Use advantage of SSE2 instructions in gaussian blur node

This gives around 30% of speedup for gaussian blur node.

Pretty much straightforward implementation inside the node
itself, but needed to implement some additional things:

- Aligned malloc. It's needed to load data onto SSE registers
  faster. based on the aligned_malloc() from Libmv with
  some additional trickery going on to support arbitrary
  alignment (this magic is needed because of MemHead).

  In the practice only 16bit alignment is supported because
  of the lack of aligned malloc with arbitrary alignment
  for OSX. Not a bit deal for now because we need 16 bytes
  alignment at this moment only. Could be tweaked further
  later.

- Memory buffers in compositor are now aligned to 16 bytes.
  Should be harmless for non-SSE cases too. just mentioning.

Reviewers: campbellbarton, lukastoenne, jbakker

Reviewed By: campbellbarton

CC: lockal

Differential Revision: https://developer.blender.org/D564

12 files changed, 362 insertions, 22 deletions

diff --git a/intern/guardedalloc/MEM_guardedalloc.h b/intern/guardedalloc/MEM_guardedalloc.h
index 4fb68965338..8c5ad77b8b6 100644
--- a/intern/guardedalloc/MEM_guardedalloc.h
+++ b/intern/guardedalloc/MEM_guardedalloc.h
@@ -120,6 +120,12 @@ extern "C" {
 	extern void *(*MEM_mallocN)(size_t len, const char *str) /* ATTR_MALLOC */ ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
 
 	/**
+	 * Allocate an aligned block of memory of size len, with tag name str. The
+	 * name must be a static, because only a pointer to it is stored !
+	 * */
+	extern void *(*MEM_mallocN_aligned)(size_t len, size_t alignment, const char *str) /* ATTR_MALLOC */ ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(3);
+
+	/**
 	 * Same as callocN, clears memory and uses mmap (disk cached) if supported.
 	 * Can be free'd with MEM_freeN as usual.
 	 * */
diff --git a/intern/guardedalloc/intern/mallocn.c b/intern/guardedalloc/intern/mallocn.c
index e85fba7a6d0..b0d252cca14 100644
--- a/intern/guardedalloc/intern/mallocn.c
+++ b/intern/guardedalloc/intern/mallocn.c
@@ -41,6 +41,7 @@ void *(*MEM_reallocN_id)(void *vmemh, size_t len, const char *str) = MEM_lockfre
 void *(*MEM_recallocN_id)(void *vmemh, size_t len, const char *str) = MEM_lockfree_recallocN_id;
 void *(*MEM_callocN)(size_t len, const char *str) = MEM_lockfree_callocN;
 void *(*MEM_mallocN)(size_t len, const char *str) = MEM_lockfree_mallocN;
+void *(*MEM_mallocN_aligned)(size_t len, size_t alignment, const char *str) = MEM_lockfree_mallocN_aligned;
 void *(*MEM_mapallocN)(size_t len, const char *str) = MEM_lockfree_mapallocN;
 void (*MEM_printmemlist_pydict)(void) = MEM_lockfree_printmemlist_pydict;
 void (*MEM_printmemlist)(void) = MEM_lockfree_printmemlist;
@@ -60,6 +61,40 @@ uintptr_t (*MEM_get_peak_memory)(void) = MEM_lockfree_get_peak_memory;
 const char *(*MEM_name_ptr)(void *vmemh) = MEM_lockfree_name_ptr;
 #endif
 
+void *aligned_malloc(size_t size, size_t alignment)
+{
+#ifdef _WIN32
+	return _aligned_malloc(size, alignment);
+#elif defined(__APPLE__)
+	/* On Mac OS X, both the heap and the stack are guaranteed 16-byte aligned so
+	 * they work natively with SSE types with no further work.
+	 */
+	assert(alignment == 16);
+	return malloc(size);
+#elif defined(__FreeBSD__) || defined(__NetBSD__)
+	void *result;
+
+	if (posix_memalign(&result, alignment, size)) {
+		/* non-zero means allocation error
+		 * either no allocation or bad alignment value
+		 */
+		return NULL;
+	}
+	return result;
+#else  /* This is for Linux. */
+	return memalign(alignment, size);
+#endif
+}
+
+void aligned_free(void *ptr)
+{
+#ifdef _WIN32
+	_aligned_free(ptr);
+#else
+	free(ptr);
+#endif
+}
+
 void MEM_use_guarded_allocator(void)
 {
 	MEM_allocN_len = MEM_guarded_allocN_len;
@@ -69,6 +104,7 @@ void MEM_use_guarded_allocator(void)
 	MEM_recallocN_id = MEM_guarded_recallocN_id;
 	MEM_callocN = MEM_guarded_callocN;
 	MEM_mallocN = MEM_guarded_mallocN;
+	MEM_mallocN_aligned = MEM_guarded_mallocN_aligned;
 	MEM_mapallocN = MEM_guarded_mapallocN;
 	MEM_printmemlist_pydict = MEM_guarded_printmemlist_pydict;
 	MEM_printmemlist = MEM_guarded_printmemlist;
diff --git a/intern/guardedalloc/intern/mallocn_guarded_impl.c b/intern/guardedalloc/intern/mallocn_guarded_impl.c
index 172c79d50cd..206390e0710 100644
--- a/intern/guardedalloc/intern/mallocn_guarded_impl.c
+++ b/intern/guardedalloc/intern/mallocn_guarded_impl.c
@@ -113,7 +113,10 @@ typedef struct MemHead {
 	const char *name;
 	const char *nextname;
 	int tag2;
-	int mmap;  /* if true, memory was mmapped */
+	short mmap;  /* if true, memory was mmapped */
+	short alignment;  /* if non-zero aligned alloc was used
+	                   * and alignment is stored here.
+	                   */
 #ifdef DEBUG_MEMCOUNTER
 	int _count;
 #endif
@@ -128,6 +131,8 @@ typedef struct MemHead {
 #endif
 } MemHead;
 
