[x265] [PATCH 2 of 3] asm: ipfilter_ss[FILTER_V_S_S_8]

Min Chen chenm003 at 163.com
Mon Nov 4 12:05:32 CET 2013


# HG changeset patch
# User Min Chen <chenm003 at 163.com>
# Date 1383563104 -28800
# Node ID 539ad4851359f96591f612f8b7b6fb0483e5a48c
# Parent  2a7a5766fbd84436cdbbd5018a34db92e62553a1
asm: ipfilter_ss[FILTER_V_S_S_8]

diff -r 2a7a5766fbd8 -r 539ad4851359 source/Lib/TLibCommon/TComPrediction.cpp
--- a/source/Lib/TLibCommon/TComPrediction.cpp	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/Lib/TLibCommon/TComPrediction.cpp	Mon Nov 04 19:05:04 2013 +0800
@@ -537,7 +537,7 @@
         int filterSize = NTAPS_LUMA;
         int halfFilterSize = (filterSize >> 1);
         primitives.ipfilter_ps[FILTER_H_P_S_8](ref - (halfFilterSize - 1) * refStride, refStride, m_immedVals, tmpStride, width, height + filterSize - 1, g_lumaFilter[xFrac]);
-        primitives.ipfilter_ss[FILTER_V_S_S_8](m_immedVals + (halfFilterSize - 1) * tmpStride, tmpStride, dst, dstStride, width, height, g_lumaFilter[yFrac]);
+        primitives.ipfilter_ss[FILTER_V_S_S_8](m_immedVals + (halfFilterSize - 1) * tmpStride, tmpStride, dst, dstStride, width, height, yFrac);
     }
 }
 
@@ -643,9 +643,9 @@
         int filterSize = NTAPS_CHROMA;
         int halfFilterSize = (filterSize >> 1);
         primitives.ipfilter_ps[FILTER_H_P_S_4](refCb - (halfFilterSize - 1) * refStride, refStride, m_immedVals, extStride, cxWidth, cxHeight + filterSize - 1, g_chromaFilter[xFrac]);
-        primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCb, dstStride, cxWidth, cxHeight, g_chromaFilter[yFrac]);
+        primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCb, dstStride, cxWidth, cxHeight, yFrac);
         primitives.ipfilter_ps[FILTER_H_P_S_4](refCr - (halfFilterSize - 1) * refStride, refStride, m_immedVals, extStride, cxWidth, cxHeight + filterSize - 1, g_chromaFilter[xFrac]);
-        primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCr, dstStride, cxWidth, cxHeight, g_chromaFilter[yFrac]);
+        primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCr, dstStride, cxWidth, cxHeight, yFrac);
     }
 }
 
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/ipfilter.cpp
--- a/source/common/ipfilter.cpp	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/common/ipfilter.cpp	Mon Nov 04 19:05:04 2013 +0800
@@ -120,8 +120,9 @@
 }
 
 template<int N>
-void filterVertical_ss_c(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, int16_t const *c)
+void filterVertical_ss_c(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coefIdx)
 {
+    const int16_t *const c = (N == 8 ? g_lumaFilter[coefIdx] : g_chromaFilter[coefIdx]);
     int shift = IF_FILTER_PREC;
     int row, col;
     src -= (N / 2 - 1) * srcStride;
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/primitives.h
--- a/source/common/primitives.h	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/common/primitives.h	Mon Nov 04 19:05:04 2013 +0800
@@ -166,7 +166,7 @@
 typedef void (*ipfilter_pp_t)(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, const int16_t *coeff);
 typedef void (*ipfilter_ps_t)(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int16_t *coeff);
 typedef void (*ipfilter_sp_t)(int16_t *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, const int coeffIdx);
-typedef void (*ipfilter_ss_t)(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int16_t *coeff);
+typedef void (*ipfilter_ss_t)(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coeffIdx);
 typedef void (*ipfilter_p2s_t)(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height);
 typedef void (*ipfilter_s2p_t)(int16_t *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height);
 typedef void (*blockcpy_pp_t)(int bx, int by, pixel *dst, intptr_t dstride, pixel *src, intptr_t sstride); // dst is aligned
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/common/x86/asm-primitives.cpp	Mon Nov 04 19:05:04 2013 +0800
@@ -279,6 +279,7 @@
         SA8D_INTER_FROM_BLOCK(sse2);
 
         p.cvt32to16_shr = x265_cvt32to16_shr_sse2;
+        p.ipfilter_ss[FILTER_V_S_S_8] = x265_interp_8tap_v_ss_sse2;
     }
     if (cpuMask & X265_CPU_SSSE3)
     {
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/common/x86/ipfilter8.asm	Mon Nov 04 19:05:04 2013 +0800
@@ -2593,3 +2593,144 @@
     jnz         .loopH
 
