[x265] [PATCH] add testbench for psyCost_ss and asm for psyCost_ss_4x4: improve 1989c->515c

Divya Manivannan divya at multicorewareinc.com
Fri Jan 9 08:58:04 CET 2015


# HG changeset patch
# User Divya Manivannan <divya at multicorewareinc.com>
# Date 1420790181 -19800
#      Fri Jan 09 13:26:21 2015 +0530
# Node ID 0f4b677cea64254d0b8f77ccc84c785bf832698d
# Parent  c99e1a309bd1690be9a0a407050d97d95ccab05a
add testbench for psyCost_ss and asm for psyCost_ss_4x4: improve 1989c->515c

diff -r c99e1a309bd1 -r 0f4b677cea64 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Fri Jan 09 13:09:39 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp	Fri Jan 09 13:26:21 2015 +0530
@@ -1430,6 +1430,7 @@
         p.psy_cost_pp[BLOCK_32x32] = x265_psyCost_pp_32x32_sse4;
         p.psy_cost_pp[BLOCK_64x64] = x265_psyCost_pp_64x64_sse4;
 #endif
+        p.psy_cost_ss[BLOCK_4x4] = x265_psyCost_ss_4x4_sse4;
     }
     if (cpuMask & X265_CPU_XOP)
     {
@@ -1716,6 +1717,7 @@
         p.psy_cost_pp[BLOCK_32x32] = x265_psyCost_pp_32x32_sse4;
         p.psy_cost_pp[BLOCK_64x64] = x265_psyCost_pp_64x64_sse4;
 #endif
+        p.psy_cost_ss[BLOCK_4x4] = x265_psyCost_ss_4x4_sse4;
     }
     if (cpuMask & X265_CPU_AVX)
     {
diff -r c99e1a309bd1 -r 0f4b677cea64 source/common/x86/pixel-a.asm
--- a/source/common/x86/pixel-a.asm	Fri Jan 09 13:09:39 2015 +0530
+++ b/source/common/x86/pixel-a.asm	Fri Jan 09 13:26:21 2015 +0530
@@ -7569,3 +7569,157 @@
     RET
 %endif ; HIGH_BIT_DEPTH
 %endif
+
+;---------------------------------------------------------------------------------------------------------------------
+;int psyCost_ss(const int16_t* source, intptr_t sstride, const int16_t* recon, intptr_t rstride)
+;---------------------------------------------------------------------------------------------------------------------
+INIT_XMM sse4
+cglobal psyCost_ss_4x4, 4, 5, 8
+
+    add             r1, r1
+    lea             r4, [3 * r1]
+    movddup         m0, [r0]
+    movddup         m1, [r0 + r1]
+    movddup         m2, [r0 + r1 * 2]
+    movddup         m3, [r0 + r4]
+
+    pabsw           m4, m0
+    pabsw           m5, m1
+    paddw           m5, m4
+    pabsw           m4, m2
+    paddw           m5, m4
+    pabsw           m4, m3
+    paddw           m5, m4
+    pmaddwd         m5, [pw_1]
+    psrldq          m4, m5, 4
+    paddd           m5, m4
+    psrld           m6, m5, 2
+
+    mova            m4, [hmul_8w]
+    pmaddwd         m0, m4
+    pmaddwd         m1, m4
+    pmaddwd         m2, m4
+    pmaddwd         m3, m4
+
+    psrldq          m4, m0, 4
+    psubd           m5, m0, m4
+    paddd           m0, m4
+    shufps          m0, m5, 10001000b
+
+    psrldq          m4, m1, 4
+    psubd           m5, m1, m4
+    paddd           m1, m4
+    shufps          m1, m5, 10001000b
+
+    psrldq          m4, m2, 4
+    psubd           m5, m2, m4
+    paddd           m2, m4
+    shufps          m2, m5, 10001000b
+
+    psrldq          m4, m3, 4
+    psubd           m5, m3, m4
+    paddd           m3, m4
+    shufps          m3, m5, 10001000b
+
+    mova            m4, m0
+    paddd           m0, m1
+    psubd           m1, m4
+    mova            m4, m2
+    paddd           m2, m3
+    psubd           m3, m4
+    mova            m4, m0
+    paddd           m0, m2
+    psubd           m2, m4
+    mova            m4, m1
+    paddd           m1, m3
+    psubd           m3, m4
+
+    pabsd           m0, m0
+    pabsd           m2, m2
+    pabsd           m1, m1
+    pabsd           m3, m3
+    paddd           m0, m2
+    paddd           m1, m3
+    paddd           m0, m1
+    movhlps         m1, m0
+    paddd           m0, m1
+    psrldq          m1, m0, 4
+    paddd           m0, m1
+    psrld           m0, 1
+    psubd           m7, m0, m6
+
+    add             r3, r3
+    lea             r4, [3 * r3]
+    movddup         m0, [r2]
+    movddup         m1, [r2 + r3]
+    movddup         m2, [r2 + r3 * 2]
+    movddup         m3, [r2 + r4]
+
+    pabsw           m4, m0
+    pabsw           m5, m1
+    paddw           m5, m4
+    pabsw           m4, m2
+    paddw           m5, m4
+    pabsw           m4, m3
+    paddw           m5, m4
+    pmaddwd         m5, [pw_1]
+    psrldq          m4, m5, 4
+    paddd           m5, m4
+    psrld           m6, m5, 2
+
+    mova            m4, [hmul_8w]
+    pmaddwd         m0, m4
+    pmaddwd         m1, m4
+    pmaddwd         m2, m4
+    pmaddwd         m3, m4
+
+    psrldq          m4, m0, 4
+    psubd           m5, m0, m4
+    paddd           m0, m4
+    shufps          m0, m5, 10001000b
+
+    psrldq          m4, m1, 4
+    psubd           m5, m1, m4
+    paddd           m1, m4
+    shufps          m1, m5, 10001000b
+
+    psrldq          m4, m2, 4
+    psubd           m5, m2, m4
+    paddd           m2, m4
+    shufps          m2, m5, 10001000b
+
+    psrldq          m4, m3, 4
+    psubd           m5, m3, m4
+    paddd           m3, m4
+    shufps          m3, m5, 10001000b
+
+    mova            m4, m0
+    paddd           m0, m1
+    psubd           m1, m4
+    mova            m4, m2
+    paddd           m2, m3
+    psubd           m3, m4
+    mova            m4, m0
+    paddd           m0, m2
+    psubd           m2, m4
+    mova            m4, m1
+    paddd           m1, m3
+    psubd           m3, m4
+
+    pabsd           m0, m0
+    pabsd           m2, m2
+    pabsd           m1, m1
+    pabsd           m3, m3
+    paddd           m0, m2
+    paddd           m1, m3
+    paddd           m0, m1
+    movhlps         m1, m0
+    paddd           m0, m1
+    psrldq          m1, m0, 4
+    paddd           m0, m1
+    psrld           m0, 1
+    psubd           m0, m6
+    psubd           m7, m0
+    pabsd           m0, m7
+    movd            eax, m0
+    RET
diff -r c99e1a309bd1 -r 0f4b677cea64 source/common/x86/pixel.h
--- a/source/common/x86/pixel.h	Fri Jan 09 13:09:39 2015 +0530
+++ b/source/common/x86/pixel.h	Fri Jan 09 13:26:21 2015 +0530
@@ -223,6 +223,7 @@
 int x265_psyCost_pp_16x16_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
 int x265_psyCost_pp_32x32_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
 int x265_psyCost_pp_64x64_sse4(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride);
+int x265_psyCost_ss_4x4_sse4(const int16_t* source, intptr_t sstride, const int16_t* recon, intptr_t rstride);
 
