[x265] [PATCH] aoCuOrgE2: asm code

praveen at multicorewareinc.com praveen at multicorewareinc.com
Tue Jan 6 12:06:13 CET 2015


# HG changeset patch
# User Praveen Tiwari
# Date 1420541256 -19800
# Node ID 382dc33423b4d18ff7babbe8f97cbba58f77876b
# Parent  feebd0ecda691aeaf9265c7cb20897169df6866a
aoCuOrgE2: asm code

diff -r feebd0ecda69 -r 382dc33423b4 source/common/loopfilter.cpp
--- a/source/common/loopfilter.cpp	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/common/loopfilter.cpp	Tue Jan 06 16:17:36 2015 +0530
@@ -57,6 +57,19 @@
     }
 }
 
+void processSaoCUE2(pixel * rec, int8_t * bufft, int8_t * buff1, int8_t * offsetEo, int width, intptr_t stride)
+{
+    int x = 0;
+    for (x = 0; x < width; x++)
+    {
+        int8_t signDown = signOf(rec[x] - rec[x + stride + 1]);
+        int edgeType = signDown + buff1[x] + 2;
+        bufft[x + 1] = -signDown;
+        short v = rec[x] + offsetEo[edgeType];
+        rec[x] = (pixel)(v < 0 ? 0 : (v > (PIXEL_MAX)) ? (PIXEL_MAX) : v);
+    }
+}
+
 void processSaoCUB0(pixel* rec, const int8_t* offset, int ctuWidth, int ctuHeight, intptr_t stride)
 {
     #define SAO_BO_BITS 5
@@ -81,6 +94,7 @@
 void Setup_C_LoopFilterPrimitives(EncoderPrimitives &p)
 {
     p.saoCuOrgE0 = processSaoCUE0;
+    p.saoCuOrgE2 = processSaoCUE2;
     p.saoCuOrgB0 = processSaoCUB0;
     p.sign = calSign;
 }
diff -r feebd0ecda69 -r 382dc33423b4 source/common/primitives.h
--- a/source/common/primitives.h	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/common/primitives.h	Tue Jan 06 16:17:36 2015 +0530
@@ -191,6 +191,7 @@
 typedef void (*addAvg_t)(const int16_t* src0, const int16_t* src1, pixel* dst, intptr_t src0Stride, intptr_t src1Stride, intptr_t dstStride);
 
 typedef void (*saoCuOrgE0_t)(pixel* rec, int8_t* offsetEo, int width, int8_t signLeft);
+typedef void (*saoCuOrgE2_t)(pixel* rec, int8_t* pBufft, int8_t* pBuff1, int8_t* offsetEo, int lcuWidth, intptr_t stride);
 typedef void (*saoCuOrgB0_t)(pixel* rec, const int8_t* offsetBo, int ctuWidth, int ctuHeight, intptr_t stride);
 typedef void (*sign_t)(int8_t *dst, const pixel *src1, const pixel *src2, const int endX);
 typedef void (*planecopy_cp_t) (const uint8_t* src, intptr_t srcStride, pixel* dst, intptr_t dstStride, int width, int height, int shift);
@@ -244,6 +245,7 @@
 
     sign_t                sign;
     saoCuOrgE0_t          saoCuOrgE0;
+    saoCuOrgE2_t          saoCuOrgE2;
     saoCuOrgB0_t          saoCuOrgB0;
 
     downscale_t           frameInitLowres;
diff -r feebd0ecda69 -r 382dc33423b4 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Tue Jan 06 16:17:36 2015 +0530
@@ -1650,6 +1650,7 @@
     {
         p.sign = x265_calSign_sse4;
         p.saoCuOrgE0 = x265_saoCuOrgE0_sse4;
+        p.saoCuOrgE2 = x265_saoCuOrgE2_sse4;
         p.saoCuOrgB0 = x265_saoCuOrgB0_sse4;
 
         LUMA_ADDAVG(_sse4);
diff -r feebd0ecda69 -r 382dc33423b4 source/common/x86/loopfilter.asm
--- a/source/common/x86/loopfilter.asm	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/common/x86/loopfilter.asm	Tue Jan 06 16:17:36 2015 +0530
@@ -35,6 +35,7 @@
 cextern pb_1
 cextern pb_128
 cextern pb_2
+cextern pw_2
 
 
 ;============================================================================================================
@@ -85,6 +86,58 @@
     jnz        .loop
     RET
 