     RET
+
+
+;-------------------------------------------------------------------------------------------------------------
+; void interp_8tap_v_ss(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coefIdx)
+;-------------------------------------------------------------------------------------------------------------
+INIT_XMM sse2
+
+%if ARCH_X86_64
+cglobal interp_8tap_v_ss, 4, 7+1, 8
+%define tmp_r4d     r7d
+%define tmp_r5d     r8d
+%else
+cglobal interp_8tap_v_ss, 4, 7, 8, 0-2*4
+%define tmp_r4d     dword [rsp + 0*4]
+%define tmp_r5d     dword [rsp + 1*4]
+%endif
+
+    ; load width, height and filterIdx
+    mov         r4d, r4m
+    mov         r5d, r5m
+    mov         r6d, r6m
+
+    ; convert to word stride
+    add         r1, r1
+    add         r3, r3
+
+    ; stort to temporary memory or register
+    shr         r4d, 2
+    mov         tmp_r4d, r4d
+    shr         r5d, 2
+    mov         tmp_r5d, r5d
+
+    shl         r6d, 6
+%ifdef PIC
+    lea         r5, [tab_LumaCoeffV]
+    lea         r6, [r5 + r6]
+%else
+    lea         r6, [tab_LumaCoeffV + r6]
+%endif
+
+    lea         r4, [r1 * 3]
+    sub         r0, r4
+
+.loopH:
+    ; load width
+    mov         r4d, tmp_r4d
+
+.loopW:
+
+    movh        m0, [r0]                    ; m0 = [0]
+    movh        m1, [r0 + r1]               ; m1 = [1]
+    lea         r0, [r0 + r1 * 2]
+    punpcklwd   m0, m1
+    pmaddwd     m0, [r6 + 0 * 16]           ; m0 = [0+1]            = R0
+
+    movh        m2, [r0]                    ; m2 = [2]
+    movh        m3, [r0 + r1]               ; m3 = [3]
+    lea         r0, [r0 + r1 * 2]
+    punpcklwd   m1, m2
+    pmaddwd     m1, [r6 + 0 * 16]           ; m1 = [1+2]            = R1
+    punpcklwd   m2, m3                      ; m2 = [2 3]
+    pmaddwd     m7, m2, [r6 + 1 * 16]       ;
+    paddd       m0, m7                      ; m0 = [0+1+2+3]        = R0
+    pmaddwd     m2, [r6 + 0 * 16]           ; m2 = [2+3]            = R2
+
+    movh        m4, [r0]                    ; m4 = [4]
+    movh        m5, [r0 + r1]               ; m5 = [5]
+    lea         r0, [r0 + r1 * 2]
+    punpcklwd   m3, m4                      ; m3 = [3 4]
+    pmaddwd     m7, m3, [r6 + 1 * 16]
+    paddd       m1, m7                      ; m1 = [1+2+3+4]        = R1
+    pmaddwd     m3, [r6 + 0 * 16]           ; m3 = [3+4]            = R3
+    punpcklwd   m4, m5                      ; m4 = [4 5]
+    pmaddwd     m7, m4, [r6 + 2 * 16]
+    paddd       m0, m7                      ; m0 = [0+1+2+3+4+5]    = R0
+    pmaddwd     m4, [r6 + 1 * 16]
+    paddd       m2, m4                      ; m2 = [2+3+4+5]        = R2
+
+    movh        m6, [r0]                    ; m6 = [6]
+    movh        m7, [r0 + r1]               ; m7 = [7]
+    lea         r0, [r0 + r1 * 2]
+    punpcklwd   m5, m6                      ; m5 = [5 6]
+    pmaddwd     m4, m5, [r6 + 2 * 16]
+    paddd       m1, m4                      ; m1 = [1+2+3+4+5+6]    = R1
+    pmaddwd     m5, [r6 + 1 * 16]
+    paddd       m3, m5                      ; m3 = [3+4+5+6]        = R3
+    punpcklwd   m6, m7                      ; m6 = [6 7]
+    pmaddwd     m4, m6, [r6 + 3 * 16]
+    paddd       m0, m4                      ; m0 = [0+1+2+3+4+5+6+7]= R0
+    pmaddwd     m6, [r6 + 2 * 16]
+    paddd       m2, m6                      ; m2 = [2+3+4+5+6+7]    = R2
+    psrad       m0, 6
+    packssdw    m0, m0
+    movh        [r2], m0                    ; store [0]
+
+    movh        m4, [r0]                    ; m4 = [8]
+    movh        m5, [r0 + r1]               ; m5 = [9]
+    punpcklwd   m7, m4                      ; m7 = [7 8]
+    pmaddwd     m6, m7, [r6 + 3 * 16]
+    paddd       m1, m6                      ; m1 = [1+2+3+4+5+6+7+8]= R1
+    pmaddwd     m7, [r6 + 2 * 16]
+    paddd       m3, m7                      ; m3 = [3+4+5+6+7+8]    = R3
+    psrad       m1, 6
+    packssdw    m1, m1
+    movh        [r2 + r3], m1               ; store [1]
+    punpcklwd   m4, m5                      ; m4 = [8 9]
+    pmaddwd     m4, [r6 + 3 * 16]
+    paddd       m2, m4                      ; m2 = [2+3+4+5+6+7+8+9]= R2
+    psrad       m2, 6
+    packssdw    m2, m2
+    movh        [r2 + r3 * 2], m2           ; store [2]
+    lea         r2, [r2 + r3 * 2]
+
+    movh        m4, [r0 + r1 * 2]           ; m4 = [10]
+    punpcklwd   m5, m4                      ; m5 = [9 10]
+    pmaddwd     m5, [r6 + 3 * 16]
+    paddd       m3, m5                      ; m3 = [3+4+5+6+7+8+9+10]=R3
+    psrad       m3, 6
+    packssdw    m3, m3
+    movh        [r2 + r3], m3               ; store [3]
+
+    lea         r5, [r1 * 8 - 8]
+    sub         r0, r5
+    lea         r5, [r3 * 2 - 8]
+    sub         r2, r5
+
+    dec         r4d
+    jnz         .loopW
+
+    ; move to next row
+    mov         r4d, tmp_r4d
+    shl         r4d, 3
+    lea         r0, [r0 + r1 * 4]
+    sub         r0, r4
+    lea         r2, [r2 + r3 * 4]
+    sub         r2, r4
+
+    dec         tmp_r5d
+    jnz         .loopH
+
+    RET
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/pixel.h
--- a/source/common/x86/pixel.h	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/common/x86/pixel.h	Mon Nov 04 19:05:04 2013 +0800
@@ -214,6 +214,7 @@
 uint64_t x265_pixel_sa8d_satd_16x16_avx2(pixel *pix1, intptr_t stride1, pixel *pix2, intptr_t stride2);
 
