Uint8 *fg = __builtin_assume_aligned(p->fg.pixels, 16);
+	Uint8 *bg = __builtin_assume_aligned(p->bg.pixels, 16);
+	Uint8 *palette = __builtin_assume_aligned((Uint8*)p->palette, 16);
+	uint8x16x4_t pal = vld4q_u8(palette); enum { R, G, B, A };
+	int i;
+
+	for(i = 0; i < p->width * p->height; i += 16, fg += 16, bg += 16, rgba += 16) {
+		uint8x16_t fg8 = vld1q_u8(fg);
+		uint8x16_t bg8 = vld1q_u8(bg);
+		uint8x16_t bgmask = vceqzq_u8(fg8);
+		uint8x16_t px8 = vorrq_u8(vandq_u8(bg8, bgmask), vandq_u8(fg8, vceqzq_u8(bgmask)));
+		uint8x16x4_t px = {
+			vqtbl1q_u8(pal.val[R], px8),
+			vqtbl1q_u8(pal.val[G], px8),
+			vqtbl1q_u8(pal.val[B], px8),
+			vqtbl1q_u8(pal.val[A], px8),
+		};
+		vst4q_u8((uint8_t*)rgba, px);
+	}
+}