+;======================================================================================================================================================
+; void saoCuOrgE2(pixel * rec, int8_t * bufft, int8_t * buff1, int8_t * offsetEo, int lcuWidth, intptr_t stride)
+;======================================================================================================================================================
+INIT_XMM sse4
+cglobal saoCuOrgE2, 5, 7, 8, rec, bufft, buff1, offsetEo, lcuWidth
+
+    mov         r6,    16
+    mov         r5d,   r5m
+    pxor        m0,    m0                      ; m0 = 0
+    mova        m6,    [pb_2]                  ; m6 = [2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2]
+    mova        m7,    [pb_128]
+    shr         r4d,   4
+    inc         r1q
+
+    .loop
+         movu        m1,    [r0]                    ; m1 = rec[x]
+         movu        m2,    [r0 + r5 + 1]           ; m2 = rec[x + stride + 1]
+         pxor        m3,    m1,    m7
+         pxor        m4,    m2,    m7
+         pcmpgtb     m2,    m3,    m4
+         pcmpgtb     m4,    m3
+         pand        m2,    [pb_1]
+         por         m2,    m4
+         movu        m3,    [r2]                    ; m3 = buff1
+
+         paddb       m3,    m2
+         paddb       m3,    m6                      ; m3 = edgeType
+
+         movu        m4,    [r3]                    ; m4 = offsetEo
+         pshufb      m4,    m3
+
+         psubb       m3,    m0,    m2
+         movu        [r1],  m3
+
+         pmovzxbw    m2,    m1
+         punpckhbw   m1,    m0
+         pmovsxbw    m3,    m4
+         punpckhbw   m4,    m4
+         psraw       m4,    8
+
+         paddw       m2,    m3
+         paddw       m1,    m4
+         packuswb    m2,    m1
+         movu        [r0],  m2
+
+         add         r0,    r6
+         add         r1,    r6
+         add         r2,    r6
+         dec         r4d
+         jnz         .loop
+    RET
+
 ;=====================================================================================
 ; void saoCuOrgB0(pixel* rec, const pixel* offset, int lcuWidth, int lcuHeight, int stride)
 ;=====================================================================================
diff -r feebd0ecda69 -r 382dc33423b4 source/common/x86/loopfilter.h
--- a/source/common/x86/loopfilter.h	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/common/x86/loopfilter.h	Tue Jan 06 16:17:36 2015 +0530
@@ -26,6 +26,7 @@
 #define X265_LOOPFILTER_H
 
 void x265_saoCuOrgE0_sse4(pixel * rec, int8_t * offsetEo, int endX, int8_t signLeft);
+void x265_saoCuOrgE2_sse4(pixel* rec, int8_t* pBufft, int8_t* pBuff1, int8_t* offsetEo, int lcuWidth, intptr_t stride);
 void x265_saoCuOrgB0_sse4(pixel* rec, const int8_t* offsetBo, int ctuWidth, int ctuHeight, intptr_t stride);
 void x265_calSign_sse4(int8_t *dst, const pixel *src1, const pixel *src2, const int endX);
 
diff -r feebd0ecda69 -r 382dc33423b4 source/encoder/sao.cpp
--- a/source/encoder/sao.cpp	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/encoder/sao.cpp	Tue Jan 06 16:17:36 2015 +0530
@@ -385,23 +385,54 @@
                 upBuff1[x] = signOf(rec[x] - tmpU[x - 1]);
         }
 
-        for (y = startY; y < endY; y++)
+        if (ctuWidth & 15)
         {
-            upBufft[startX] = signOf(rec[stride + startX] - tmpL[y]);
-            for (x = startX; x < endX; x++)
-            {
-                int8_t signDown = signOf(rec[x] - rec[x + stride + 1]);
-                int edgeType = signDown + upBuff1[x] + 2;
-                upBufft[x + 1] = -signDown;
-                rec[x] = m_clipTable[rec[x] + m_offsetEo[edgeType]];
-            }
+             for (y = startY; y < endY; y++)
+             {
+                 upBufft[startX] = signOf(rec[stride + startX] - tmpL[y]);
+                 for (x = startX; x < endX; x++)
+                 {
+                     int8_t signDown = signOf(rec[x] - rec[x + stride + 1]);
+                     int edgeType = signDown + upBuff1[x] + 2;
+                     upBufft[x + 1] = -signDown;
+                     rec[x] = m_clipTable[rec[x] + m_offsetEo[edgeType]];
+                 }
 
-            std::swap(upBuff1, upBufft);
+                 std::swap(upBuff1, upBufft);
 
-            rec += stride;
+                 rec += stride;
+             }
         }
+         else
+         {
+             for (y = startY; y < endY; y++)
+             {
+                 int8_t iSignDown2 = signOf(rec[stride + startX] - tmpL[y]);
+                 pixel firstPxl = rec[0];  // copy first Pxl
+                 pixel lastPxl = rec[ctuWidth - 1];
+                 int8_t one = upBufft[1];
+                 int8_t two = upBufft[endX + 1];
 