+typedef MemHead MemHeadAligned;
+
 /* for openmp threading asserts, saves time troubleshooting
  * we may need to extend this if blender code starts using MEM_
  * functions inside OpenMP correctly with omp_set_lock() */
@@ -187,7 +192,7 @@ static const char *check_memlist(MemHead *memh);
 
 #define MEMNEXT(x) \
 	((MemHead *)(((char *) x) - ((char *) &(((MemHead *)0)->next))))
-	
+
 /* --------------------------------------------------------------------- */
 /* vars                                                                  */
 /* --------------------------------------------------------------------- */
@@ -325,10 +330,12 @@ void *MEM_guarded_dupallocN(const void *vmemh)
 		memh--;
 
 #ifndef DEBUG_MEMDUPLINAME
-		if (memh->mmap)
+		if (UNLIKELY(memh->mmap))
+			newp = MEM_guarded_mapallocN(memh->len, "dupli_mapalloc");
+		else if (LIKELY(memh->alignment == 0))
 			newp = MEM_guarded_mapallocN(memh->len, "dupli_mapalloc");
 		else
-			newp = MEM_guarded_mallocN(memh->len, "dupli_alloc");
+			newp = MEM_guarded_mallocN_aligned(memh->len, (size_t) memh->alignment, "dupli_alloc");
 
 		if (newp == NULL) return NULL;
 #else
@@ -336,14 +343,18 @@ void *MEM_guarded_dupallocN(const void *vmemh)
 			MemHead *nmemh;
 			char *name = malloc(strlen(memh->name) + 24);
 
-			if (memh->mmap) {
+			if (UNLIKELY(memh->mmap)) {
 				sprintf(name, "%s %s", "dupli_mapalloc", memh->name);
 				newp = MEM_guarded_mapallocN(memh->len, name);
 			}
-			else {
+			else if (LIKELY(memh->alignment == 0)) {
 				sprintf(name, "%s %s", "dupli_alloc", memh->name);
 				newp = MEM_guarded_mallocN(memh->len, name);
 			}
+			else {
+				sprintf(name, "%s %s", "dupli_alloc", memh->name);
+				newp = MEM_guarded_mallocN_aligned(memh->len, (size_t) memh->alignment, name);
+			}
 
 			if (newp == NULL) return NULL;
 
@@ -368,7 +379,13 @@ void *MEM_guarded_reallocN_id(void *vmemh, size_t len, const char *str)
 		MemHead *memh = vmemh;
 		memh--;
 
-		newp = MEM_guarded_mallocN(len, memh->name);
+		if (LIKELY(memh->alignment == 0)) {
+			newp = MEM_guarded_mallocN(len, memh->name);
+		}
+		else {
+			newp = MEM_guarded_mallocN_aligned(len, (size_t) memh->alignment, memh->name);
+		}
+
 		if (newp) {
 			if (len < memh->len) {
 				/* shrink */
@@ -397,7 +414,13 @@ void *MEM_guarded_recallocN_id(void *vmemh, size_t len, const char *str)
 		MemHead *memh = vmemh;
 		memh--;
 
-		newp = MEM_guarded_mallocN(len, memh->name);
+		if (LIKELY(memh->alignment == 0)) {
+			newp = MEM_guarded_mallocN(len, memh->name);
+		}
+		else {
+			newp = MEM_guarded_mallocN_aligned(len, (size_t) memh->alignment, memh->name);
+		}
+
 		if (newp) {
 			if (len < memh->len) {
 				/* shrink */
@@ -464,6 +487,7 @@ static void make_memhead_header(MemHead *memh, size_t len, const char *str)
 	memh->nextname = NULL;
 	memh->len = len;
 	memh->mmap = 0;
+	memh->alignment = 0;
 	memh->tag2 = MEMTAG2;
 
