<div dir="ltr"><br><div class="gmail_extra"><br><br><div class="gmail_quote">On Mon, Nov 4, 2013 at 5:05 AM, Min Chen <span dir="ltr"><<a href="mailto:chenm003@163.com" target="_blank">chenm003@163.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"># HG changeset patch<br>
# User Min Chen <<a href="mailto:chenm003@163.com">chenm003@163.com</a>><br>
# Date 1383563104 -28800<br>
# Node ID 539ad4851359f96591f612f8b7b6fb0483e5a48c<br>
# Parent 2a7a5766fbd84436cdbbd5018a34db92e62553a1<br>
asm: ipfilter_ss[FILTER_V_S_S_8]<br></blockquote><div><br></div><div>When this patch is applied, the testbench enters an infinite loop in this function</div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/Lib/TLibCommon/TComPrediction.cpp<br>
--- a/source/Lib/TLibCommon/TComPrediction.cpp Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/Lib/TLibCommon/TComPrediction.cpp Mon Nov 04 19:05:04 2013 +0800<br>
@@ -537,7 +537,7 @@<br>
int filterSize = NTAPS_LUMA;<br>
int halfFilterSize = (filterSize >> 1);<br>
primitives.ipfilter_ps[FILTER_H_P_S_8](ref - (halfFilterSize - 1) * refStride, refStride, m_immedVals, tmpStride, width, height + filterSize - 1, g_lumaFilter[xFrac]);<br>
- primitives.ipfilter_ss[FILTER_V_S_S_8](m_immedVals + (halfFilterSize - 1) * tmpStride, tmpStride, dst, dstStride, width, height, g_lumaFilter[yFrac]);<br>
+ primitives.ipfilter_ss[FILTER_V_S_S_8](m_immedVals + (halfFilterSize - 1) * tmpStride, tmpStride, dst, dstStride, width, height, yFrac);<br>
}<br>
}<br>
<br>
@@ -643,9 +643,9 @@<br>
int filterSize = NTAPS_CHROMA;<br>
int halfFilterSize = (filterSize >> 1);<br>
primitives.ipfilter_ps[FILTER_H_P_S_4](refCb - (halfFilterSize - 1) * refStride, refStride, m_immedVals, extStride, cxWidth, cxHeight + filterSize - 1, g_chromaFilter[xFrac]);<br>
- primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCb, dstStride, cxWidth, cxHeight, g_chromaFilter[yFrac]);<br>
+ primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCb, dstStride, cxWidth, cxHeight, yFrac);<br>
primitives.ipfilter_ps[FILTER_H_P_S_4](refCr - (halfFilterSize - 1) * refStride, refStride, m_immedVals, extStride, cxWidth, cxHeight + filterSize - 1, g_chromaFilter[xFrac]);<br>
- primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCr, dstStride, cxWidth, cxHeight, g_chromaFilter[yFrac]);<br>
+ primitives.ipfilter_ss[FILTER_V_S_S_4](m_immedVals + (halfFilterSize - 1) * extStride, extStride, dstCr, dstStride, cxWidth, cxHeight, yFrac);<br>
}<br>
}<br>
<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/ipfilter.cpp<br>
--- a/source/common/ipfilter.cpp Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/common/ipfilter.cpp Mon Nov 04 19:05:04 2013 +0800<br>
@@ -120,8 +120,9 @@<br>
}<br>
<br>
template<int N><br>
-void filterVertical_ss_c(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, int16_t const *c)<br>
+void filterVertical_ss_c(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coefIdx)<br>
{<br>
