From 6613d80d72931b13cc008c3dcc8ee90a39bec8f5 Mon Sep 17 00:00:00 2001
From: Graham Madarasz <graham@lindenlab.com>
Date: Mon, 11 Mar 2013 14:19:05 -0700
Subject: Clean up moving llalignedarray and fast memcpy to llcommon
---
indra/llcommon/llmemory.h | 24 +++++++++++++++++++-----
1 file changed, 19 insertions(+), 5 deletions(-)
(limited to 'indra/llcommon')
diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h
index 4938775e2b..61e30f11cc 100644
--- a/indra/llcommon/llmemory.h
+++ b/indra/llcommon/llmemory.h
@@ -201,24 +201,36 @@ inline void ll_memcpy_nonaliased_aligned_16(char* __restrict dst, const char* __
if (bytes > 64)
{
+
+ // Find start of 64b aligned area within block
+ //
void* begin_64 = LL_NEXT_ALIGNED_ADDRESS_64(dst);
//at least 64 bytes before the end of the destination, switch to 16 byte copies
void* end_64 = end-64;
-
+
+ // Prefetch the head of the 64b area now
+ //
_mm_prefetch((char*)begin_64, _MM_HINT_NTA);
_mm_prefetch((char*)begin_64 + 64, _MM_HINT_NTA);
_mm_prefetch((char*)begin_64 + 128, _MM_HINT_NTA);
_mm_prefetch((char*)begin_64 + 192, _MM_HINT_NTA);
-
+
+ // Copy 16b chunks until we're 64b aligned
+ //
while (dst < begin_64)
{
_mm_store_ps((F32*)dst, _mm_load_ps((F32*)src));
- dst += 4;
- src += 4;
+ dst += 16;
+ src += 16;
}
-
+
+ // Copy 64b chunks up to your tail
+ //
+ // might be good to shmoo the 512b prefetch offset
+ // (characterize performance for various values)
+ //
while (dst < end_64)
{
_mm_prefetch((char*)src + 512, _MM_HINT_NTA);
@@ -232,6 +244,8 @@ inline void ll_memcpy_nonaliased_aligned_16(char* __restrict dst, const char* __
}
}
+ // Copy remainder 16b tail chunks (or ALL 16b chunks for sub-64b copies)
+ //
while (dst < end)
{
_mm_store_ps((F32*)dst, _mm_load_ps((F32*)src));
--
cgit v1.2.3