 #ifdef DEBUG_MEMDUPLINAME
@@ -514,6 +538,54 @@ void *MEM_guarded_mallocN(size_t len, const char *str)
 	return NULL;
 }
 
+void *MEM_guarded_mallocN_aligned(size_t len, size_t alignment, const char *str)
+{
+	MemHead *memh;
+
+	/* It's possible that MemHead's size is not properly aligned,
+	 * do extra padding to deal with this.
+	 *
+	 * We only support small alignments which fits into short in
+	 * order to save some bits in MemHead structure.
+	 */
+	short extra_padding = (short)MEMHEAD_ALIGN_PADDING(alignment);
+
+	/* Huge alignment values doesn't make sense and they
+	 * wouldn't fit into 'short' used in the MemHead.
+	 */
+	assert(alignment < 1024);
+
+	/* We only support alignment to a power of two. */
+	assert(IS_POW2(alignment));
+
+	len = SIZET_ALIGN_4(len);
+
+	memh = (MemHead *)aligned_malloc(len + (size_t)extra_padding + sizeof(MemHead) + sizeof(MemTail), alignment);
+
+	if (LIKELY(memh)) {
+		/* We keep padding in the beginning of MemHead,
+		 * this way it's always possible to get MemHead
+		 * from the data pointer.
+		 */
+		memh = (MemHead *)((char *)memh + extra_padding);
+
+		make_memhead_header(memh, len, str);
+		memh->alignment = (short) alignment;
+		if (UNLIKELY(malloc_debug_memset && len))
+			memset(memh + 1, 255, len);
+
+#ifdef DEBUG_MEMCOUNTER
+		if (_mallocn_count == DEBUG_MEMCOUNTER_ERROR_VAL)
+			memcount_raise(__func__);
+		memh->_count = _mallocn_count++;
+#endif
+		return (++memh);
+	}
+	print_error("aligned_malloc returns null: len=" SIZET_FORMAT " in %s, total %u\n",
+	            SIZET_ARG(len), str, (unsigned int) mem_in_use);
+	return NULL;
+}
+
 void *MEM_guarded_callocN(size_t len, const char *str)
 {
 	MemHead *memh;
@@ -953,7 +1025,12 @@ static void rem_memblock(MemHead *memh)
 	else {
 		if (UNLIKELY(malloc_debug_memset && memh->len))
 			memset(memh + 1, 255, memh->len);
-		free(memh);
+		if (LIKELY(memh->alignment == 0)) {
+			free(memh);
+		}
+		else {
+			aligned_free(MEMHEAD_REAL_PTR(memh));
+		}
 	}
 }
 
diff --git a/intern/guardedalloc/intern/mallocn_intern.h b/intern/guardedalloc/intern/mallocn_intern.h
index 7c8922dd407..a69bcf3d27b 100644
--- a/intern/guardedalloc/intern/mallocn_intern.h
+++ b/intern/guardedalloc/intern/mallocn_intern.h
@@ -85,6 +85,35 @@
 #  define UNLIKELY(x)     (x)
 #endif
 
