From 057ce7a549c4ace67f5bcf94538e372005f23bde Mon Sep 17 00:00:00 2001
From: veejay <>
Date: Mon, 18 Sep 2023 01:29:36 +0200
Subject: [PATCH] ARMv8-a RPI4 memset, cacheline size WIP

---
 veejay-current/veejay-core/libvjmem/memcpy.c | 30 +++++++++++++-------
 veejay-current/veejay-core/libvjmem/vj-x86.c |  2 +-
 veejay-current/veejay-core/libvjmem/vjmem.h  |  1 +
 3 files changed, 22 insertions(+), 11 deletions(-)

diff --git a/veejay-current/veejay-core/libvjmem/memcpy.c b/veejay-current/veejay-core/libvjmem/memcpy.c
index 37e40111..1ff76a60 100644
--- a/veejay-current/veejay-core/libvjmem/memcpy.c
+++ b/veejay-current/veejay-core/libvjmem/memcpy.c
@@ -1756,28 +1756,38 @@ static void fast_memset(void * to, int val, size_t len)
 #ifdef HAVE_ARM_ASIMD
 
 void memset_asimd_v3(void *dst, uint8_t val, size_t len) {
-    uint8x16_t value = vdupq_n_u8(val);
+    uint8x16_t value = vdupq_n_u8(val);  
     uint8_t *dst_bytes = (uint8_t *)dst;
-    size_t num_blocks = len / 16;
+
+    const int CACHE_LINE_SIZE = cpu_cache_size();
+    size_t offset = (CACHE_LINE_SIZE - ((uintptr_t)dst_bytes % CACHE_LINE_SIZE)) % CACHE_LINE_SIZE;
+
+    for (size_t i = 0; i < offset; i++) {
+        *dst_bytes++ = val;
+        len--;
+    }
 
     value = vld1q_u8(&val);
 
-    for (size_t i = 0; i < num_blocks; i++) {
-        uint8x16_t dst_data = vld1q_u8(dst_bytes);
+    size_t num_lines = len / CACHE_LINE_SIZE;
 
-        dst_data = vorrq_u8(dst_data, value);
-
-        vst1q_u8(dst_bytes, dst_data);
-        dst_bytes += 16;
+    for (size_t i = 0; i < num_lines; i++) {
+        for (size_t j = 0; j < CACHE_LINE_SIZE / 16; j++) {
+            uint8x16_t dst_data = vld1q_u8(dst_bytes);
+            dst_data = vorrq_u8(dst_data, value);
+            vst1q_u8(dst_bytes, dst_data);
+            dst_bytes += 16;
+        }
     }
 
-    size_t remaining_bytes = len % 16;
-
+    size_t remaining_bytes = len % CACHE_LINE_SIZE;
     for (size_t i = 0; i < remaining_bytes; i++) {
         *dst_bytes++ = val;
     }
 }
 
+
+
 void memset_asimd(void *dst, uint8_t val, size_t len) {
     uint8x16_t value = vdupq_n_u8(val);
 	uint8_t *dst_bytes = (uint8_t *)dst;
diff --git a/veejay-current/veejay-core/libvjmem/vj-x86.c b/veejay-current/veejay-core/libvjmem/vj-x86.c
index 8fe7afb8..8935a33d 100644
--- a/veejay-current/veejay-core/libvjmem/vj-x86.c
+++ b/veejay-current/veejay-core/libvjmem/vj-x86.c
@@ -33,7 +33,7 @@ extern void yuyv_plane_init();
 extern void benchmark_tasks(int n_tasks, long n_frames, int w, int h);
 extern void init_parallel_tasks(int n_tasks);
 static int MEM_ALIGNMENT_SIZE = 0;
-static int CACHE_LINE_SIZE = 64;
+extern int CACHE_LINE_SIZE = 64;
 
 
 static int has_cpuid(void)
diff --git a/veejay-current/veejay-core/libvjmem/vjmem.h b/veejay-current/veejay-core/libvjmem/vjmem.h
index 24953d08..1f19ed10 100644
--- a/veejay-current/veejay-core/libvjmem/vjmem.h
+++ b/veejay-current/veejay-core/libvjmem/vjmem.h
@@ -65,4 +65,5 @@ extern void *vj_simple_pool_alloc( void *ptr, size_t s );
 extern void *vj_simple_pool_init( size_t s );
 extern void vj_simple_pool_reset( void *ptr );
 extern void vj_mem_destroy();
+
 #endif