[x265] [PATCH] asm: code for pixel_sse_sp_24x32

[murugan at multicorewareinc.com]

# HG changeset patch
# User Murugan Vairavel <murugan at multicorewareinc.com>
# Date 1385559769 -19800
#      Wed Nov 27 19:12:49 2013 +0530
# Branch stable
# Node ID 2f66bfd1237aa1c2b859cd31e1084e84524dea59
# Parent  f8eef9dd7e408748e505a7a7d3253b32cb6702bd
asm: code for pixel_sse_sp_24x32

diff -r f8eef9dd7e40 -r 2f66bfd1237a source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Wed Nov 27 19:11:33 2013 +0530
+++ b/source/common/x86/asm-primitives.cpp	Wed Nov 27 19:12:49 2013 +0530
@@ -651,6 +651,7 @@
         p.sse_sp[LUMA_16x16] = x265_pixel_ssd_sp_16x16_sse4;
         p.sse_sp[LUMA_16x32] = x265_pixel_ssd_sp_16x32_sse4;
         p.sse_sp[LUMA_16x64] = x265_pixel_ssd_sp_16x64_sse4;
+        p.sse_sp[LUMA_24x32] = x265_pixel_ssd_sp_24x32_sse4;
         p.sse_sp[LUMA_32x8] = x265_pixel_ssd_sp_32x8_sse4;
         p.sse_sp[LUMA_32x16] = x265_pixel_ssd_sp_32x16_sse4;
         p.sse_sp[LUMA_32x24] = x265_pixel_ssd_sp_32x24_sse4;
diff -r f8eef9dd7e40 -r 2f66bfd1237a source/common/x86/pixel-a.asm
--- a/source/common/x86/pixel-a.asm	Wed Nov 27 19:11:33 2013 +0530
+++ b/source/common/x86/pixel-a.asm	Wed Nov 27 19:12:49 2013 +0530
@@ -1534,6 +1534,51 @@
     RET
 
 ;-----------------------------------------------------------------------------
+; int pixel_ssd_sp_24x32( int16_t *, intptr_t, uint8_t *, intptr_t )
+;-----------------------------------------------------------------------------
+INIT_XMM sse4
+cglobal pixel_ssd_sp_24x32, 4, 7, 8, src1, stride1, src2, stride2
+    pxor     m6,     m6
+    pxor     m7,     m7
+    add      r1,     r1
+    lea      r4,     [r1 * 2]
+    mov      r5,     r0
+    mov      r6,     r2
+    call     pixel_ssd_sp_16x16_internal
+    lea      r0,     [r0 + r4]
+    lea      r2,     [r2 + 2 * r3]
+    call     pixel_ssd_sp_16x16_internal
+    lea      r0,     [r5 + 32]
+    lea      r2,     [r6 + 16]
+    lea      r4,     [r1 * 3]
+    lea      r5,     [r3 * 3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    lea      r0,     [r0 + 4 * r1]
+    lea      r2,     [r2 + 4 * r3]
+    call     pixel_ssd_sp_8x4_internal
+    HADDD    m7,     m1
+    movd     eax,    m7
+    RET
+
+;-----------------------------------------------------------------------------
 ; int pixel_ssd_32x8( uint8_t *, intptr_t, uint8_t *, intptr_t )
 ;-----------------------------------------------------------------------------
 INIT_XMM sse4
diff -r f8eef9dd7e40 -r 2f66bfd1237a source/common/x86/pixel.h
--- a/source/common/x86/pixel.h	Wed Nov 27 19:11:33 2013 +0530
+++ b/source/common/x86/pixel.h	Wed Nov 27 19:12:49 2013 +0530
@@ -411,6 +411,7 @@
 int x265_pixel_ssd_sp_16x16_sse4(int16_t *, intptr_t, pixel *, intptr_t);
 int x265_pixel_ssd_sp_16x32_sse4(int16_t *, intptr_t, pixel *, intptr_t);
 int x265_pixel_ssd_sp_16x64_sse4(int16_t *, intptr_t, pixel *, intptr_t);
+int x265_pixel_ssd_sp_24x32_sse4(int16_t *, intptr_t, pixel *, intptr_t);
 int x265_pixel_ssd_sp_32x8_sse4(int16_t *, intptr_t, pixel *, intptr_t);
 int x265_pixel_ssd_sp_32x16_sse4(int16_t *, intptr_t, pixel *, intptr_t);
 int x265_pixel_ssd_sp_32x24_sse4(int16_t *, intptr_t, pixel *, intptr_t);