+#if !defined(__APPLE__) && !defined(__FreeBSD__) && !defined(__NetBSD__)
+// Needed for memalign on Linux and _aligned_alloc on Windows.
+#  ifdef FREE_WINDOWS
+/* make sure _aligned_malloc is included */
+#    ifdef __MSVCRT_VERSION__
+#      undef __MSVCRT_VERSION__
+#    endif
+
+#    define __MSVCRT_VERSION__ 0x0700
+#  endif  // FREE_WINDOWS
+
+#  include <malloc.h>
+#else
+// Apple's malloc is 16-byte aligned, and does not have malloc.h, so include
+// stdilb instead.
+#  include <cstdlib>
+#endif
+
+#define IS_POW2(a) (((a) & ((a) - 1)) == 0)
+
+/* Extra padding which needs to be applied on MemHead to make it aligned. */
+#define MEMHEAD_ALIGN_PADDING(alignment) ((size_t)alignment - (sizeof(MemHeadAligned) % (size_t)alignment))
+
+/* Real pointer returned by the malloc or aligned_alloc. */
+#define MEMHEAD_REAL_PTR(memh) ((char *)memh - MEMHEAD_ALIGN_PADDING(memh->alignment))
+
+void *aligned_malloc(size_t size, size_t alignment);
+void aligned_free(void *ptr);
+
 /* Prototypes for counted allocator functions */
 size_t MEM_lockfree_allocN_len(const void *vmemh) ATTR_WARN_UNUSED_RESULT;
 void MEM_lockfree_freeN(void *vmemh);
@@ -93,6 +122,7 @@ void *MEM_lockfree_reallocN_id(void *vmemh, size_t len, const char *UNUSED(str))
 void *MEM_lockfree_recallocN_id(void *vmemh, size_t len, const char *UNUSED(str))  ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(2);
 void *MEM_lockfree_callocN(size_t len, const char *UNUSED(str))  ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
 void *MEM_lockfree_mallocN(size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
+void *MEM_lockfree_mallocN_aligned(size_t len, size_t alignment, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(3);
 void *MEM_lockfree_mapallocN(size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
 void MEM_lockfree_printmemlist_pydict(void);
 void MEM_lockfree_printmemlist(void);
@@ -119,6 +149,7 @@ void *MEM_guarded_reallocN_id(void *vmemh, size_t len, const char *UNUSED(str))
 void *MEM_guarded_recallocN_id(void *vmemh, size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(2);
 void *MEM_guarded_callocN(size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
 void *MEM_guarded_mallocN(size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
+void *MEM_guarded_mallocN_aligned(size_t len, size_t alignment, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(3);
 void *MEM_guarded_mapallocN(size_t len, const char *UNUSED(str)) ATTR_MALLOC ATTR_WARN_UNUSED_RESULT ATTR_ALLOC_SIZE(1) ATTR_NONNULL(2);
 void MEM_guarded_printmemlist_pydict(void);
 void MEM_guarded_printmemlist(void);
diff --git a/intern/guardedalloc/intern/mallocn_lockfree_impl.c b/intern/guardedalloc/intern/mallocn_lockfree_impl.c
index 6fc01807af3..c76caff0d74 100644
--- a/intern/guardedalloc/intern/mallocn_lockfree_impl.c
+++ b/intern/guardedalloc/intern/mallocn_lockfree_impl.c
@@ -46,6 +46,11 @@ typedef struct MemHead {
 	size_t len;
 } MemHead;
 
+typedef struct MemHeadAligned {
+	short alignment;
+	size_t len;
+} MemHeadAligned;
+
 static unsigned int totblock = 0;
 static size_t mem_in_use = 0, mmap_in_use = 0, peak_mem = 0;
 static bool malloc_debug_memset = false;
@@ -54,9 +59,17 @@ static void (*error_callback)(const char *) = NULL;
 static void (*thread_lock_callback)(void) = NULL;
 static void (*thread_unlock_callback)(void) = NULL;
 
+enum {
+	MEMHEAD_MMAP_FLAG = 1,
+	MEMHEAD_ALIGN_FLAG = 2,
+};
+
 #define MEMHEAD_FROM_PTR(ptr) (((MemHead*) vmemh) - 1)
 #define PTR_FROM_MEMHEAD(memhead) (memhead + 1)
-#define MEMHEAD_IS_MMAP(memhead) ((memhead)->len & (size_t) 1)
+#define MEMHEAD_ALIGNED_FROM_PTR(ptr) (((MemHeadAligned*) vmemh) - 1)
+#define PTR_FROM_MEMHEAD_ALIGNED(memhead) (memhead + 1)
+#define MEMHEAD_IS_MMAP(memhead) ((memhead)->len & (size_t) MEMHEAD_MMAP_FLAG)
+#define MEMHEAD_IS_ALIGNED(memhead) ((memhead)->len & (size_t) MEMHEAD_ALIGN_FLAG)
 
 #ifdef __GNUC__
 __attribute__ ((format(printf, 1, 2)))
@@ -93,7 +106,7 @@ static void mem_unlock_thread(void)
 size_t MEM_lockfree_allocN_len(const void *vmemh)
 {
 	if (vmemh) {
-		return MEMHEAD_FROM_PTR(vmemh)->len & ~((size_t) 1);
+		return MEMHEAD_FROM_PTR(vmemh)->len & ~((size_t) (MEMHEAD_MMAP_FLAG | MEMHEAD_ALIGN_FLAG));
 	}
 	else {
 		return 0;
@@ -124,7 +137,13 @@ void MEM_lockfree_freeN(void *vmemh)
 		if (UNLIKELY(malloc_debug_memset && len)) {
 			memset(memh + 1, 255, len);
 		}
-		free(memh);
+		if (UNLIKELY(MEMHEAD_IS_ALIGNED(memh))) {
+			MemHeadAligned *memh_aligned = MEMHEAD_ALIGNED_FROM_PTR(vmemh);
+			aligned_free(MEMHEAD_REAL_PTR(memh_aligned));
+		}
+		else {
+			free(memh);
+		}
 	}
 }
 
