[x265] [PATCH 4 of 4] asm: chroma_p2s to replace ipfilter_p2s
Min Chen
chenm003 at 163.com
Thu Oct 31 14:03:02 CET 2013
# HG changeset patch
# User Min Chen <chenm003 at 163.com>
# Date 1383224503 -28800
# Node ID 4a40c4069ad12bc72a1c443b45a91c65d319d35d
# Parent 21dbf988079b0e33265ae48578c26347cc779fbe
asm: chroma_p2s to replace ipfilter_p2s
diff -r 21dbf988079b -r 4a40c4069ad1 source/Lib/TLibCommon/TComPrediction.cpp
--- a/source/Lib/TLibCommon/TComPrediction.cpp Thu Oct 31 21:01:29 2013 +0800
+++ b/source/Lib/TLibCommon/TComPrediction.cpp Thu Oct 31 21:01:43 2013 +0800
@@ -619,10 +619,13 @@
uint32_t cxWidth = width >> 1;
uint32_t cxHeight = height >> 1;
+ assert(dstStride == MAX_CU_SIZE / 2);
+ assert(((cxWidth | cxHeight) % 2) == 0);
+
if ((yFrac | xFrac) == 0)
{
- primitives.ipfilter_p2s(refCb, refStride, dstCb, dstStride, cxWidth, cxHeight);
- primitives.ipfilter_p2s(refCr, refStride, dstCr, dstStride, cxWidth, cxHeight);
+ primitives.chroma_p2s(refCb, refStride, dstCb, cxWidth, cxHeight);
+ primitives.chroma_p2s(refCr, refStride, dstCr, cxWidth, cxHeight);
}
else if (yFrac == 0)
{
diff -r 21dbf988079b -r 4a40c4069ad1 source/common/ipfilter.cpp
--- a/source/common/ipfilter.cpp Thu Oct 31 21:01:29 2013 +0800
+++ b/source/common/ipfilter.cpp Thu Oct 31 21:01:43 2013 +0800
@@ -264,6 +264,7 @@
}
}
+template<int dstStride>
void filterConvertPelToShort_c(pixel *src, intptr_t srcStride, int16_t *dst, int width, int height)
{
int shift = IF_INTERNAL_PREC - X265_DEPTH;
@@ -278,7 +279,7 @@
}
src += srcStride;
- dst += MAX_CU_SIZE;
+ dst += dstStride;
}
}
@@ -489,7 +490,8 @@
p.ipfilter_p2s = filterConvertPelToShort_c;
p.ipfilter_s2p = filterConvertShortToPel_c;
- p.luma_p2s = filterConvertPelToShort_c;
+ p.luma_p2s = filterConvertPelToShort_c<MAX_CU_SIZE>;
+ p.chroma_p2s = filterConvertPelToShort_c<MAX_CU_SIZE/2>;
p.extendRowBorder = extendCURowColBorder;
}
diff -r 21dbf988079b -r 4a40c4069ad1 source/common/primitives.h
--- a/source/common/primitives.h Thu Oct 31 21:01:29 2013 +0800
+++ b/source/common/primitives.h Thu Oct 31 21:01:43 2013 +0800
@@ -254,6 +254,7 @@
filter_pp_t luma_vpp[NUM_LUMA_PARTITIONS];
filter_hv_pp_t luma_hvpp[NUM_LUMA_PARTITIONS];
filter_p2s_t luma_p2s;
+ filter_p2s_t chroma_p2s;
intra_dc_t intra_pred_dc;
intra_planar_t intra_pred_planar;
diff -r 21dbf988079b -r 4a40c4069ad1 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Thu Oct 31 21:01:29 2013 +0800
+++ b/source/common/x86/asm-primitives.cpp Thu Oct 31 21:01:43 2013 +0800
@@ -318,6 +318,7 @@
p.luma_hvpp[LUMA_8x8] = x265_interp_8tap_hv_pp_8x8_ssse3;
p.ipfilter_sp[FILTER_V_S_P_8] = x265_interp_8tap_v_sp_ssse3;
p.luma_p2s = x265_luma_p2s_ssse3;
+ p.chroma_p2s = x265_chroma_p2s_ssse3;
}
if (cpuMask & X265_CPU_SSE4)
{
diff -r 21dbf988079b -r 4a40c4069ad1 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Thu Oct 31 21:01:29 2013 +0800
+++ b/source/common/x86/ipfilter8.asm Thu Oct 31 21:01:43 2013 +0800
@@ -2124,3 +2124,61 @@
jnz .loopH
RET
+
+
+; TODO: combin of U and V is more performance, but need more register
+; TODO: use two path for height alignment to 4 and otherwise may improvement 10% performance, but code is more complex, so I disable it
+INIT_XMM ssse3
+cglobal chroma_p2s, 3, 7, 6
+
+ ; load width and height
+ mov r3d, r3m
+ mov r4d, r4m
+
+ ; load constant
+ mova m4, [tab_c_128]
+ mova m5, [tab_c_64_n64]
+
+.loopH:
+
+ xor r5d, r5d
+.loopW:
+ lea r6, [r0 + r5]
+
+ movh m0, [r6]
+ punpcklbw m0, m4
+ pmaddubsw m0, m5
+
+ movh m1, [r6 + r1]
+ punpcklbw m1, m6
+ pmaddubsw m1, m7
+
+ add r5d, 8
+ cmp r5d, r3d
+ lea r6, [r2 + r5 * 2]
+ jg .width2
+ movu [r6 + FENC_STRIDE / 2 * 0 - 16], m0
+ movu [r6 + FENC_STRIDE / 2 * 2 - 16], m1
+ je .nextH
+ jmp .loopW
+
+.width4:
+ cmp r3d, 4
+ jl .width2
+ movh [r6 + FENC_STRIDE / 2 * 0 - 16], m0
+ movh [r6 + FENC_STRIDE / 2 * 2 - 16], m1
+ lea r6, [r6 + 8]
+ jz .nextH
+
+.width2:
+ movd [r6 + FENC_STRIDE / 2 * 0 - 16], m0
+ movd [r6 + FENC_STRIDE / 2 * 2 - 16], m1
+
+.nextH:
+ lea r0, [r0 + r1 * 2]
+ add r2, FENC_STRIDE / 2 * 4
+
+ sub r4d, 2
+ jnz .loopH
+
+ RET
diff -r 21dbf988079b -r 4a40c4069ad1 source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h Thu Oct 31 21:01:29 2013 +0800
+++ b/source/common/x86/ipfilter8.h Thu Oct 31 21:01:43 2013 +0800
@@ -91,6 +91,7 @@
void x265_interp_8tap_hv_pp_8x8_ssse3(pixel * src, intptr_t srcStride, pixel * dst, intptr_t dstStride, int idxX, int idxY);
void x265_interp_8tap_v_sp_ssse3(int16_t *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int width, int height, const int coeffIdx);
void x265_luma_p2s_ssse3(pixel *src, intptr_t srcStride, int16_t *dst, int width, int height);
+void x265_chroma_p2s_ssse3(pixel *src, intptr_t srcStride, int16_t *dst, int width, int height);
#undef SETUP_CHROMA_FUNC_DEF
#undef SETUP_LUMA_FUNC_DEF
diff -r 21dbf988079b -r 4a40c4069ad1 source/test/ipfilterharness.cpp
--- a/source/test/ipfilterharness.cpp Thu Oct 31 21:01:29 2013 +0800
+++ b/source/test/ipfilterharness.cpp Thu Oct 31 21:01:43 2013 +0800
@@ -240,14 +240,15 @@
return true;
}
-bool IPFilterHarness::check_IPFilter_primitive(filter_p2s_t ref, filter_p2s_t opt)
+bool IPFilterHarness::check_IPFilter_primitive(filter_p2s_t ref, filter_p2s_t opt, int isChroma)
{
- int16_t rand_srcStride;
+ intptr_t rand_srcStride;
+ const int min_size = isChroma ? 2 : 4;
for (int i = 0; i <= 1000; i++)
{
- int16_t rand_height = (int16_t)rand() % 100; // Randomly generated Height
- int16_t rand_width = (int16_t)rand() % 100; // Randomly generated Width
+ int rand_height = (int16_t)rand() % 100; // Randomly generated Height
+ int rand_width = (int16_t)rand() % 100; // Randomly generated Width
memset(IPF_vec_output_s, 0, ipf_t_size); // Initialize output buffer to zero
memset(IPF_C_output_s, 0, ipf_t_size); // Initialize output buffer to zero
@@ -256,13 +257,13 @@
if (rand_srcStride < rand_width)
rand_srcStride = rand_width;
- rand_width %= 4;
- if (rand_width < 4)
- rand_width = 4;
+ rand_width %= min_size;
+ if (rand_width < min_size)
+ rand_width = min_size;
- rand_height %= 4;
- if (rand_height < 4)
- rand_height = 4;
+ rand_height %= min_size;
+ if (rand_height < min_size)
+ rand_height = min_size;
ref(pixel_buff,
rand_srcStride,
@@ -461,7 +462,16 @@
if (opt.luma_p2s)
{
- if (!check_IPFilter_primitive(ref.luma_p2s, opt.luma_p2s))
+ if (!check_IPFilter_primitive(ref.luma_p2s, opt.luma_p2s, 0))
+ {
+ printf("ipfilter_p2s failed\n");
+ return false;
+ }
+ }
+
+ if (opt.chroma_p2s)
+ {
+ if (!check_IPFilter_primitive(ref.chroma_p2s, opt.chroma_p2s, 1))
{
printf("ipfilter_p2s failed\n");
return false;
@@ -586,6 +596,13 @@
pixel_buff, srcStride, IPF_vec_output_s, width, height);
}
+ if (opt.chroma_p2s)
+ {
+ printf("chroma_p2s\t");
+ REPORT_SPEEDUP(opt.chroma_p2s, ref.chroma_p2s,
+ pixel_buff, srcStride, IPF_vec_output_s, width, height);
+ }
+
if (opt.ipfilter_s2p)
{
printf("ipfilter_s2p\t");
diff -r 21dbf988079b -r 4a40c4069ad1 source/test/ipfilterharness.h
--- a/source/test/ipfilterharness.h Thu Oct 31 21:01:29 2013 +0800
+++ b/source/test/ipfilterharness.h Thu Oct 31 21:01:43 2013 +0800
@@ -45,7 +45,7 @@
bool check_IPFilter_primitive(ipfilter_ps_t ref, ipfilter_ps_t opt);
bool check_IPFilter_primitive(ipfilter_sp_t ref, ipfilter_sp_t opt);
bool check_IPFilter_primitive(ipfilter_p2s_t ref, ipfilter_p2s_t opt);
- bool check_IPFilter_primitive(filter_p2s_t ref, filter_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_IPFilterChroma_primitive(filter_pp_t ref, filter_pp_t opt);
bool check_IPFilterLuma_primitive(filter_pp_t ref, filter_pp_t opt);
More information about the x265-devel
mailing list