 #undef DECL_PIXELS
 #undef DECL_HEVC_SSD
diff -r c99e1a309bd1 -r 0f4b677cea64 source/test/pixelharness.cpp
--- a/source/test/pixelharness.cpp	Fri Jan 09 13:09:39 2015 +0530
+++ b/source/test/pixelharness.cpp	Fri Jan 09 13:26:21 2015 +0530
@@ -1089,6 +1089,28 @@
     return true;
 }
 
+bool PixelHarness::check_psyCost_ss(pixelcmp_ss_t ref, pixelcmp_ss_t opt)
+{
+    int j = 0, index1, index2, optres, refres;
+    intptr_t stride = STRIDE;
+
+    for (int i = 0; i < ITERS; i++)
+    {
+        index1 = rand() % TEST_CASES;
+        index2 = rand() % TEST_CASES;
+        optres = (int)checked(opt, short_test_buff[index1], stride, short_test_buff[index2] + j, stride);
+        refres = ref(short_test_buff[index1], stride, short_test_buff[index2] + j, stride);
+
+        if (optres != refres)
+            return false;
+
+        reportfail();
+        j += INCR;
+    }
+
+    return true;
+}
+
 bool PixelHarness::check_saoCuOrgB0_t(saoCuOrgB0_t ref, saoCuOrgB0_t opt)
 {
     ALIGN_VAR_16(pixel, ref_dest[64 * 64]);
@@ -1470,6 +1492,15 @@
                 return false;
             }
         }
+
+        if (opt.psy_cost_ss[i])
+        {
+            if (!check_psyCost_ss(ref.psy_cost_ss[i], opt.psy_cost_ss[i]))
+            {
+                printf("\npsy_cost_ss[%dx%d] failed!\n", 4 << i, 4 << i);
+                return false;
+            }
+        }
     }
 
     if (opt.weight_pp)
@@ -1862,6 +1893,12 @@
             HEADER("psy_cost_pp[%dx%d]", 4 << i, 4 << i);
             REPORT_SPEEDUP(opt.psy_cost_pp[i], ref.psy_cost_pp[i], pbuf1, STRIDE, pbuf2, STRIDE);
         }
+
+        if (opt.psy_cost_ss[i])
+        {
+            HEADER("psy_cost_ss[%dx%d]", 4 << i, 4 << i);
+            REPORT_SPEEDUP(opt.psy_cost_ss[i], ref.psy_cost_ss[i], sbuf1, STRIDE, sbuf2, STRIDE);
+        }
     }
 
     if (opt.weight_pp)
diff -r c99e1a309bd1 -r 0f4b677cea64 source/test/pixelharness.h
--- a/source/test/pixelharness.h	Fri Jan 09 13:09:39 2015 +0530
+++ b/source/test/pixelharness.h	Fri Jan 09 13:26:21 2015 +0530
@@ -101,6 +101,7 @@
     bool check_planecopy_sp(planecopy_sp_t ref, planecopy_sp_t opt);
     bool check_planecopy_cp(planecopy_cp_t ref, planecopy_cp_t opt);
     bool check_psyCost_pp(pixelcmp_t ref, pixelcmp_t opt);
+    bool check_psyCost_ss(pixelcmp_ss_t ref, pixelcmp_ss_t opt);
     bool check_calSign(sign_t ref, sign_t opt);
 
 public:


More information about the x265-devel mailing list