@@ -134,9 +153,16 @@ void *MEM_lockfree_dupallocN(const void *vmemh)
 	if (vmemh) {
 		MemHead *memh = MEMHEAD_FROM_PTR(vmemh);
 		const size_t prev_size = MEM_allocN_len(vmemh);
-		if (MEMHEAD_IS_MMAP(memh)) {
+		if (UNLIKELY(MEMHEAD_IS_MMAP(memh))) {
 			newp = MEM_lockfree_mapallocN(prev_size, "dupli_mapalloc");
 		}
+		else if (UNLIKELY(MEMHEAD_IS_ALIGNED(memh))) {
+			MemHeadAligned *memh_aligned = MEMHEAD_ALIGNED_FROM_PTR(vmemh);
+			newp = MEM_lockfree_mallocN_aligned(
+				prev_size,
+				(size_t)memh_aligned->alignment,
+				"dupli_malloc");
+		}
 		else {
 			newp = MEM_lockfree_mallocN(prev_size, "dupli_malloc");
 		}
@@ -150,9 +176,20 @@ void *MEM_lockfree_reallocN_id(void *vmemh, size_t len, const char *str)
 	void *newp = NULL;
 
 	if (vmemh) {
+		MemHead *memh = MEMHEAD_FROM_PTR(vmemh);
 		size_t old_len = MEM_allocN_len(vmemh);
 
-		newp = MEM_lockfree_mallocN(len, "realloc");
+		if (LIKELY(!MEMHEAD_IS_ALIGNED(memh))) {
+			newp = MEM_lockfree_mallocN(len, "realloc");
+		}
+		else {
+			MemHeadAligned *memh_aligned = MEMHEAD_ALIGNED_FROM_PTR(vmemh);
+			newp = MEM_lockfree_mallocN_aligned(
+				old_len,
+				(size_t)memh_aligned->alignment,
+				"realloc");
+		}
+
 		if (newp) {
 			if (len < old_len) {
 				/* shrink */
@@ -178,9 +215,19 @@ void *MEM_lockfree_recallocN_id(void *vmemh, size_t len, const char *str)
 	void *newp = NULL;
 
 	if (vmemh) {
+		MemHead *memh = MEMHEAD_FROM_PTR(vmemh);
 		size_t old_len = MEM_allocN_len(vmemh);
 
-		newp = MEM_lockfree_mallocN(len, "recalloc");
+		if (LIKELY(!MEMHEAD_IS_ALIGNED(memh))) {
+			newp = MEM_lockfree_mallocN(len, "recalloc");
+		}
+		else {
+			MemHeadAligned *memh_aligned = MEMHEAD_ALIGNED_FROM_PTR(vmemh);
+			newp = MEM_lockfree_mallocN_aligned(old_len,
+			                                    (size_t)memh_aligned->alignment,
+			                                    "recalloc");
+		}
+
 		if (newp) {
 			if (len < old_len) {
 				/* shrink */
@@ -256,6 +303,57 @@ void *MEM_lockfree_mallocN(size_t len, const char *str)
 	return NULL;
 }
 
+void *MEM_lockfree_mallocN_aligned(size_t len, size_t alignment, const char *str)
+{
+	MemHeadAligned *memh;
+
+	/* It's possible that MemHead's size is not properly aligned,
+	 * do extra padding to deal with this.
+	 *
+	 * We only support small alignments which fits into short in
+	 * order to save some bits in MemHead structure.
+	 */
+	size_t extra_padding = MEMHEAD_ALIGN_PADDING(alignment);
+
+	/* Huge alignment values doesn't make sense and they
+	 * wouldn't fit into 'short' used in the MemHead.
+	 */
+	assert(alignment < 1024);
+
+	/* We only support alignment to a power of two. */
+	assert(IS_POW2(alignment));
+
+	len = SIZET_ALIGN_4(len);
+
+	memh = (MemHeadAligned *)aligned_malloc(
+		len + extra_padding + sizeof(MemHeadAligned), alignment);
+
+	if (LIKELY(memh)) {
+		/* We keep padding in the beginning of MemHead,
+		 * this way it's always possible to get MemHead
+		 * from the data pointer.
+		 */
+		memh = (MemHeadAligned *)((char *)memh + extra_padding);
+
+		if (UNLIKELY(malloc_debug_memset && len)) {
+			memset(memh + 1, 255, len);
+		}
+
+		memh->len = len | (size_t) MEMHEAD_ALIGN_FLAG;
+		memh->alignment = (short) alignment;
+		atomic_add_u(&totblock, 1);
+		atomic_add_z(&mem_in_use, len);
+
+		/* TODO(sergey): Not strictly speaking thread-safe. */
+		peak_mem = mem_in_use > peak_mem ? mem_in_use : peak_mem;
+
+		return PTR_FROM_MEMHEAD(memh);
+	}
+	print_error("Malloc returns null: len=" SIZET_FORMAT " in %s, total %u\n",
+	            SIZET_ARG(len), str, (unsigned int) mem_in_use);
+	return NULL;
+}
+
 void *MEM_lockfree_mapallocN(size_t len, const char *str)
 {
 	MemHead *memh;
@@ -279,7 +377,7 @@ void *MEM_lockfree_mapallocN(size_t len, const char *str)
 #endif
 