 void x265_cvt32to16_shr_sse2(int16_t *dst, int *src, intptr_t, int, int);
+void x265_interp_8tap_v_ss_sse2(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coefIdx);
 
 #define DECL_HEVC_SSD(suffix) \
     int x265_pixel_ssd_32x64_ ## suffix(pixel *, intptr_t, pixel *, intptr_t); \
diff -r 2a7a5766fbd8 -r 539ad4851359 source/encoder/motion.cpp
--- a/source/encoder/motion.cpp	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/encoder/motion.cpp	Mon Nov 04 19:05:04 2013 +0800
@@ -1213,7 +1213,7 @@
             int filterSize = NTAPS_LUMA;
             int halfFilterSize = (filterSize >> 1);
             primitives.ipfilter_ps[FILTER_H_P_S_8](fref - (halfFilterSize - 1) * ref->lumaStride, ref->lumaStride, immedVal, blockwidth, blockwidth, realHeight + filterSize - 1, g_lumaFilter[xFrac]);
-            primitives.ipfilter_ss[FILTER_V_S_S_8](immedVal + (halfFilterSize - 1) * blockwidth, blockwidth, immedVal2, FENC_STRIDE, blockwidth, realHeight, g_lumaFilter[yFrac]);
+            primitives.ipfilter_ss[FILTER_V_S_S_8](immedVal + (halfFilterSize - 1) * blockwidth, blockwidth, immedVal2, FENC_STRIDE, blockwidth, realHeight, yFrac);
             primitives.weightpUni(immedVal2, subpelbuf, FENC_STRIDE, FENC_STRIDE, blockwidth, realHeight, ref->weight, local_round, local_shift, ref->offset);
         }
     }
diff -r 2a7a5766fbd8 -r 539ad4851359 source/test/ipfilterharness.cpp
--- a/source/test/ipfilterharness.cpp	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/test/ipfilterharness.cpp	Mon Nov 04 19:05:04 2013 +0800
@@ -318,6 +318,68 @@
     return true;
 }
 
