From 34d16392f3238a1a8d9ee035ef4c829d3ab2a1dc Mon Sep 17 00:00:00 2001
From: xy <yuzhuohuang@qq.com>
Date: Sat, 28 Jul 2012 11:00:02 +0800
Subject: [PATCH] Blur Optimization. [Part 5]

---
 src/subtitles/xy_filter.cpp          |  57 ++++++++++---
 test/unit_test/test_xy_filter.h      |  25 +++++-
 test/unit_test/xy_filter_benchmark.h | 153 ++++++++++++++++++++++++++++++++++-
 3 files changed, 220 insertions(+), 15 deletions(-)

diff --git a/src/subtitles/xy_filter.cpp b/src/subtitles/xy_filter.cpp
index 77f6531..9dee499 100644
--- a/src/subtitles/xy_filter.cpp
+++ b/src/subtitles/xy_filter.cpp
@@ -113,15 +113,15 @@ void xy_filter_c(float *dst, int width, int height, int stride, const float *fil
     sum = _mm_add_ps(sum, f4_1)
 
 
-__forceinline void xy_filter_one_line_sse_v4(float *dst, int width, const float *filter, int filter_width)
+__forceinline void xy_filter_one_line_sse_v6(float *dst, int width, const float *filter, int filter_width)
 {
     int xx_fix = width > filter_width ? 0 : filter_width - width;
     const float *filter_start = filter;
     float *dst2 = dst - filter_width;
     float *dst_endr = dst + width;
-    float *dst_end0 = dst_endr - filter_width;
+    float *dst_end0 = dst_endr - filter_width - 4;
     float *dst_endl = dst - xx_fix;
-    ASSERT(xx_fix==0 || dst_end0==dst_endl);
+    ASSERT(xx_fix==0 || dst_end0==dst_endl-4);
 
     ASSERT(filter_start == filter);
     filter_start += filter_width;
@@ -171,17 +171,49 @@ __forceinline void xy_filter_one_line_sse_v4(float *dst, int width, const float
         _mm_store_ps(dst2, sum);
     }
     ASSERT(filter_start == filter);
-    for (;dst2<dst_end0;dst2+=4)
+    for (;dst2<dst_end0;dst2+=8)
     {
-        const float *src = dst2;
+        const float *src = dst2;           
+        const float* f=filter_start;
 
-        //filter 4
+        //filter 8
         __m128 src4 = _mm_load_ps(src);/*1 2 3 4*/
+        src+=4;
+        __m128 src_5_8;
         __m128 sum = _mm_setzero_ps();
-        for(const float* f=filter_start;f<filter_end;f+=4)
+        __m128 sum2 = _mm_setzero_ps();
+        __m128 f4 = _mm_load_ps(f);
+        f+=4;
+        src_5_8 = _mm_load_ps(src);/*5 6 7 8*/
+        src+=4;
+        { XY_FILTER_4(src4, src_5_8, f4, sum); }
+        for(;f<filter_end;f+=4,src+=4)
         {
-            src+=4;
-            __m128 src_5_8 = _mm_load_ps(src);/*5 6 7 8*/
+            src4 = _mm_load_ps(src);/*1 2 3 4*/
+            __m128 tmp = src_5_8;//important! 
+            { XY_FILTER_4(tmp, src4, f4, sum2); }
+
+            f4 = _mm_load_ps(f);
+            { XY_FILTER_4(src_5_8, src4, f4, sum); }
+            src_5_8 = src4;
+        }
+        src4 = _mm_load_ps(src);/*1 2 3 4*/
+        { XY_FILTER_4(src_5_8, src4, f4, sum2); }
+
+        //store result
+        _mm_store_ps(dst2, sum);
+        _mm_store_ps(dst2+4, sum2);
+    }
+    if (dst2==dst_end0)
+    {
+        const float *src = dst2;
+        //filter 4
+        __m128 src4 = _mm_load_ps(src);//1 2 3 4
+        src+=4;
+        __m128 sum = _mm_setzero_ps();
+        for(const float* f=filter_start;f<filter_end;f+=4,src+=4)
+        {
+            __m128 src_5_8 = _mm_load_ps(src);//5 6 7 8
             __m128 f4 = _mm_load_ps(f);
 