 	if (memh != (MemHead *)-1) {
-		memh->len = len | (size_t) 1;
+		memh->len = len | (size_t) MEMHEAD_MMAP_FLAG;
 		atomic_add_u(&totblock, 1);
 		atomic_add_z(&mem_in_use, len);
 		atomic_add_z(&mmap_in_use, len);
diff --git a/source/blender/compositor/intern/COM_MemoryBuffer.cpp b/source/blender/compositor/intern/COM_MemoryBuffer.cpp
index 04828bfe3f8..c1916f4a68f 100644
--- a/source/blender/compositor/intern/COM_MemoryBuffer.cpp
+++ b/source/blender/compositor/intern/COM_MemoryBuffer.cpp
@@ -46,7 +46,7 @@ MemoryBuffer::MemoryBuffer(MemoryProxy *memoryProxy, unsigned int chunkNumber, r
 	BLI_rcti_init(&this->m_rect, rect->xmin, rect->xmax, rect->ymin, rect->ymax);
 	this->m_memoryProxy = memoryProxy;
 	this->m_chunkNumber = chunkNumber;
-	this->m_buffer = (float *)MEM_mallocN(sizeof(float) * determineBufferSize() * COM_NUMBER_OF_CHANNELS, "COM_MemoryBuffer");
+	this->m_buffer = (float *)MEM_mallocN_aligned(sizeof(float) * determineBufferSize() * COM_NUMBER_OF_CHANNELS, 16, "COM_MemoryBuffer");
 	this->m_state = COM_MB_ALLOCATED;
 	this->m_datatype = COM_DT_COLOR;
 	this->m_chunkWidth = this->m_rect.xmax - this->m_rect.xmin;
@@ -57,7 +57,7 @@ MemoryBuffer::MemoryBuffer(MemoryProxy *memoryProxy, rcti *rect)
 	BLI_rcti_init(&this->m_rect, rect->xmin, rect->xmax, rect->ymin, rect->ymax);
 	this->m_memoryProxy = memoryProxy;
 	this->m_chunkNumber = -1;
-	this->m_buffer = (float *)MEM_mallocN(sizeof(float) * determineBufferSize() * COM_NUMBER_OF_CHANNELS, "COM_MemoryBuffer");
+	this->m_buffer = (float *)MEM_mallocN_aligned(sizeof(float) * determineBufferSize() * COM_NUMBER_OF_CHANNELS, 16, "COM_MemoryBuffer");
 	this->m_state = COM_MB_TEMPORARILY;
 	this->m_datatype = COM_DT_COLOR;
 	this->m_chunkWidth = this->m_rect.xmax - this->m_rect.xmin;
diff --git a/source/blender/compositor/operations/COM_BlurBaseOperation.cpp b/source/blender/compositor/operations/COM_BlurBaseOperation.cpp
index e7af9319f88..d5aafc7c2ae 100644
--- a/source/blender/compositor/operations/COM_BlurBaseOperation.cpp
+++ b/source/blender/compositor/operations/COM_BlurBaseOperation.cpp
@@ -91,6 +91,18 @@ float *BlurBaseOperation::make_gausstab(float rad, int size)
 	return gausstab;
 }
 
+#ifdef __SSE2__
+__m128 *BlurBaseOperation::convert_gausstab_sse(const float *gausstab, float rad, int size)
+{
+	int n = 2 * size + 1;
+	__m128 *gausstab_sse = (__m128 *) MEM_mallocN_aligned(sizeof(__m128) * n, 16, "gausstab sse");
+	for (int i = 0; i < n; ++i) {
+		gausstab_sse[i] = _mm_set1_ps(gausstab[i]);
+	}
+	return gausstab_sse;
+}
+#endif
+
 /* normalized distance from the current (inverted so 1.0 is close and 0.0 is far)
  * 'ease' is applied after, looks nicer */
 float *BlurBaseOperation::make_dist_fac_inverse(float rad, int size, int falloff)
diff --git a/source/blender/compositor/operations/COM_BlurBaseOperation.h b/source/blender/compositor/operations/COM_BlurBaseOperation.h
index 052a525ef2c..e97dd4d766d 100644
--- a/source/blender/compositor/operations/COM_BlurBaseOperation.h
+++ b/source/blender/compositor/operations/COM_BlurBaseOperation.h
@@ -27,6 +27,10 @@
 
 #define MAX_GAUSSTAB_RADIUS 30000
 
+#ifdef __SSE2__
+#  include <emmintrin.h>
+#endif
+
 class BlurBaseOperation : public NodeOperation, public QualityStepHelper {
 private:
 
@@ -34,6 +38,9 @@ protected:
 