+bool IPFilterHarness::check_IPFilter_primitive(ipfilter_ss_t ref, ipfilter_ss_t opt, int isChroma)
+{
+    int rand_val, rand_srcStride, rand_dstStride;
+    const int min_size = isChroma ? 2 : 4;
+
+    // NOTE: refill data to avoid overflow
+    const int max_filter_val = 64 * (1 << 8);
+    for (int i = 0; i < ipf_t_size; i++)
+    {
+        short_buff[i] = rand() % (2 * max_filter_val) - max_filter_val;
+    }
+
+    for (int i = 0; i <= 1000; i++)
+    {
+        int rand_height = rand() % 100;                 // Randomly generated Height
+        int rand_width = rand() % 100;                  // Randomly generated Width
+
+        memset(IPF_vec_output_s, 0xCD, ipf_t_size);      // Initialize output buffer to zero
+        memset(IPF_C_output_s, 0xCD, ipf_t_size);        // Initialize output buffer to zero
+
+        rand_val = rand() % 4;                      // Random offset in the filter
+        rand_srcStride = rand() % 100;              // Randomly generated srcStride
+        rand_dstStride = rand() % 100;              // Randomly generated dstStride
+
+        rand_width &= ~(min_size - 1);
+        if (rand_width < min_size)
+            rand_width = min_size;
+
+        rand_height &= ~(min_size - 1);
+        if (rand_height < min_size)
+            rand_height = min_size;
+
+        if (rand_srcStride < rand_width)
+            rand_srcStride = rand_width;
+
+        if (rand_dstStride < rand_width)
+            rand_dstStride = rand_width;
+
+        ref(short_buff + 3 * rand_srcStride,
+            rand_srcStride,
+            IPF_C_output_s,
+            rand_dstStride,
+            rand_width,
+            rand_height, rand_val
+            );
+        opt(short_buff + 3 * rand_srcStride,
+            rand_srcStride,
+            IPF_vec_output_s,
+            rand_dstStride,
+            rand_width,
+            rand_height, rand_val
+            );
+
+        if (memcmp(IPF_C_output_s, IPF_vec_output_s, ipf_t_size * sizeof(int16_t)))
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 bool IPFilterHarness::check_IPFilterChroma_primitive(filter_pp_t ref, filter_pp_t opt)
 {
     int rand_srcStride, rand_dstStride, rand_coeffIdx;
@@ -452,6 +514,18 @@
         }
     }
 
+    for (int value = 0; value < NUM_IPFILTER_S_S; value++)
+    {
+        if (opt.ipfilter_ss[value])
+        {
+            if (!check_IPFilter_primitive(ref.ipfilter_ss[value], opt.ipfilter_ss[value], (value == FILTER_V_S_S_4)))
+            {
+                printf("ipfilter_ss %d failed\n", 8 / (value + 1));
+                return false;
+            }
+        }
+    }
+
     if (opt.ipfilter_p2s)
     {
         if (!check_IPFilter_primitive(ref.ipfilter_p2s, opt.ipfilter_p2s))
@@ -583,6 +657,17 @@
         }
     }
 
+    for (int value = 0; value < NUM_IPFILTER_S_S; value++)
+    {
+        if (opt.ipfilter_ss[value])
+        {
+            printf("ipfilter_ss %d\t", 8 / (value + 1));
+            REPORT_SPEEDUP(opt.ipfilter_ss[value], ref.ipfilter_ss[value],
+                           short_buff + maxVerticalfilterHalfDistance * srcStride, srcStride,
+                           IPF_vec_output_s, dstStride, width, height, val);
+        }
+    }
+
     if (opt.ipfilter_p2s)
     {
         printf("ipfilter_p2s\t");
diff -r 2a7a5766fbd8 -r 539ad4851359 source/test/ipfilterharness.h
--- a/source/test/ipfilterharness.h	Mon Nov 04 19:04:43 2013 +0800
+++ b/source/test/ipfilterharness.h	Mon Nov 04 19:05:04 2013 +0800
@@ -47,6 +47,7 @@
     bool check_IPFilter_primitive(ipfilter_p2s_t ref, ipfilter_p2s_t opt);
     bool check_IPFilter_primitive(filter_p2s_t ref, filter_p2s_t opt, int isChroma);
     bool check_IPFilter_primitive(ipfilter_s2p_t ref, ipfilter_s2p_t opt);
+    bool check_IPFilter_primitive(ipfilter_ss_t ref, ipfilter_ss_t opt, int isChroma);
     bool check_IPFilterChroma_primitive(filter_pp_t ref, filter_pp_t opt);
     bool check_IPFilterLuma_primitive(filter_pp_t ref, filter_pp_t opt);
     bool check_IPFilterLumaHV_primitive(filter_hv_pp_t ref, filter_hv_pp_t opt);



More information about the x265-devel mailing list