+ const int16_t *const c = (N == 8 ? g_lumaFilter[coefIdx] : g_chromaFilter[coefIdx]);<br>
int shift = IF_FILTER_PREC;<br>
int row, col;<br>
src -= (N / 2 - 1) * srcStride;<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/primitives.h<br>
--- a/source/common/primitives.h Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/common/primitives.h Mon Nov 04 19:05:04 2013 +0800<br>
@@ -166,7 +166,7 @@<br>
typedef void (*ipfilter_pp_t)(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, const int16_t *coeff);<br>
typedef void (*ipfilter_ps_t)(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int16_t *coeff);<br>
typedef void (*ipfilter_sp_t)(int16_t *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, const int coeffIdx);<br>
-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);<br>
+typedef void (*ipfilter_ss_t)(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coeffIdx);<br>
typedef void (*ipfilter_p2s_t)(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height);<br>
typedef void (*ipfilter_s2p_t)(int16_t *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height);<br>
typedef void (*blockcpy_pp_t)(int bx, int by, pixel *dst, intptr_t dstride, pixel *src, intptr_t sstride); // dst is aligned<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/asm-primitives.cpp<br>
--- a/source/common/x86/asm-primitives.cpp Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/common/x86/asm-primitives.cpp Mon Nov 04 19:05:04 2013 +0800<br>
@@ -279,6 +279,7 @@<br>
SA8D_INTER_FROM_BLOCK(sse2);<br>
<br>
p.cvt32to16_shr = x265_cvt32to16_shr_sse2;<br>
+ p.ipfilter_ss[FILTER_V_S_S_8] = x265_interp_8tap_v_ss_sse2;<br>
}<br>
if (cpuMask & X265_CPU_SSSE3)<br>
{<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/ipfilter8.asm<br>
--- a/source/common/x86/ipfilter8.asm Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/common/x86/ipfilter8.asm Mon Nov 04 19:05:04 2013 +0800<br>
@@ -2593,3 +2593,144 @@<br>
jnz .loopH<br>
<br>
RET<br>
+<br>
+<br>
+;-------------------------------------------------------------------------------------------------------------<br>
+; void interp_8tap_v_ss(int16_t *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride, int width, int height, const int coefIdx)<br>
+;-------------------------------------------------------------------------------------------------------------<br>
+INIT_XMM sse2<br>
+<br>
+%if ARCH_X86_64<br>
+cglobal interp_8tap_v_ss, 4, 7+1, 8<br>
+%define tmp_r4d r7d<br>
+%define tmp_r5d r8d<br>
+%else<br>
+cglobal interp_8tap_v_ss, 4, 7, 8, 0-2*4<br>
+%define tmp_r4d dword [rsp + 0*4]<br>
+%define tmp_r5d dword [rsp + 1*4]<br>
+%endif<br>
+<br>
+ ; load width, height and filterIdx<br>
+ mov r4d, r4m<br>
+ mov r5d, r5m<br>
+ mov r6d, r6m<br>
+<br>
+ ; convert to word stride<br>
+ add r1, r1<br>
+ add r3, r3<br>
+<br>
+ ; stort to temporary memory or register<br>
+ shr r4d, 2<br>
+ mov tmp_r4d, r4d<br>
+ shr r5d, 2<br>
+ mov tmp_r5d, r5d<br>
+<br>
+ shl r6d, 6<br>
+%ifdef PIC<br>
+ lea r5, [tab_LumaCoeffV]<br>
+ lea r6, [r5 + r6]<br>
+%else<br>
+ lea r6, [tab_LumaCoeffV + r6]<br>
+%endif<br>
+<br>
+ lea r4, [r1 * 3]<br>
+ sub r0, r4<br>
+<br>
+.loopH:<br>
+ ; load width<br>
+ mov r4d, tmp_r4d<br>
+<br>
+.loopW:<br>
+<br>
+ movh m0, [r0] ; m0 = [0]<br>
+ movh m1, [r0 + r1] ; m1 = [1]<br>
+ lea r0, [r0 + r1 * 2]<br>
+ punpcklwd m0, m1<br>
+ pmaddwd m0, [r6 + 0 * 16] ; m0 = [0+1] = R0<br>
+<br>
+ movh m2, [r0] ; m2 = [2]<br>
+ movh m3, [r0 + r1] ; m3 = [3]<br>
+ lea r0, [r0 + r1 * 2]<br>
+ punpcklwd m1, m2<br>
+ pmaddwd m1, [r6 + 0 * 16] ; m1 = [1+2] = R1<br>
+ punpcklwd m2, m3 ; m2 = [2 3]<br>
+ pmaddwd m7, m2, [r6 + 1 * 16] ;<br>
+ paddd m0, m7 ; m0 = [0+1+2+3] = R0<br>
+ pmaddwd m2, [r6 + 0 * 16] ; m2 = [2+3] = R2<br>
+<br>
+ movh m4, [r0] ; m4 = [4]<br>
+ movh m5, [r0 + r1] ; m5 = [5]<br>
+ lea r0, [r0 + r1 * 2]<br>
+ punpcklwd m3, m4 ; m3 = [3 4]<br>
+ pmaddwd m7, m3, [r6 + 1 * 16]<br>
+ paddd m1, m7 ; m1 = [1+2+3+4] = R1<br>
+ pmaddwd m3, [r6 + 0 * 16] ; m3 = [3+4] = R3<br>
+ punpcklwd m4, m5 ; m4 = [4 5]<br>
+ pmaddwd m7, m4, [r6 + 2 * 16]<br>
+ paddd m0, m7 ; m0 = [0+1+2+3+4+5] = R0<br>
+ pmaddwd m4, [r6 + 1 * 16]<br>
+ paddd m2, m4 ; m2 = [2+3+4+5] = R2<br>
+<br>
+ movh m6, [r0] ; m6 = [6]<br>
+ movh m7, [r0 + r1] ; m7 = [7]<br>
+ lea r0, [r0 + r1 * 2]<br>
+ punpcklwd m5, m6 ; m5 = [5 6]<br>
+ pmaddwd m4, m5, [r6 + 2 * 16]<br>
+ paddd m1, m4 ; m1 = [1+2+3+4+5+6] = R1<br>
+ pmaddwd m5, [r6 + 1 * 16]<br>
+ paddd m3, m5 ; m3 = [3+4+5+6] = R3<br>
+ punpcklwd m6, m7 ; m6 = [6 7]<br>
+ pmaddwd m4, m6, [r6 + 3 * 16]<br>
+ paddd m0, m4 ; m0 = [0+1+2+3+4+5+6+7]= R0<br>
+ pmaddwd m6, [r6 + 2 * 16]<br>
+ paddd m2, m6 ; m2 = [2+3+4+5+6+7] = R2<br>
+ psrad m0, 6<br>
+ packssdw m0, m0<br>
+ movh [r2], m0 ; store [0]<br>
+<br>
+ movh m4, [r0] ; m4 = [8]<br>
+ movh m5, [r0 + r1] ; m5 = [9]<br>
+ punpcklwd m7, m4 ; m7 = [7 8]<br>
+ pmaddwd m6, m7, [r6 + 3 * 16]<br>
+ paddd m1, m6 ; m1 = [1+2+3+4+5+6+7+8]= R1<br>
+ pmaddwd m7, [r6 + 2 * 16]<br>
+ paddd m3, m7 ; m3 = [3+4+5+6+7+8] = R3<br>
+ psrad m1, 6<br>
+ packssdw m1, m1<br>
+ movh [r2 + r3], m1 ; store [1]<br>
+ punpcklwd m4, m5 ; m4 = [8 9]<br>
+ pmaddwd m4, [r6 + 3 * 16]<br>
+ paddd m2, m4 ; m2 = [2+3+4+5+6+7+8+9]= R2<br>
+ psrad m2, 6<br>
+ packssdw m2, m2<br>
+ movh [r2 + r3 * 2], m2 ; store [2]<br>
+ lea r2, [r2 + r3 * 2]<br>
+<br>
+ movh m4, [r0 + r1 * 2] ; m4 = [10]<br>
+ punpcklwd m5, m4 ; m5 = [9 10]<br>
+ pmaddwd m5, [r6 + 3 * 16]<br>
+ paddd m3, m5 ; m3 = [3+4+5+6+7+8+9+10]=R3<br>
+ psrad m3, 6<br>
+ packssdw m3, m3<br>
+ movh [r2 + r3], m3 ; store [3]<br>
+<br>
+ lea r5, [r1 * 8 - 8]<br>
+ sub r0, r5<br>
+ lea r5, [r3 * 2 - 8]<br>
+ sub r2, r5<br>
+<br>
+ dec r4d<br>
+ jnz .loopW<br>
+<br>
+ ; move to next row<br>
+ mov r4d, tmp_r4d<br>
+ shl r4d, 3<br>
+ lea r0, [r0 + r1 * 4]<br>
+ sub r0, r4<br>
+ lea r2, [r2 + r3 * 4]<br>
+ sub r2, r4<br>
+<br>
+ dec tmp_r5d<br>
+ jnz .loopH<br>
+<br>
+ RET<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/common/x86/pixel.h<br>
--- a/source/common/x86/pixel.h Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/common/x86/pixel.h Mon Nov 04 19:05:04 2013 +0800<br>
@@ -214,6 +214,7 @@<br>
uint64_t x265_pixel_sa8d_satd_16x16_avx2(pixel *pix1, intptr_t stride1, pixel *pix2, intptr_t stride2);<br>
<br>
void x265_cvt32to16_shr_sse2(int16_t *dst, int *src, intptr_t, int, int);<br>
+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);<br>
<br>
#define DECL_HEVC_SSD(suffix) \<br>
int x265_pixel_ssd_32x64_ ## suffix(pixel *, intptr_t, pixel *, intptr_t); \<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/encoder/motion.cpp<br>
--- a/source/encoder/motion.cpp Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/encoder/motion.cpp Mon Nov 04 19:05:04 2013 +0800<br>
@@ -1213,7 +1213,7 @@<br>
int filterSize = NTAPS_LUMA;<br>
int halfFilterSize = (filterSize >> 1);<br>
primitives.ipfilter_ps[FILTER_H_P_S_8](fref - (halfFilterSize - 1) * ref->lumaStride, ref->lumaStride, immedVal, blockwidth, blockwidth, realHeight + filterSize - 1, g_lumaFilter[xFrac]);<br>
- primitives.ipfilter_ss[FILTER_V_S_S_8](immedVal + (halfFilterSize - 1) * blockwidth, blockwidth, immedVal2, FENC_STRIDE, blockwidth, realHeight, g_lumaFilter[yFrac]);<br>
+ primitives.ipfilter_ss[FILTER_V_S_S_8](immedVal + (halfFilterSize - 1) * blockwidth, blockwidth, immedVal2, FENC_STRIDE, blockwidth, realHeight, yFrac);<br>
primitives.weightpUni(immedVal2, subpelbuf, FENC_STRIDE, FENC_STRIDE, blockwidth, realHeight, ref->weight, local_round, local_shift, ref->offset);<br>
}<br>
}<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/test/ipfilterharness.cpp<br>
--- a/source/test/ipfilterharness.cpp Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/test/ipfilterharness.cpp Mon Nov 04 19:05:04 2013 +0800<br>
@@ -318,6 +318,68 @@<br>
return true;<br>
}<br>
<br>
+bool IPFilterHarness::check_IPFilter_primitive(ipfilter_ss_t ref, ipfilter_ss_t opt, int isChroma)<br>
+{<br>
+ int rand_val, rand_srcStride, rand_dstStride;<br>
+ const int min_size = isChroma ? 2 : 4;<br>
+<br>
+ // NOTE: refill data to avoid overflow<br>
+ const int max_filter_val = 64 * (1 << 8);<br>
+ for (int i = 0; i < ipf_t_size; i++)<br>
+ {<br>
+ short_buff[i] = rand() % (2 * max_filter_val) - max_filter_val;<br>
+ }<br>
+<br>
+ for (int i = 0; i <= 1000; i++)<br>
+ {<br>
+ int rand_height = rand() % 100; // Randomly generated Height<br>
+ int rand_width = rand() % 100; // Randomly generated Width<br>
+<br>
+ memset(IPF_vec_output_s, 0xCD, ipf_t_size); // Initialize output buffer to zero<br>
+ memset(IPF_C_output_s, 0xCD, ipf_t_size); // Initialize output buffer to zero<br>
+<br>
+ rand_val = rand() % 4; // Random offset in the filter<br>
+ rand_srcStride = rand() % 100; // Randomly generated srcStride<br>
+ rand_dstStride = rand() % 100; // Randomly generated dstStride<br>
+<br>
+ rand_width &= ~(min_size - 1);<br>
+ if (rand_width < min_size)<br>
+ rand_width = min_size;<br>
+<br>
+ rand_height &= ~(min_size - 1);<br>
+ if (rand_height < min_size)<br>