 	BlurBaseOperation(DataType data_type);
 	float *make_gausstab(float rad, int size);
+#ifdef __SSE2__
+	__m128 *convert_gausstab_sse(const float *gaustab, float rad, int size);
+#endif
 	float *make_dist_fac_inverse(float rad, int size, int falloff);
 
 	void updateSize();
diff --git a/source/blender/compositor/operations/COM_GaussianXBlurOperation.cpp b/source/blender/compositor/operations/COM_GaussianXBlurOperation.cpp
index d08924ca4ef..0aefba3bb7c 100644
--- a/source/blender/compositor/operations/COM_GaussianXBlurOperation.cpp
+++ b/source/blender/compositor/operations/COM_GaussianXBlurOperation.cpp
@@ -31,6 +31,9 @@ extern "C" {
 GaussianXBlurOperation::GaussianXBlurOperation() : BlurBaseOperation(COM_DT_COLOR)
 {
 	this->m_gausstab = NULL;
+#ifdef __SSE2__
+	this->m_gausstab_sse = NULL;
+#endif
 	this->m_filtersize = 0;
 }
 
@@ -54,8 +57,14 @@ void GaussianXBlurOperation::initExecution()
 	if (this->m_sizeavailable) {
 		float rad = max_ff(m_size * m_data.sizex, 0.0f);
 		m_filtersize = min_ii(ceil(rad), MAX_GAUSSTAB_RADIUS);
-		
+
+		/* TODO(sergey): De-duplicate with the case below and Y blur. */
 		this->m_gausstab = BlurBaseOperation::make_gausstab(rad, m_filtersize);
+#ifdef __SSE2__
+		this->m_gausstab_sse = BlurBaseOperation::convert_gausstab_sse(this->m_gausstab,
+		                                                               rad,
+		                                                               m_filtersize);
+#endif
 	}
 }
 
@@ -65,8 +74,13 @@ void GaussianXBlurOperation::updateGauss()
 		updateSize();
 		float rad = max_ff(m_size * m_data.sizex, 0.0f);
 		m_filtersize = min_ii(ceil(rad), MAX_GAUSSTAB_RADIUS);
-		
+
 		this->m_gausstab = BlurBaseOperation::make_gausstab(rad, m_filtersize);
+#ifdef __SSE2__
+		this->m_gausstab_sse = BlurBaseOperation::convert_gausstab_sse(this->m_gausstab,
+		                                                               rad,
+		                                                               m_filtersize);
+#endif
 	}
 }
 
@@ -88,12 +102,25 @@ void GaussianXBlurOperation::executePixel(float output[4], int x, int y, void *d
 	int step = getStep();
 	int offsetadd = getOffsetAdd();
 	int bufferindex = ((xmin - bufferstartx) * 4) + ((ymin - bufferstarty) * 4 * bufferwidth);
+
+#ifdef __SSE2__
+	__m128 accum_r = _mm_load_ps(color_accum);
+	for (int nx = xmin, index = (xmin - x) + this->m_filtersize; nx < xmax; nx += step, index += step) {
+		__m128 reg_a = _mm_load_ps(&buffer[bufferindex]);
+		reg_a = _mm_mul_ps(reg_a, this->m_gausstab_sse[index]);
+		accum_r = _mm_add_ps(accum_r, reg_a);
+		multiplier_accum += this->m_gausstab[index];
+		bufferindex += offsetadd;
+	}
+	_mm_store_ps(color_accum, accum_r);
+#else
 	for (int nx = xmin, index = (xmin - x) + this->m_filtersize; nx < xmax; nx += step, index += step) {
 		const float multiplier = this->m_gausstab[index];
 		madd_v4_v4fl(color_accum, &buffer[bufferindex], multiplier);
 		multiplier_accum += multiplier;
 		bufferindex += offsetadd;
 	}
+#endif
 	mul_v4_v4fl(output, color_accum, 1.0f / multiplier_accum);
 }
 
@@ -105,6 +132,12 @@ void GaussianXBlurOperation::deinitExecution()
 		MEM_freeN(this->m_gausstab);
 		this->m_gausstab = NULL;
 	}
+#ifdef __SSE2__
+	if (this->m_gausstab_sse) {
+		MEM_freeN(this->m_gausstab_sse);
+		this->m_gausstab_sse = NULL;
+	}
+#endif
 
