From c772deadb7c8491f08103a4aa33bd98c264e2097 Mon Sep 17 00:00:00 2001
From: veejay <>
Date: Mon, 18 Sep 2023 15:01:55 +0200
Subject: [PATCH] WIP memset neon 32

---
 veejay-current/veejay-core/libvjmem/memcpy.c | 36 +++++++++++++++++++-
 1 file changed, 35 insertions(+), 1 deletion(-)

diff --git a/veejay-current/veejay-core/libvjmem/memcpy.c b/veejay-current/veejay-core/libvjmem/memcpy.c
index 3a988c01..8d8d0af2 100644
--- a/veejay-current/veejay-core/libvjmem/memcpy.c
+++ b/veejay-current/veejay-core/libvjmem/memcpy.c
@@ -1945,6 +1945,40 @@ void memset_asimd_64(uint8_t *dst, uint8_t value, size_t size) {
         remaining_bytes--;
     }
 }
+
+void memset_asimd_32(uint8_t *dst, uint8_t value, size_t size) {
+ 	uint8x16_t value_v = vdupq_n_u8(value);
+
+    size_t num_blocks = size / 32;
+    size_t remaining_bytes = size % 32;
+
+    for (size_t i = 0; i < num_blocks; i++) {
+        vst1q_u8(dst, value_v); 
+		dst += 16;
+		vst1q_u8(dst, value_v);
+		dst += 16; 
+	}
+
+	while (remaining_bytes >= 16) {
+        vst1q_u8(dst, value_v);
+        dst += 16;
+        remaining_bytes -= 16;
+    }
+
+    while (remaining_bytes >= 8) {
+        uint64x1_t value_u64 = vdup_n_u64(*((uint64_t*)&value));
+        vst1_u8(dst, vreinterpret_u8_u64(value_u64));
+        dst += 8;
+        remaining_bytes -= 8;
+    }
+
+    while (remaining_bytes > 0) {
+        *dst = value;
+        dst++;
+        remaining_bytes--;
+    }
+}
+
 #endif
 
 static struct {
@@ -2036,7 +2070,7 @@ static struct {
 	{ "Advanced SIMD memset()", (void*) memset_asimd, 0, AV_CPU_FLAG_ARMV8 },
 	{ "Advanced SIMD memset() v4", (void*) memset_asimd_v4, 0, AV_CPU_FLAG_ARMV8 },	
 	{ "Advanced SIMD memset() with line size of 64", (void*) memset_asimd_64, 0, AV_CPU_FLAG_ARMV8 },	
-	
+	{ "Advanced SIMD memset() with line size of 32", (void*) memset_asimd_64, 0, AV_CPU_FLAG_ARMV8 },	
 //	{ "Advanced SIMD memset() v3", (void*) memset_asimd_v3, 0, AV_CPU_FLAG_ARMV8 },	
 //	{ "Advanced SIMD memset() v2", (void*) memset_asimd_v2, 0, AV_CPU_FLAG_ARMV8 },