             { XY_FILTER_4(src4, src_5_8, f4, sum); }
@@ -189,6 +221,7 @@ __forceinline void xy_filter_one_line_sse_v4(float *dst, int width, const float
         }
         //store result
         _mm_store_ps(dst2, sum);
+        dst2+=4;
     }
     for (;dst2<dst_endr;dst2+=4)//right margin
     {
@@ -221,7 +254,7 @@ __forceinline void xy_filter_one_line_sse_v4(float *dst, int width, const float
 /****
  * See @xy_filter_c
  **/
-void xy_filter_sse_v4(float *dst, int width, int height, int stride, const float *filter, int filter_width)
+void xy_filter_sse_v6(float *dst, int width, int height, int stride, const float *filter, int filter_width)
 {
     ASSERT( stride>=4*(width+filter_width) );
     ASSERT( ((stride|(4*width)|(4*filter_width)|reinterpret_cast<int>(dst)|reinterpret_cast<int>(filter))&15)==0 );
@@ -230,7 +263,7 @@ void xy_filter_sse_v4(float *dst, int width, int height, int stride, const float
     BYTE* end = dst_byte + height*stride;
     for( ; dst_byte<end; dst_byte+=stride )
     {
-        xy_filter_one_line_sse_v4(reinterpret_cast<float*>(dst_byte), width, filter, filter_width);
+        xy_filter_one_line_sse_v6(reinterpret_cast<float*>(dst_byte), width, filter, filter_width);
     }
 }
 
@@ -386,7 +419,7 @@ void xy_filter_sse(float *dst, int width, int height, int stride, const float *f
     }
     else
     {
-        xy_filter_sse_v4(dst, width, height, stride, filter, filter_width);
+        xy_filter_sse_v6(dst, width, height, stride, filter, filter_width);
     }   
 }
 
diff --git a/test/unit_test/test_xy_filter.h b/test/unit_test/test_xy_filter.h
index b0bc1c6..835de7c 100644
--- a/test/unit_test/test_xy_filter.h
+++ b/test/unit_test/test_xy_filter.h
@@ -13,6 +13,7 @@ void xy_filter_one_line_sse(float *dst, int width, const float *filter, int filt
 {
     xy_filter_sse(dst, width, 1, ((width*4)+15)&~15, filter, filter_width);
 }
+void xy_filter_one_line_sse_v6(float *dst, int width, const float *filter, int filter_width);
 
 template<typename T>
 std::ostream& GetBufString(std::ostream& os, T* buf, int len)
@@ -188,7 +189,7 @@ TEST_F(XyFilterTest, c_vs_sse)
                 data1.copy(data0);
                 data2.copy(data0);
                 xy_filter_one_line_c(data1.data, data1.w, data1.filter_f, data1.ex_filter_width);
-                xy_filter_one_line_sse(data2.data, data2.w, data2.filter_f, data2.ex_filter_width);
+                xy_filter_one_line_sse_v6(data2.data, data2.w, data2.filter_f, data2.ex_filter_width);
                 ASSERT_EQ(true, data2.operator==(data1))
                     <<LOG_VAR(i)<<LOG_VAR(WIDTH)<<LOG_VAR(FILTER_WIDTH)<<std::endl
                     <<"data0:"<<data0
@@ -199,4 +200,26 @@ TEST_F(XyFilterTest, c_vs_sse)
 }
 