+ rand_height = min_size;<br>
+<br>
+ if (rand_srcStride < rand_width)<br>
+ rand_srcStride = rand_width;<br>
+<br>
+ if (rand_dstStride < rand_width)<br>
+ rand_dstStride = rand_width;<br>
+<br>
+ ref(short_buff + 3 * rand_srcStride,<br>
+ rand_srcStride,<br>
+ IPF_C_output_s,<br>
+ rand_dstStride,<br>
+ rand_width,<br>
+ rand_height, rand_val<br>
+ );<br>
+ opt(short_buff + 3 * rand_srcStride,<br>
+ rand_srcStride,<br>
+ IPF_vec_output_s,<br>
+ rand_dstStride,<br>
+ rand_width,<br>
+ rand_height, rand_val<br>
+ );<br>
+<br>
+ if (memcmp(IPF_C_output_s, IPF_vec_output_s, ipf_t_size * sizeof(int16_t)))<br>
+ {<br>
+ return false;<br>
+ }<br>
+ }<br>
+<br>
+ return true;<br>
+}<br>
+<br>
bool IPFilterHarness::check_IPFilterChroma_primitive(filter_pp_t ref, filter_pp_t opt)<br>
{<br>
int rand_srcStride, rand_dstStride, rand_coeffIdx;<br>
@@ -452,6 +514,18 @@<br>
}<br>
}<br>
<br>
+ for (int value = 0; value < NUM_IPFILTER_S_S; value++)<br>
+ {<br>
+ if (opt.ipfilter_ss[value])<br>
+ {<br>
+ if (!check_IPFilter_primitive(ref.ipfilter_ss[value], opt.ipfilter_ss[value], (value == FILTER_V_S_S_4)))<br>
+ {<br>
+ printf("ipfilter_ss %d failed\n", 8 / (value + 1));<br>
+ return false;<br>
+ }<br>
+ }<br>
+ }<br>
+<br>
if (opt.ipfilter_p2s)<br>
{<br>
if (!check_IPFilter_primitive(ref.ipfilter_p2s, opt.ipfilter_p2s))<br>
@@ -583,6 +657,17 @@<br>
}<br>
}<br>
<br>
+ for (int value = 0; value < NUM_IPFILTER_S_S; value++)<br>
+ {<br>
+ if (opt.ipfilter_ss[value])<br>
+ {<br>
+ printf("ipfilter_ss %d\t", 8 / (value + 1));<br>
+ REPORT_SPEEDUP(opt.ipfilter_ss[value], ref.ipfilter_ss[value],<br>
+ short_buff + maxVerticalfilterHalfDistance * srcStride, srcStride,<br>
+ IPF_vec_output_s, dstStride, width, height, val);<br>
+ }<br>
+ }<br>
+<br>
if (opt.ipfilter_p2s)<br>
{<br>
printf("ipfilter_p2s\t");<br>
diff -r 2a7a5766fbd8 -r 539ad4851359 source/test/ipfilterharness.h<br>
--- a/source/test/ipfilterharness.h Mon Nov 04 19:04:43 2013 +0800<br>
+++ b/source/test/ipfilterharness.h Mon Nov 04 19:05:04 2013 +0800<br>
@@ -47,6 +47,7 @@<br>
bool check_IPFilter_primitive(ipfilter_p2s_t ref, ipfilter_p2s_t opt);<br>
bool check_IPFilter_primitive(filter_p2s_t ref, filter_p2s_t opt, int isChroma);<br>
bool check_IPFilter_primitive(ipfilter_s2p_t ref, ipfilter_s2p_t opt);<br>
+ bool check_IPFilter_primitive(ipfilter_ss_t ref, ipfilter_ss_t opt, int isChroma);<br>
bool check_IPFilterChroma_primitive(filter_pp_t ref, filter_pp_t opt);<br>
bool check_IPFilterLuma_primitive(filter_pp_t ref, filter_pp_t opt);<br>
bool check_IPFilterLumaHV_primitive(filter_hv_pp_t ref, filter_hv_pp_t opt);<br>
<br>
_______________________________________________<br>
x265-devel mailing list<br>
<a href="mailto:x265-devel@videolan.org">x265-devel@videolan.org</a><br>
<a href="https://mailman.videolan.org/listinfo/x265-devel" target="_blank">https://mailman.videolan.org/listinfo/x265-devel</a><br>
</blockquote></div><br><br clear="all"><div><br></div>-- <br>Steve Borho
</div></div>