-        break;
+                 primitives.saoCuOrgE2(rec, upBufft, upBuff1, m_offsetEo, ctuWidth, stride);
+                 if (!lpelx)
+                 {
+                     rec[0] = firstPxl;
+                     upBufft[1] = one;
+                 }
+
+                 if (rpelx == picWidth)
+                 {
+                     rec[ctuWidth - 1] = lastPxl;
+                     upBufft[endX + 1] = two;
+                 }
+
+                 upBufft[startX] = iSignDown2;
+
+                 std::swap(upBuff1, upBufft);
+                 rec += stride;
+             }
+         }
+         break;
     }
     case SAO_EO_3: // dir: 45
     {
diff -r feebd0ecda69 -r 382dc33423b4 source/test/pixelharness.cpp
--- a/source/test/pixelharness.cpp	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/test/pixelharness.cpp	Tue Jan 06 16:17:36 2015 +0530
@@ -65,7 +65,9 @@
         sbuf1[i] = (rand() % (2 * SMAX + 1)) - SMAX - 1; //max(SHORT_MIN, min(rand(), SMAX));
         sbuf2[i] = (rand() % (2 * SMAX + 1)) - SMAX - 1; //max(SHORT_MIN, min(rand(), SMAX));
         ibuf1[i] = (rand() % (2 * SMAX + 1)) - SMAX - 1;
-        psbuf1[i] = (rand() % 65) - 32;                   // range is between -32 to 32
+        psbuf1[i] = psbuf4[i] = (rand() % 65) - 32;                   // range is between -32 to 32
+        psbuf2[i] = (rand() % 3) - 1;
+        psbuf3[i] = (rand() % 129) - 128;
         sbuf3[i] = rand() % PIXEL_MAX; // for blockcopy only
     }
 }
@@ -917,6 +919,37 @@
     return true;
 }
 
+bool PixelHarness::check_saoCuOrgE2_t(saoCuOrgE2_t ref, saoCuOrgE2_t opt)
+{
+    ALIGN_VAR_16(pixel, ref_dest[64 * 64]);
+    ALIGN_VAR_16(pixel, opt_dest[64 * 64]);
+
+    memset(ref_dest, 0xCD, sizeof(ref_dest));
+    memset(opt_dest, 0xCD, sizeof(opt_dest));
+
+    int j = 0;
+
+    for (int i = 0; i < ITERS; i++)
+    {
+        int width = 16 * (rand() % 4 + 1);
+        int stride = width + 1;
+
+        ref(ref_dest, psbuf1 + j, psbuf2 + j, psbuf3 + j, width, stride);
+        checked(opt, opt_dest, psbuf4 + j, psbuf2 + j, psbuf3 + j, width, stride);
+
+        if (memcmp(psbuf1 + j, psbuf4 + j, width * sizeof(int8_t)))
+            return false;
+
+        if (memcmp(ref_dest, opt_dest, 64 * 64 * sizeof(pixel)))
+            return false;
+
+        reportfail();
+        j += INCR;
+    }
+
+    return true;
+}
+
 bool PixelHarness::check_planecopy_sp(planecopy_sp_t ref, planecopy_sp_t opt)
 {
     ALIGN_VAR_16(pixel, ref_dest[64 * 64]);
@@ -1463,6 +1496,15 @@
         }
     }
 
+    if (opt.saoCuOrgE2)
+    {
+        if (!check_saoCuOrgE2_t(ref.saoCuOrgE2, opt.saoCuOrgE2))
+        {
+            printf("SAO_EO_2 failed\n");
+            return false;
+        }
+    }
+
     if (opt.saoCuOrgB0)
     {
         if (!check_saoCuOrgB0_t(ref.saoCuOrgB0, opt.saoCuOrgB0))
@@ -1801,6 +1843,12 @@
         REPORT_SPEEDUP(opt.saoCuOrgE0, ref.saoCuOrgE0, pbuf1, psbuf1, 64, 1);
     }
 
+    if (opt.saoCuOrgE2)
+    {
+        HEADER0("SAO_EO_2");
+        REPORT_SPEEDUP(opt.saoCuOrgE2, ref.saoCuOrgE2, pbuf1, psbuf1, psbuf2, psbuf3, 64, 64);
+    }
+
     if (opt.saoCuOrgB0)
     {
         HEADER0("SAO_BO_0");
diff -r feebd0ecda69 -r 382dc33423b4 source/test/pixelharness.h
--- a/source/test/pixelharness.h	Mon Jan 05 18:57:20 2015 +0530
+++ b/source/test/pixelharness.h	Tue Jan 06 16:17:36 2015 +0530
@@ -47,6 +47,9 @@
     pixel    pbuf4[BUFFSIZE];
     int      ibuf1[BUFFSIZE];
     int8_t   psbuf1[BUFFSIZE];
+    int8_t   psbuf2[BUFFSIZE];
+    int8_t   psbuf3[BUFFSIZE];
+    int8_t   psbuf4[BUFFSIZE];
 
     int16_t  sbuf1[BUFFSIZE];
     int16_t  sbuf2[BUFFSIZE];
@@ -90,6 +93,7 @@
     bool check_ssim_end(ssim_end4_t ref, ssim_end4_t opt);
     bool check_addAvg(addAvg_t, addAvg_t);
     bool check_saoCuOrgE0_t(saoCuOrgE0_t ref, saoCuOrgE0_t opt);
+    bool check_saoCuOrgE2_t(saoCuOrgE2_t ref, saoCuOrgE2_t opt);
     bool check_saoCuOrgB0_t(saoCuOrgB0_t ref, saoCuOrgB0_t opt);
     bool check_planecopy_sp(planecopy_sp_t ref, planecopy_sp_t opt);
     bool check_planecopy_cp(planecopy_cp_t ref, planecopy_cp_t opt);


More information about the x265-devel mailing list