 
+TEST_F(XyFilterTest, sse_vs_sse_v6)
+{
+    int LOOP_NUM = 10;
+    for(int WIDTH=4;WIDTH<256;WIDTH+=4)
+        for (int FILTER_WIDTH=3;FILTER_WIDTH<32;FILTER_WIDTH++)
+        {
+            for(int i=0;i<LOOP_NUM;i++)
+            {
+                data0.FillRandData(WIDTH, 1, FILTER_WIDTH);
+                data1.copy(data0);
+                data2.copy(data0);
+                xy_filter_one_line_sse(data1.data, data1.w, data1.filter_f, data1.ex_filter_width);
+                xy_filter_one_line_sse_v6(data2.data, data2.w, data2.filter_f, data2.ex_filter_width);
+                ASSERT_EQ(true, data2.operator==(data1))
+                    <<LOG_VAR(i)<<LOG_VAR(WIDTH)<<LOG_VAR(FILTER_WIDTH)<<std::endl
+                    <<"data0:"<<data0
+                    <<"data1:"<<data1
+                    <<"data2:"<<data2;
+            }
+        }
+}
+
 #endif // __TEST_XY_FILTER_0B6AC4E9_AA51_4EF9_B255_90792DC07DB9_H__
diff --git a/test/unit_test/xy_filter_benchmark.h b/test/unit_test/xy_filter_benchmark.h
index 8604b11..24f33ba 100644
--- a/test/unit_test/xy_filter_benchmark.h
+++ b/test/unit_test/xy_filter_benchmark.h
@@ -917,6 +917,156 @@ void xy_filter_sse_v5(float *dst, int width, int height, int stride, const float
 
 // v5 end
 
+__forceinline void xy_filter_one_line_sse_v6(float *dst, int width, const float *filter, int filter_width)
+{
+    int xx_fix = width > filter_width ? 0 : filter_width - width;
+    const float *filter_start = filter;
+    float *dst2 = dst - filter_width;
+    float *dst_endr = dst + width;
+    float *dst_end0 = dst_endr - filter_width - 4;
+    float *dst_endl = dst - xx_fix;
+    ASSERT(xx_fix==0 || dst_end0==dst_endl-4);
+
+    ASSERT(filter_start == filter);
+    filter_start += filter_width;
+    const float *filter_end = filter_start;
+
+    for (;dst2<dst_endl;dst2+=4)//left margin
+    {
+        const float *src = dst;
+        filter_start -= 4;
+
+        //filter 4
+        __m128 src4 = _mm_setzero_ps();/*1 2 3 4*/
+        __m128 sum = _mm_setzero_ps();
+        for(const float* f=filter_start;f<filter_end;f+=4,src+=4)
+        {   
+            __m128 src_5_8 = _mm_load_ps(src);/*5 6 7 8*/
+            __m128 f4 = _mm_load_ps(f);
+
+            { XY_FILTER_4(src4, src_5_8, f4, sum); }
+
+            src4 = src_5_8;
+        }
+        //store result
+        _mm_store_ps(dst2, sum);
+    }
+    for (;dst2<dst;dst2+=4)//if width < filter_width
+    {
+        const float *src = dst;
+        filter_start-=4;
+        filter_end-=4;
+
+        __m128 src4 = _mm_setzero_ps();/*1 2 3 4*/
+        __m128 sum = _mm_setzero_ps();
+        __m128 src_5_8, f4;
+        for(const float* f=filter_start;f<filter_end;f+=4,src+=4)
+        {   
+            src_5_8 = _mm_load_ps(src);/*5 6 7 8*/
+            f4 = _mm_load_ps(f);
+
+            { XY_FILTER_4(src4, src_5_8, f4, sum); }
+            src4 = src_5_8;
+        }
+        src_5_8 = _mm_setzero_ps();
+        f4 = _mm_load_ps(filter_end);
+        { XY_FILTER_4(src4, src_5_8, f4, sum); }
+        //store result
+        _mm_store_ps(dst2, sum);
+    }
+    ASSERT(filter_start == filter);
+    for (;dst2<dst_end0;dst2+=8)
+    {
+        const float *src = dst2;           
+        const float* f=filter_start;
+
+        //filter 4
+        __m128 src4 = _mm_load_ps(src);/*1 2 3 4*/
+        src+=4;
+        __m128 src_5_8;
+        __m128 sum = _mm_setzero_ps();
+        __m128 sum2 = _mm_setzero_ps();
+        __m128 f4 = _mm_load_ps(f);
+        f+=4;
+        src_5_8 = _mm_load_ps(src);/*5 6 7 8*/
+        src+=4;
+        { XY_FILTER_4(src4, src_5_8, f4, sum); }
+        for(;f<filter_end;f+=4,src+=4)
+        {
+            src4 = _mm_load_ps(src);/*1 2 3 4*/
+            __m128 tmp = src_5_8;//important! 
+            { XY_FILTER_4(tmp, src4, f4, sum2); }
+
+            f4 = _mm_load_ps(f);
+            { XY_FILTER_4(src_5_8, src4, f4, sum); }
+            src_5_8 = src4;
+        }
+        src4 = _mm_load_ps(src);/*1 2 3 4*/
+        { XY_FILTER_4(src_5_8, src4, f4, sum2); }
+
+        //store result
+        _mm_store_ps(dst2, sum);
+        _mm_store_ps(dst2+4, sum2);
+    }
+    if (dst2==dst_end0)
+    {
+        const float *src = dst2;
+        //filter 4
+        __m128 src4 = _mm_load_ps(src);//1 2 3 4
+        src+=4;
+        __m128 sum = _mm_setzero_ps();
+        for(const float* f=filter_start;f<filter_end;f+=4,src+=4)
+        {
+            __m128 src_5_8 = _mm_load_ps(src);//5 6 7 8
+            __m128 f4 = _mm_load_ps(f);
+
+            { XY_FILTER_4(src4, src_5_8, f4, sum); }
+            src4 = src_5_8;
+        }
+        //store result
+        _mm_store_ps(dst2, sum);
+        dst2+=4;
+    }
+    for (;dst2<dst_endr;dst2+=4)//right margin
+    {
+        const float *src = dst2;
+        filter_end-=4;
+
+        //filter 4
+        __m128 src4 = _mm_load_ps(src);//1 2 3 4
+        __m128 sum = _mm_setzero_ps();
+        __m128 src_5_8, f4;
+        for(const float* f=filter_start;f<filter_end;f+=4)
+        {
+            src+=4;
+            src_5_8 = _mm_load_ps(src);//5 6 7 8
+            f4 = _mm_load_ps(f);
+
+            { XY_FILTER_4(src4, src_5_8, f4, sum); }
+
+            src4 = src_5_8;
+            //move new 4 in_n_out to old 4 in_n_out
+        }
+        src_5_8 = _mm_setzero_ps();
+        f4 = _mm_load_ps(filter_end);
+        { XY_FILTER_4(src4, src_5_8, f4, sum); }
+        //store result
+        _mm_store_ps(dst2, sum);
+    }
+}
+
+void xy_filter_sse_v6(float *dst, int width, int height, int stride, const float *filter, int filter_width)
+{
+    ASSERT( stride>=4*(width+filter_width) );
+    ASSERT( ((stride|(4*width)|(4*filter_width)|reinterpret_cast<int>(dst)|reinterpret_cast<int>(filter))&15)==0 );
+
+    BYTE* dst_byte = reinterpret_cast<BYTE*>(dst);
+    BYTE* end = dst_byte + height*stride;
+    for( ; dst_byte<end; dst_byte+=stride )
+    {
+        xy_filter_one_line_sse_v6(reinterpret_cast<float*>(dst_byte), width, filter, filter_width);
+    }
+}
 
 class XyFilterTest : public ::testing::Test 
 {
@@ -998,9 +1148,8 @@ TEST_F(XyFilterTest, function ## _ ## width ## _ ## height ## _ ## FILTER_LENGTH
 }
 
 #define DUAL_TEST(width, height, FILTER_LENGTH, loop_num) \
-    FilterTest(width, height, FILTER_LENGTH, loop_num, xy_filter_sse_v2) \
     FilterTest(width, height, FILTER_LENGTH, loop_num, xy_filter_sse_v4) \
-    FilterTest(width, height, FILTER_LENGTH, loop_num, xy_filter_sse_v5)
+    FilterTest(width, height, FILTER_LENGTH, loop_num, xy_filter_sse_v6)
 
 DUAL_TEST(128, 16, 28, 20000)
     DUAL_TEST(256, 16, 28, 20000)
-- 
2.11.4.GIT