 	deinitMutex();
 }
diff --git a/source/blender/compositor/operations/COM_GaussianXBlurOperation.h b/source/blender/compositor/operations/COM_GaussianXBlurOperation.h
index 6442f214138..e391320a007 100644
--- a/source/blender/compositor/operations/COM_GaussianXBlurOperation.h
+++ b/source/blender/compositor/operations/COM_GaussianXBlurOperation.h
@@ -28,6 +28,9 @@
 class GaussianXBlurOperation : public BlurBaseOperation {
 private:
 	float *m_gausstab;
+#ifdef __SSE2__
+	__m128 *m_gausstab_sse;
+#endif
 	int m_filtersize;
 	void updateGauss();
 public:
diff --git a/source/blender/compositor/operations/COM_GaussianYBlurOperation.cpp b/source/blender/compositor/operations/COM_GaussianYBlurOperation.cpp
index 8216b79372f..a05a1ab6a23 100644
--- a/source/blender/compositor/operations/COM_GaussianYBlurOperation.cpp
+++ b/source/blender/compositor/operations/COM_GaussianYBlurOperation.cpp
@@ -31,6 +31,9 @@ extern "C" {
 GaussianYBlurOperation::GaussianYBlurOperation() : BlurBaseOperation(COM_DT_COLOR)
 {
 	this->m_gausstab = NULL;
+#ifdef __SSE2__
+	this->m_gausstab_sse = NULL;
+#endif
 	this->m_filtersize = 0;
 }
 
@@ -54,8 +57,13 @@ void GaussianYBlurOperation::initExecution()
 	if (this->m_sizeavailable) {
 		float rad = max_ff(m_size * m_data.sizey, 0.0f);
 		m_filtersize = min_ii(ceil(rad), MAX_GAUSSTAB_RADIUS);
-		
+
 		this->m_gausstab = BlurBaseOperation::make_gausstab(rad, m_filtersize);
+#ifdef __SSE2__
+		this->m_gausstab_sse = BlurBaseOperation::convert_gausstab_sse(this->m_gausstab,
+		                                                               rad,
+		                                                               m_filtersize);
+#endif
 	}
 }
 
@@ -65,8 +73,13 @@ void GaussianYBlurOperation::updateGauss()
 		updateSize();
 		float rad = max_ff(m_size * m_data.sizey, 0.0f);
 		m_filtersize = min_ii(ceil(rad), MAX_GAUSSTAB_RADIUS);
-		
+
 		this->m_gausstab = BlurBaseOperation::make_gausstab(rad, m_filtersize);
+#ifdef __SSE2__
+		this->m_gausstab_sse = BlurBaseOperation::convert_gausstab_sse(this->m_gausstab,
+		                                                               rad,
+		                                                               m_filtersize);
+#endif
 	}
 }
 
@@ -88,6 +101,20 @@ void GaussianYBlurOperation::executePixel(float output[4], int x, int y, void *d
 	int index;
 	int step = getStep();
 	const int bufferIndexx = ((xmin - bufferstartx) * 4);
+
+#ifdef __SSE2__
+	__m128 accum_r = _mm_load_ps(color_accum);
+	for (int ny = ymin; ny < ymax; ny += step) {
+		index = (ny - y) + this->m_filtersize;
+		int bufferindex = bufferIndexx + ((ny - bufferstarty) * 4 * bufferwidth);
+		const float multiplier = this->m_gausstab[index];
+		__m128 reg_a = _mm_load_ps(&buffer[bufferindex]);
+		reg_a = _mm_mul_ps(reg_a, this->m_gausstab_sse[index]);
+		accum_r = _mm_add_ps(accum_r, reg_a);
+		multiplier_accum += multiplier;
+	}
+	_mm_store_ps(color_accum, accum_r);
+#else
 	for (int ny = ymin; ny < ymax; ny += step) {
 		index = (ny - y) + this->m_filtersize;
 		int bufferindex = bufferIndexx + ((ny - bufferstarty) * 4 * bufferwidth);
@@ -95,6 +122,7 @@ void GaussianYBlurOperation::executePixel(float output[4], int x, int y, void *d
 		madd_v4_v4fl(color_accum, &buffer[bufferindex], multiplier);
 		multiplier_accum += multiplier;
 	}
+#endif
 	mul_v4_v4fl(output, color_accum, 1.0f / multiplier_accum);
 }
 
@@ -106,6 +134,12 @@ void GaussianYBlurOperation::deinitExecution()
 		MEM_freeN(this->m_gausstab);
 		this->m_gausstab = NULL;
 	}
+#ifdef __SSE2__
+	if (this->m_gausstab_sse) {
+		MEM_freeN(this->m_gausstab_sse);
+		this->m_gausstab_sse = NULL;
+	}
+#endif
 
 	deinitMutex();
 }
diff --git a/source/blender/compositor/operations/COM_GaussianYBlurOperation.h b/source/blender/compositor/operations/COM_GaussianYBlurOperation.h
index 16503360de2..22b6562077d 100644
--- a/source/blender/compositor/operations/COM_GaussianYBlurOperation.h
+++ b/source/blender/compositor/operations/COM_GaussianYBlurOperation.h
@@ -28,6 +28,9 @@
 class GaussianYBlurOperation : public BlurBaseOperation {
 private:
 	float *m_gausstab;
+#ifdef __SSE2__
+	__m128 *m_gausstab_sse;
+#endif
 	int m_filtersize;
 	void updateGauss();
 public: