Re: [libav-devel] [PATCH] x86: Port gradfun to yasm
--- libavfilter/x86/vf_gradfun.asm | 82 +- 1 file changed, 24 insertions(+), 58 deletions(-) diff --git a/libavfilter/x86/vf_gradfun.asm b/libavfilter/x86/vf_gradfun.asm index e1737dd..0ba051d 100644 --- a/libavfilter/x86/vf_gradfun.asm +++ b/libavfilter/x86/vf_gradfun.asm @@ -27,6 +27,27 @@ pw_ff: times 8 dw 0xFF SECTION .text +%macro FILTER_LINE 1 +movh m0, [r2+r0] +movh m1, [r3+r0] +punpcklbw m0, m7 +punpcklwd m1, m1 +psllw m0, 7 +psubw m1, m0 +PABSW m2, m1 +pmulhuwm2, m5 +psubw m2, m6 +pminsw m2, m7 +pmullw m2, m2 +psllw m1, 2 +paddw m0, %1 +pmulhw m1, m2 +paddw m0, m1 +psraw m0, 7 +packuswb m0, m0 +movh [r1+r0], m0 +%endmacro + INIT_MMX mmxext cglobal gradfun_filter_line, 6, 6 movh m5, r4d @@ -36,48 +57,10 @@ cglobal gradfun_filter_line, 6, 6 mova m3, [r5] mova m4, [r5+8] .loop: -movh m0, [r2+r0] -movh m1, [r3+r0] -punpcklbw m0, m7 -punpcklwd m1, m1 -psllw m0, 7 -pxor m2, m2 -psubw m1, m0 -psubw m2, m1 -pmaxswm2, m1 -pmulhuw m2, m5 -psubw m2, m6 -pminswm2, m7 -pmullwm2, m2 -paddw m0, m3 -psllw m1, 2 -pmulhwm1, m2 -paddw m0, m1 -psraw m0, 7 -packuswb m0, m0 -movh [r1+r0], m0 +FILTER_LINE m3 add r0, 4 jge .end -movh m0, [r2+r0] -movh m1, [r3+r0] -punpcklbw m0, m7 -punpcklwd m1, m1 -psllw m0, 7 -pxor m2, m2 -psubw m1, m0 -psubw m2, m1 -pmaxswm2, m1 -pmulhuw m2, m5 -psubw m2, m6 -pminswm2, m7 -pmullwm2, m2 -paddw m0, m4 -psllw m1, 2 -pmulhwm1, m2 -paddw m0, m1 -psraw m0, 7 -packuswb m0, m0 -movh [r1+r0], m0 +FILTER_LINE m4 add r0, 4 jl .loop .end: @@ -92,24 +75,7 @@ cglobal gradfun_filter_line, 6, 6, 8 punpcklqdq m5, m5 mova m4, [r5] .loop: -movh m0, [r2+r0] -movh m1, [r3+r0] -punpcklbw m0, m7 -punpcklwd m1, m1 -psllw m0, 7 -psubw m1, m0 -pabsw m2, m1 -pmulhuwm2, m5 -psubw m2, m6 -pminsw m2, m7 -pmullw m2, m2 -psllw m1, 2 -paddw m0, m4 -pmulhw m1, m2 -paddw m0, m1 -psraw m0, 7 -packuswb m0, m0 -movh [r1+r0], m0 +FILTER_LINE m4 addr0, 8 jl .loop REP_RET -- 1.8.3.2 ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 01/10] x86inc: activate REP_RET automatically
On Mon, 30 Sep 2013, Diego Biurrun wrote: On Mon, Sep 30, 2013 at 04:38:01PM +0100, Derek Buitenhuis wrote: On 9/30/2013 4:28 PM, Diego Biurrun wrote: Not sure what you are trying to achieve with the if, $(STRIP) should always be set ... If we bring back stripping, we'll need the rest of the portability parts from e0be794 as well. For now, I do not think this patch is necessary, no? Are you referring to 01/10 or the build system patch? If the latter, then I don't know what the side effect of keeping those labels around is ... The effect of keeping those labels around is making debugging harder. Because those labels are meaningless, and complicate the disassembly. Also, gdb can't tell the difference between them and function entrypoints. This new strip command is irrelevant to any usage of libav that would have used the old fully stripped version, because the old one was for non-debug use. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 01/10] x86inc: activate REP_RET automatically
On Wed, 11 Sep 2013, Hendrik Leppkes wrote:
From: Loren Merritt pengv...@akuvian.org
Now RET checks whether it immediately follows a branch, so the programmer
dosen't have to keep track of that condition.
REP_RET is still needed manually when it's a branch target, but that's much
rarer.
The implementation involves lots of spurious labels, but that's ok because we
strip them.
That's true of x264's buildsystem, but not libav yet. So import that patch too:
---
configure | 3 +++
library.mak | 1 +
2 files changed, 4 insertions(+)
diff --git a/configure b/configure
index e6f8b52..fce4d25 100755
--- a/configure
+++ b/configure
@@ -1973,6 +1973,7 @@ nm_default=nm -g
objformat=elf
pkg_config_default=pkg-config
ranlib=ranlib
+strip=strip
yasmexe=yasm
nogas=:
@@ -2231,6 +2232,7 @@ cc_default=${cross_prefix}${cc_default}
nm_default=${cross_prefix}${nm_default}
pkg_config_default=${cross_prefix}${pkg_config_default}
ranlib=${cross_prefix}${ranlib}
+strip=${cross_prefix}${strip}
sysinclude_default=${sysroot}/usr/include
@@ -4140,6 +4142,7 @@ AR=$ar
ARFLAGS=$arflags
AR_O=$ar_o
RANLIB=$ranlib
+STRIP=$strip
LN_S=$ln_s
CPPFLAGS=$CPPFLAGS
CFLAGS=$CFLAGS
diff --git a/library.mak b/library.mak
index 88d33dc..58f6360 100644
--- a/library.mak
+++ b/library.mak
@@ -25,6 +25,7 @@ $(SUBDIR)%-test.i: $(SUBDIR)%.c
$(SUBDIR)x86/%.o: $(SUBDIR)x86/%.asm
$(DEPYASM) $(YASMFLAGS) -I $(D)/ -M -o $@ $ $(@:.o=.d)
$(YASM) $(YASMFLAGS) -I $(D)/ -o $@ $
+ -@ $(if $(STRIP), $(STRIP) -wN '..@*' $@)
LIBOBJS := $(OBJS) $(SUBDIR)%.h.o $(TESTOBJS)
$(LIBOBJS) $(LIBOBJS:.o=.i): CPPFLAGS += -DHAVE_AV_CONFIG_H
--
1.8.3.2
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 1/4] lpc: remove decay argument
We never used the rolling-average mode, and this makes av_update_lls 15% faster.
---
libavcodec/lpc.c | 2 +-
libavutil/lls.c | 7 +++
libavutil/lls.h | 2 +-
3 files changed, 5 insertions(+), 6 deletions(-)
diff --git a/libavcodec/lpc.c b/libavcodec/lpc.c
index bada368..c098b0f 100644
--- a/libavcodec/lpc.c
+++ b/libavcodec/lpc.c
@@ -222,7 +222,7 @@ int ff_lpc_calc_coefs(LPCContext *s,
}else
weight++;
-avpriv_update_lls(m[pass1], var, 1.0);
+avpriv_update_lls(m[pass1], var);
}
avpriv_solve_lls(m[pass1], 0.001, 0);
}
diff --git a/libavutil/lls.c b/libavutil/lls.c
index 246189b..2061e6a 100644
--- a/libavutil/lls.c
+++ b/libavutil/lls.c
@@ -38,13 +38,12 @@ av_cold void avpriv_init_lls(LLSModel *m, int indep_count)
m-indep_count = indep_count;
}
-void avpriv_update_lls(LLSModel *m, double *var, double decay)
+void avpriv_update_lls(LLSModel *m, double *var)
{
int i, j;
for (i = 0; i = m-indep_count; i++) {
for (j = i; j = m-indep_count; j++) {
-m-covariance[i][j] *= decay;
m-covariance[i][j] += var[i] * var[j];
}
}
@@ -125,7 +124,7 @@ av_cold void av_init_lls(LLSModel *m, int indep_count)
}
void av_update_lls(LLSModel *m, double *param, double decay)
{
-avpriv_update_lls(m, param, decay);
+avpriv_update_lls(m, param);
}
void av_solve_lls(LLSModel *m, double threshold, int min_order)
{
@@ -160,7 +159,7 @@ int main(void)
var[1] = var[0] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
var[2] = var[1] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
var[3] = var[2] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
-avpriv_update_lls(m, var, 0.99);
+avpriv_update_lls(m, var);
avpriv_solve_lls(m, 0.001, 0);
for (order = 0; order 3; order++) {
eval = avpriv_evaluate_lls(m, var + 1, order);
diff --git a/libavutil/lls.h b/libavutil/lls.h
index f493076..9c71cf9 100644
--- a/libavutil/lls.h
+++ b/libavutil/lls.h
@@ -40,7 +40,7 @@ typedef struct LLSModel {
} LLSModel;
void avpriv_init_lls(LLSModel *m, int indep_count);
-void avpriv_update_lls(LLSModel *m, double *param, double decay);
+void avpriv_update_lls(LLSModel *m, double *param);
void avpriv_solve_lls(LLSModel *m, double threshold, unsigned short min_order);
double avpriv_evaluate_lls(LLSModel *m, double *param, int order);
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 0/4] lpc optimizations v2
changes: Converted API to function pointers, like we usually use for DSP functions. Removed the unused rolling-average mode from the C implementation. My asm already didn't support it. Updated doxygen, headers, and function prefixes. (Not reposting the unchanged patches.) ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 4/4] x86: lpc: simd av_evaluate_lls
1.5x-1.8x faster on sandybridge
---
libavutil/x86/lls.asm| 38 ++
libavutil/x86/lls_init.c | 3 +++
2 files changed, 41 insertions(+)
diff --git a/libavutil/x86/lls.asm b/libavutil/x86/lls.asm
index 92c00fc..92b7f95 100644
--- a/libavutil/x86/lls.asm
+++ b/libavutil/x86/lls.asm
@@ -194,3 +194,41 @@ cglobal update_lls, 3,6,8, ctx, var, count, i, j, count2
jle .loop2x1
.ret:
REP_RET
+
+
+INIT_XMM sse2
+cglobal evaluate_lls, 2,4,2, ctx, var, order, i
+; This function is often called on the same buffer as update_lls, but with
+; an offset. They can't both be aligned.
+; Load halves rather than movu to avoid store-forwarding stalls, since the
+; input was initialized immediately prior to this function using scalar
math.
+%define coefsq ctxq
+mov id, orderd
+imulorderd, MAX_VARS
+lea coefsq, [ctxq + LLSModel.coeff + orderq*8]
+movsd m0, [varq]
+movhpd m0, [varq + 8]
+mulpd m0, [coefsq]
+lea coefsq, [coefsq + iq*8]
+lea varq, [varq + iq*8]
+neg iq
+add iq, 2
+.loop:
+movsd m1, [varq + iq*8]
+movhpd m1, [varq + iq*8 + 8]
+mulpd m1, [coefsq + iq*8]
+addpd m0, m1
+add iq, 2
+jl .loop
+jg .skip1
+movsd m1, [varq + iq*8]
+mulsd m1, [coefsq + iq*8]
+addpd m0, m1
+.skip1:
+movhlps m1, m0
+addsd m0, m1
+%if ARCH_X86_32
+movsd r0m, m0
+fld qword r0m
+%endif
+RET
diff --git a/libavutil/x86/lls_init.c b/libavutil/x86/lls_init.c
index 8a80f83..888bc54 100644
--- a/libavutil/x86/lls_init.c
+++ b/libavutil/x86/lls_init.c
@@ -25,12 +25,15 @@
void ff_update_lls_sse2(LLSModel *m, double *var);
void ff_update_lls_avx(LLSModel *m, double *var);
+double ff_evaluate_lls_sse2(LLSModel *m, double *var, int order);
av_cold void ff_init_lls_x86(LLSModel *m)
{
int cpu_flags = av_get_cpu_flags();
if (EXTERNAL_SSE2(cpu_flags)) {
m-update_lls = ff_update_lls_sse2;
+if (m-indep_count = 4)
+m-evaluate_lls = ff_evaluate_lls_sse2;
}
if (EXTERNAL_AVX(cpu_flags)) {
m-update_lls = ff_update_lls_avx;
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 2/4] lpc: use function pointers, in preparation for asm
---
libavcodec/lpc.c | 4 ++--
libavutil/lls.c | 26 ++
libavutil/lls.h | 15 +--
3 files changed, 29 insertions(+), 16 deletions(-)
diff --git a/libavcodec/lpc.c b/libavcodec/lpc.c
index c098b0f..f60976b 100644
--- a/libavcodec/lpc.c
+++ b/libavcodec/lpc.c
@@ -212,7 +212,7 @@ int ff_lpc_calc_coefs(LPCContext *s,
if(pass){
double eval, inv, rinv;
-eval= avpriv_evaluate_lls(m[(pass-1)1], var+1,
max_order-1);
+eval= m[(pass-1)1].evaluate_lls(m[(pass-1)1], var+1,
max_order-1);
eval= (512pass) + fabs(eval - var[0]);
inv = 1/eval;
rinv = sqrt(inv);
@@ -222,7 +222,7 @@ int ff_lpc_calc_coefs(LPCContext *s,
}else
weight++;
-avpriv_update_lls(m[pass1], var);
+m[pass1].update_lls(m[pass1], var);
}
avpriv_solve_lls(m[pass1], 0.001, 0);
}
diff --git a/libavutil/lls.c b/libavutil/lls.c
index 2061e6a..5a3e448 100644
--- a/libavutil/lls.c
+++ b/libavutil/lls.c
@@ -32,13 +32,7 @@
#include version.h
#include lls.h
-av_cold void avpriv_init_lls(LLSModel *m, int indep_count)
-{
-memset(m, 0, sizeof(LLSModel));
-m-indep_count = indep_count;
-}
-
-void avpriv_update_lls(LLSModel *m, double *var)
+static void update_lls(LLSModel *m, double *var)
{
int i, j;
@@ -106,7 +100,7 @@ void avpriv_solve_lls(LLSModel *m, double threshold,
unsigned short min_order)
}
}
-double avpriv_evaluate_lls(LLSModel *m, double *param, int order)
+static double evaluate_lls(LLSModel *m, double *param, int order)
{
int i;
double out = 0;
@@ -117,6 +111,14 @@ double avpriv_evaluate_lls(LLSModel *m, double *param, int
order)
return out;
}
+av_cold void avpriv_init_lls(LLSModel *m, int indep_count)
+{
+memset(m, 0, sizeof(LLSModel));
+m-indep_count = indep_count;
+m-update_lls = update_lls;
+m-evaluate_lls = evaluate_lls;
+}
+
#if FF_API_LLS_PRIVATE
av_cold void av_init_lls(LLSModel *m, int indep_count)
{
@@ -124,7 +126,7 @@ av_cold void av_init_lls(LLSModel *m, int indep_count)
}
void av_update_lls(LLSModel *m, double *param, double decay)
{
-avpriv_update_lls(m, param);
+m-update_lls(m, param);
}
void av_solve_lls(LLSModel *m, double threshold, int min_order)
{
@@ -132,7 +134,7 @@ void av_solve_lls(LLSModel *m, double threshold, int
min_order)
}
double av_evaluate_lls(LLSModel *m, double *param, int order)
{
-return avpriv_evaluate_lls(m, param, order);
+return m-evaluate_lls(m, param, order);
}
#endif /* FF_API_LLS_PRIVATE */
@@ -159,10 +161,10 @@ int main(void)
var[1] = var[0] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
var[2] = var[1] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
var[3] = var[2] + av_lfg_get(lfg) / (double) UINT_MAX - 0.5;
-avpriv_update_lls(m, var);
+m.update_lls(m, var);
avpriv_solve_lls(m, 0.001, 0);
for (order = 0; order 3; order++) {
-eval = avpriv_evaluate_lls(m, var + 1, order);
+eval = m.evaluate_lls(m, var + 1, order);
printf(real:%9f order:%d pred:%9f var:%f coeffs:%f %9f %9f\n,
var[0], order, eval, sqrt(m.variance[order] / (i + 1)),
m.coeff[order][0], m.coeff[order][1],
diff --git a/libavutil/lls.h b/libavutil/lls.h
index 9c71cf9..8183440 100644
--- a/libavutil/lls.h
+++ b/libavutil/lls.h
@@ -37,12 +37,23 @@ typedef struct LLSModel {
double coeff[MAX_VARS][MAX_VARS];
double variance[MAX_VARS];
int indep_count;
+/**
+ * Take the outer-product of var[] with itself, and add to the covariance
matrix.
+ * @param m this context
+ * @param var training samples, starting with the value to be predicted
+ */
+void (*update_lls)(struct LLSModel *m, double *var);
+/**
+ * Inner product of var[] and the LPC coefs.
+ * @param m this context
+ * @param var training samples, excluding the value to be predicted
+ * @param order lpc order
+ */
+double (*evaluate_lls)(struct LLSModel *m, double *var, int order);
} LLSModel;
void avpriv_init_lls(LLSModel *m, int indep_count);
-void avpriv_update_lls(LLSModel *m, double *param);
void avpriv_solve_lls(LLSModel *m, double threshold, unsigned short min_order);
-double avpriv_evaluate_lls(LLSModel *m, double *param, int order);
#if FF_API_LLS_PRIVATE
void av_init_lls(LLSModel *m, int indep_count);
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 3/4] x86: lpc: simd av_update_lls
4x-6x faster on sandybridge
---
libavcodec/lpc.c | 4 +-
libavutil/lls.c | 8 +-
libavutil/lls.h | 12 ++-
libavutil/x86/Makefile | 2 +
libavutil/x86/lls.asm| 196 +++
libavutil/x86/lls_init.c | 38 +
6 files changed, 253 insertions(+), 7 deletions(-)
create mode 100644 libavutil/x86/lls.asm
create mode 100644 libavutil/x86/lls_init.c
diff --git a/libavcodec/lpc.c b/libavcodec/lpc.c
index f60976b..144bbed 100644
--- a/libavcodec/lpc.c
+++ b/libavcodec/lpc.c
@@ -200,7 +200,9 @@ int ff_lpc_calc_coefs(LPCContext *s,
ref[i] = fabs(lpc[i][i]);
} else if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
LLSModel m[2];
-double var[MAX_LPC_ORDER+1], av_uninit(weight);
+LOCAL_ALIGNED(32, double, var, [FFALIGN(MAX_LPC_ORDER+1,4)]);
+double av_uninit(weight);
+memset(var, 0, FFALIGN(MAX_LPC_ORDER+1,4)*sizeof(*var));
for(pass=0; passlpc_passes; pass++){
avpriv_init_lls(m[pass1], max_order);
diff --git a/libavutil/lls.c b/libavutil/lls.c
index 5a3e448..f87c2cd 100644
--- a/libavutil/lls.c
+++ b/libavutil/lls.c
@@ -46,8 +46,8 @@ static void update_lls(LLSModel *m, double *var)
void avpriv_solve_lls(LLSModel *m, double threshold, unsigned short min_order)
{
int i, j, k;
-double (*factor)[MAX_VARS + 1] = (void *) m-covariance[1][0];
-double (*covar) [MAX_VARS + 1] = (void *) m-covariance[1][1];
+double (*factor)[MAX_VARS_ALIGN] = (void *) m-covariance[1][0];
+double (*covar) [MAX_VARS_ALIGN] = (void *) m-covariance[1][1];
double *covar_y= m-covariance[0];
int count = m-indep_count;
@@ -117,6 +117,8 @@ av_cold void avpriv_init_lls(LLSModel *m, int indep_count)
m-indep_count = indep_count;
m-update_lls = update_lls;
m-evaluate_lls = evaluate_lls;
+if (ARCH_X86)
+ff_init_lls_x86(m);
}
#if FF_API_LLS_PRIVATE
@@ -154,7 +156,7 @@ int main(void)
avpriv_init_lls(m, 3);
for (i = 0; i 100; i++) {
-double var[4];
+LOCAL_ALIGNED(32, double, var, [4]);
double eval;
var[0] = (av_lfg_get(lfg) / (double) UINT_MAX - 0.5) * 2;
diff --git a/libavutil/lls.h b/libavutil/lls.h
index 8183440..27c0d5e 100644
--- a/libavutil/lls.h
+++ b/libavutil/lls.h
@@ -23,9 +23,12 @@
#ifndef AVUTIL_LLS_H
#define AVUTIL_LLS_H
+#include common.h
+#include mem.h
#include version.h
#define MAX_VARS 32
+#define MAX_VARS_ALIGN FFALIGN(MAX_VARS+1,4)
//FIXME avoid direct access to LLSModel from outside
@@ -33,26 +36,29 @@
* Linear least squares model.
*/
typedef struct LLSModel {
-double covariance[MAX_VARS + 1][MAX_VARS + 1];
-double coeff[MAX_VARS][MAX_VARS];
+DECLARE_ALIGNED(32, double, covariance[MAX_VARS_ALIGN][MAX_VARS_ALIGN]);
+DECLARE_ALIGNED(32, double, coeff[MAX_VARS][MAX_VARS]);
double variance[MAX_VARS];
int indep_count;
/**
* Take the outer-product of var[] with itself, and add to the covariance
matrix.
* @param m this context
* @param var training samples, starting with the value to be predicted
+ *32-byte aligned, and any padding elements must be initialized
+ *(i.e not denormal/nan).
*/
void (*update_lls)(struct LLSModel *m, double *var);
/**
* Inner product of var[] and the LPC coefs.
* @param m this context
- * @param var training samples, excluding the value to be predicted
+ * @param var training samples, excluding the value to be predicted.
unaligned.
* @param order lpc order
*/
double (*evaluate_lls)(struct LLSModel *m, double *var, int order);
} LLSModel;
void avpriv_init_lls(LLSModel *m, int indep_count);
+void ff_init_lls_x86(LLSModel *m);
void avpriv_solve_lls(LLSModel *m, double threshold, unsigned short min_order);
#if FF_API_LLS_PRIVATE
diff --git a/libavutil/x86/Makefile b/libavutil/x86/Makefile
index ae07470..1e19082 100644
--- a/libavutil/x86/Makefile
+++ b/libavutil/x86/Makefile
@@ -1,6 +1,8 @@
OBJS += x86/cpu.o \
x86/float_dsp_init.o\
+x86/lls_init.o \
YASM-OBJS += x86/cpuid.o\
x86/emms.o \
x86/float_dsp.o\
+ x86/lls.o \
diff --git a/libavutil/x86/lls.asm b/libavutil/x86/lls.asm
new file mode 100644
index 000..92c00fc
--- /dev/null
+++ b/libavutil/x86/lls.asm
@@ -0,0 +1,196 @@
+;**
+;* linear least squares model
+;*
+;* Copyright (c) 2013 Loren Merritt
+;*
+;* This file is part of Libav
Re: [libav-devel] [PATCH 2/5] x86: lpc: simd av_evaluate_lls
On Sun, 16 Jun 2013, Ronald S. Bultje wrote: On Sat, Jun 15, 2013 at 5:53 PM, Loren Merritt lor...@u.washington.eduwrote: 1.5x-1.8x faster on sandybridge --- libavutil/lls.c | 3 +++ libavutil/lls.h | 1 + libavutil/x86/lls.asm| 31 +++ libavutil/x86/lls_init.c | 6 +- 4 files changed, 40 insertions(+), 1 deletion(-) diff --git a/libavutil/lls.c b/libavutil/lls.c index eb500af..8f1aff1 100644 --- a/libavutil/lls.c +++ b/libavutil/lls.c @@ -119,6 +119,9 @@ double avpriv_evaluate_lls(LLSModel *m, double *param, int order) int i; double out = 0; +if (m-evaluate_lls) +return m-evaluate_lls(m-coeff[order], param, order); Is there a special reason you didn't assign the default code as default implementation for evaluate_lls (as in: evaluate_lls_c), as is commonly done? No special reason, I just forgot to update those when I realised I didn't have to preserve the api. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/5] x86: lpc: simd av_update_lls
On Sun, 16 Jun 2013, Anton Khirnov wrote: On Sun, 16 Jun 2013 00:53:43 +, Loren Merritt lor...@u.washington.edu wrote: FIXME: This adds alignment constraints and changes a struct that used to be public. External use of LLS is already deprecated, but I guess we can't actually enable this optimization until the next major version bump. I don't think the header was ever installed, so it wasn't really public. I wouldn't bother with compatibility in this case. Ok with me. But what does FF_API_LLS_PRIVATE mean, if not these functions were public, and are now transitioning to private? Why were they ifdeffed rather than removed when someone decided they shouldn't have been exported? --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 2/5] x86: lpc: simd av_evaluate_lls
On Sun, 16 Jun 2013, Jason Garrett-Glaser wrote: +movsd m0, [varq] +movhpd m0, [varq+8] movu isn't faster or at least the same? I'd think it'd be better on Haswell at least, but I'm surprised it hurts on Sandy Bridge (if it does?) movu can store-forwarding stall, since the parent function initializes that array with scalar ops. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 1/5] x86: lpc: simd av_update_lls
/float_dsp.o\ + x86/lls.o \ diff --git a/libavutil/x86/lls.asm b/libavutil/x86/lls.asm new file mode 100644 index 000..f44f7ae --- /dev/null +++ b/libavutil/x86/lls.asm @@ -0,0 +1,184 @@ +;** +;* linear least squares model +;* +;* Copyright (c) 2013 Loren Merritt +;* +;* This file is part of Libav. +;* +;* Libav is free software; you can redistribute it and/or +;* modify it under the terms of the GNU Lesser General Public +;* License as published by the Free Software Foundation; either +;* version 2.1 of the License, or (at your option) any later version. +;* +;* Libav is distributed in the hope that it will be useful, +;* but WITHOUT ANY WARRANTY; without even the implied warranty of +;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU +;* Lesser General Public License for more details. +;* +;* You should have received a copy of the GNU Lesser General Public +;* License along with Libav; if not, write to the Free Software +;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA +;** + +%include x86util.asm + +SECTION .text + +%define MAX_VARS 32 +%define COVAR_STRIDE (MAX_VARS+4)*8 +%define COVAR(x,y) [covarq+(x)*8+y*COVAR_STRIDE] + +%macro ADDPD_MEM 2 +%if cpuflag(avx) +vaddpd %2, %1 +%else +addpd %2, %1 +%endif +mova %1, %2 +%endmacro + +INIT_XMM sse2 +%define movdqa movaps +cglobal update_lls, 3,5,8, covar, var, i, j, covar2 +lea varq, [varq+iq*8] +neg iq +mov covar2q, covarq +.loopi: +; Compute all 3 pairwise products of a 2x2 block that lies on the diagonal +movam1, [varq+iq*8] +movam3, [varq+iq*8+16] +pshufd m4, m1, q1010 +pshufd m5, m1, q3232 +pshufd m6, m3, q1010 +pshufd m7, m3, q3232 +mulpd m0, m1, m4 +mulpd m1, m1, m5 +lea covarq, [covar2q+16] +ADDPD_MEM COVAR(-2,0), m0 +ADDPD_MEM COVAR(-2,1), m1 +lea jq, [iq+2] +cmp jd, -2 +jg .skip4x4 +.loop4x4: +; Compute all 16 pairwise products of a 4x4 block +mulpd m0, m4, m3 +mulpd m1, m5, m3 +mulpd m2, m6, m3 +mulpd m3, m3, m7 +ADDPD_MEM COVAR(0,0), m0 +ADDPD_MEM COVAR(0,1), m1 +ADDPD_MEM COVAR(0,2), m2 +ADDPD_MEM COVAR(0,3), m3 +movam3, [varq+jq*8+16] +mulpd m0, m4, m3 +mulpd m1, m5, m3 +mulpd m2, m6, m3 +mulpd m3, m3, m7 +ADDPD_MEM COVAR(2,0), m0 +ADDPD_MEM COVAR(2,1), m1 +ADDPD_MEM COVAR(2,2), m2 +ADDPD_MEM COVAR(2,3), m3 +movam3, [varq+jq*8+32] +add covarq, 32 +add jq, 4 +cmp jd, -2 +jle .loop4x4 +.skip4x4: +testjd, jd +jg .skip2x4 +mulpd m4, m3 +mulpd m5, m3 +mulpd m6, m3 +mulpd m7, m3 +ADDPD_MEM COVAR(0,0), m4 +ADDPD_MEM COVAR(0,1), m5 +ADDPD_MEM COVAR(0,2), m6 +ADDPD_MEM COVAR(0,3), m7 +.skip2x4: +add iq, 4 +add covar2q, 4*COVAR_STRIDE+32 +cmp id, -2 +jle .loopi +testid, id +jg .ret +mov jq, iq +%define covarq covar2q +.loop2x1: +movsd m0, [varq+iq*8] +movlhps m0, m0 +mulpd m0, [varq+jq*8] +ADDPD_MEM COVAR(0,0), m0 +inc iq +add covarq, COVAR_STRIDE +testid, id +jle .loop2x1 +.ret: +REP_RET + +INIT_YMM avx +cglobal update_lls, 3,6,8, covar, var, count, i, j, count2 +lea count2d, [countq-2] +xor id, id +.loopi: +; Compute all 10 pairwise products of a 4x4 block that lies on the diagonal +movaymm1, [varq+iq*8] +vbroadcastsd ymm4, [varq+iq*8] +vbroadcastsd ymm5, [varq+iq*8+8] +vbroadcastsd ymm6, [varq+iq*8+16] +vbroadcastsd ymm7, [varq+iq*8+24] +vextractf128 xmm3, ymm1, 1 +vmulpd ymm0, ymm1, ymm4 +vmulpd ymm1, ymm1, ymm5 +vmulpd xmm2, xmm3, xmm6 +vmulpd xmm3, xmm3, xmm7 +ADDPD_MEM COVAR(iq ,0), ymm0 +ADDPD_MEM COVAR(iq ,1), ymm1 +ADDPD_MEM COVAR(iq+2,2), xmm2 +ADDPD_MEM COVAR(iq+2,3), xmm3 +lea jd, [iq+4] +cmp jd, count2d +jg .skip4x4 +.loop4x4: +; Compute all 16 pairwise products of a 4x4 block +movaymm3, [varq+jq*8] +vmulpd ymm0, ymm3, ymm4 +vmulpd ymm1, ymm3, ymm5 +vmulpd ymm2, ymm3, ymm6 +vmulpd ymm3, ymm3, ymm7 +ADDPD_MEM COVAR(jq,0), ymm0 +ADDPD_MEM COVAR(jq,1), ymm1 +ADDPD_MEM COVAR(jq,2), ymm2 +ADDPD_MEM COVAR(jq,3), ymm3 +add jd, 4 +cmp jd, count2d +jle .loop4x4 +.skip4x4: +cmp jd, countd +jg .skip2x4 +movaxmm3, [varq+jq*8] +vmulpd xmm0, xmm3, xmm4 +vmulpd xmm1, xmm3, xmm5 +vmulpd xmm2, xmm3, xmm6 +vmulpd xmm3, xmm3, xmm7 +ADDPD_MEM COVAR(jq,0), xmm0 +ADDPD_MEM COVAR(jq,1), xmm1 +ADDPD_MEM COVAR(jq,2), xmm2 +ADDPD_MEM COVAR(jq,3), xmm3 +.skip2x4
[libav-devel] [PATCH 4/5] x86: lpc: optimize lpc_compute_autocorr_sse2
4% faster on sandybridge --- libavcodec/x86/lpc.c | 10 ++ 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/libavcodec/x86/lpc.c b/libavcodec/x86/lpc.c index 7a37b88..902d364 100644 --- a/libavcodec/x86/lpc.c +++ b/libavcodec/x86/lpc.c @@ -124,11 +124,13 @@ static void lpc_compute_autocorr_sse2(const double *data, int len, int lag, movsdMANGLE(pd_1), %%xmm1\n\t 1: \n\t movapd (%3,%0), %%xmm3 \n\t -movupd -8(%4,%0), %%xmm4 \n\t +movapd (%4,%0), %%xmm4 \n\t +movsd -8(%4,%0), %%xmm5 \n\t +movlhps %%xmm4, %%xmm5 \n\t mulpd %%xmm3, %%xmm4 \n\t -mulpd(%4,%0), %%xmm3 \n\t -addpd %%xmm4, %%xmm1 \n\t -addpd %%xmm3, %%xmm0 \n\t +mulpd %%xmm3, %%xmm5 \n\t +addpd %%xmm4, %%xmm0 \n\t +addpd %%xmm5, %%xmm1 \n\t add $16,%0 \n\t jl 1b \n\t movhlps %%xmm0, %%xmm3 \n\t -- 1.8.1.5 ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 2/5] x86: lpc: simd av_evaluate_lls
1.5x-1.8x faster on sandybridge
---
libavutil/lls.c | 3 +++
libavutil/lls.h | 1 +
libavutil/x86/lls.asm| 31 +++
libavutil/x86/lls_init.c | 6 +-
4 files changed, 40 insertions(+), 1 deletion(-)
diff --git a/libavutil/lls.c b/libavutil/lls.c
index eb500af..8f1aff1 100644
--- a/libavutil/lls.c
+++ b/libavutil/lls.c
@@ -119,6 +119,9 @@ double avpriv_evaluate_lls(LLSModel *m, double *param, int
order)
int i;
double out = 0;
+if (m-evaluate_lls)
+return m-evaluate_lls(m-coeff[order], param, order);
+
for (i = 0; i = order; i++)
out += param[i] * m-coeff[order][i];
diff --git a/libavutil/lls.h b/libavutil/lls.h
index 76ff10c..8a4d318 100644
--- a/libavutil/lls.h
+++ b/libavutil/lls.h
@@ -39,6 +39,7 @@ typedef struct LLSModel {
double variance[MAX_VARS];
int indep_count;
void (*update_lls)(struct LLSModel *m, double *var, int order);
+double (*evaluate_lls)(double *coefs, double *var, int order);
} LLSModel;
void avpriv_init_lls(LLSModel *m, int indep_count);
diff --git a/libavutil/x86/lls.asm b/libavutil/x86/lls.asm
index f44f7ae..b5e04d9 100644
--- a/libavutil/x86/lls.asm
+++ b/libavutil/x86/lls.asm
@@ -182,3 +182,34 @@ cglobal update_lls, 3,6,8, covar, var, count, i, j, count2
jle .loop2x1
.ret:
REP_RET
+
+
+INIT_XMM sse2
+cglobal evaluate_lls, 3,3,3, coefs, var, i
+; This function is often called on the same buffer as update_lls, but with
an offset. They can't both be aligned.
+movsd m0, [varq]
+movhpd m0, [varq+8]
+mulpd m0, [coefsq]
+lea coefsq, [coefsq+iq*8]
+lea varq, [varq+iq*8]
+neg iq
+add iq, 2
+.loop:
+movsd m1, [varq+iq*8]
+movhpd m1, [varq+iq*8+8]
+mulpd m1, [coefsq+iq*8]
+addpd m0, m1
+add iq, 2
+jl .loop
+jg .skip1
+movsd m1, [varq+iq*8]
+mulsd m1, [coefsq+iq*8]
+addpd m0, m1
+.skip1:
+movhlps m1, m0
+addsd m0, m1
+%if ARCH_X86_32
+movsd r0m, m0
+fld qword r0m
+%endif
+RET
diff --git a/libavutil/x86/lls_init.c b/libavutil/x86/lls_init.c
index 1215b14..d65ddc8 100644
--- a/libavutil/x86/lls_init.c
+++ b/libavutil/x86/lls_init.c
@@ -25,12 +25,16 @@
void ff_update_lls_sse2(LLSModel *m, double *var, int order);
void ff_update_lls_avx(LLSModel *m, double *var, int order);
+double ff_evaluate_lls_sse2(double *coefs, double *var, int order);
void avpriv_init_lls_x86(LLSModel *m)
{
int cpu_flags = av_get_cpu_flags();
-if (EXTERNAL_SSE2(cpu_flags))
+if (EXTERNAL_SSE2(cpu_flags)) {
m-update_lls = ff_update_lls_sse2;
+if (m-indep_count = 4)
+m-evaluate_lls = ff_evaluate_lls_sse2;
+}
if (EXTERNAL_AVX(cpu_flags))
m-update_lls = ff_update_lls_avx;
}
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 3/5] x86: lpc: optimize av_solve_lls
1.2x-1.4x faster on sandybridge
---
libavutil/lls.c | 8
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/libavutil/lls.c b/libavutil/lls.c
index 8f1aff1..bcaa901 100644
--- a/libavutil/lls.c
+++ b/libavutil/lls.c
@@ -75,9 +75,9 @@ void avpriv_solve_lls(LLSModel *m, double threshold, unsigned
short min_order)
if (i == j) {
if (sum threshold)
sum = 1.0;
-factor[i][i] = sqrt(sum);
+factor[i][i] = 1.0 / sqrt(sum);
} else {
-factor[j][i] = sum / factor[i][i];
+factor[j][i] = sum * factor[i][i];
}
}
}
@@ -88,7 +88,7 @@ void avpriv_solve_lls(LLSModel *m, double threshold, unsigned
short min_order)
for (k = i - 1; k = 0; k--)
sum -= factor[i][k] * m-coeff[0][k];
-m-coeff[0][i] = sum / factor[i][i];
+m-coeff[0][i] = sum * factor[i][i];
}
for (j = count - 1; j = min_order; j--) {
@@ -98,7 +98,7 @@ void avpriv_solve_lls(LLSModel *m, double threshold, unsigned
short min_order)
for (k = i + 1; k = j; k++)
sum -= factor[k][i] * m-coeff[j][k];
-m-coeff[j][i] = sum / factor[i][i];
+m-coeff[j][i] = sum * factor[i][i];
}
m-variance[j] = covar_y[0];
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 5/5] lpc: use levinson for the first pass of multipass cholesky
Levinson is faster, and cholesky is only needed if we want to apply different
weights to different samples, which doesn't happen on the first pass.
---
libavcodec/lpc.c | 29 -
1 file changed, 16 insertions(+), 13 deletions(-)
diff --git a/libavcodec/lpc.c b/libavcodec/lpc.c
index 7247a76..58c7585 100644
--- a/libavcodec/lpc.c
+++ b/libavcodec/lpc.c
@@ -176,7 +176,7 @@ int ff_lpc_calc_coefs(LPCContext *s,
double autoc[MAX_LPC_ORDER+1];
double ref[MAX_LPC_ORDER];
double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
-int i, j, pass;
+int i, j, pass = 0;
int opt_order;
assert(max_order = MIN_LPC_ORDER max_order = MAX_LPC_ORDER
@@ -189,7 +189,7 @@ int ff_lpc_calc_coefs(LPCContext *s,
ff_lpc_init(s, blocksize, max_order, lpc_type);
}
-if (lpc_type == FF_LPC_TYPE_LEVINSON) {
+if (lpc_type == FF_LPC_TYPE_LEVINSON || (lpc_type == FF_LPC_TYPE_CHOLESKY
lpc_passes 1)) {
s-lpc_apply_welch_window(samples, blocksize, s-windowed_samples);
s-lpc_compute_autocorr(s-windowed_samples, blocksize, max_order,
autoc);
@@ -198,13 +198,20 @@ int ff_lpc_calc_coefs(LPCContext *s,
for(i=0; imax_order; i++)
ref[i] = fabs(lpc[i][i]);
-} else if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
+
+pass++;
+}
+
+if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
LLSModel m[2];
LOCAL_ALIGNED(32, double, var, [FFALIGN(MAX_LPC_ORDER+1,4)]);
double av_uninit(weight);
memset(var, 0, FFALIGN(MAX_LPC_ORDER+1,4)*sizeof(*var));
-for(pass=0; passlpc_passes; pass++){
+for(j=0; jmax_order; j++)
+m[0].coeff[max_order-1][j] = -lpc[max_order-1][j];
+
+for(; passlpc_passes; pass++){
avpriv_init_lls(m[pass1], max_order);
weight=0;
@@ -259,15 +266,11 @@ av_cold int ff_lpc_init(LPCContext *s, int blocksize, int
max_order,
s-max_order = max_order;
s-lpc_type = lpc_type;
-if (lpc_type == FF_LPC_TYPE_LEVINSON) {
-s-windowed_buffer = av_mallocz((blocksize + 2 + FFALIGN(max_order,
4)) *
-sizeof(*s-windowed_samples));
-if (!s-windowed_buffer)
-return AVERROR(ENOMEM);
-s-windowed_samples = s-windowed_buffer + FFALIGN(max_order, 4);
-} else {
-s-windowed_samples = NULL;
-}
+s-windowed_buffer = av_mallocz((blocksize + 2 + FFALIGN(max_order, 4)) *
+sizeof(*s-windowed_samples));
+if (!s-windowed_buffer)
+return AVERROR(ENOMEM);
+s-windowed_samples = s-windowed_buffer + FFALIGN(max_order, 4);
s-lpc_apply_welch_window = lpc_apply_welch_window_c;
s-lpc_compute_autocorr = lpc_compute_autocorr_c;
--
1.8.1.5
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [libav-api] High CPU load of the avcodec_decode_video2() function
On Tue, 26 Mar 2013, Ond?ej Perutka wrote: Hello again, I think I found the problem (correct me if I'm wrong :-)). I checked all functions with AVX 256 instructions and there is really no such function which is called from my application and does not have the vzeroupper instruction. After a few days of disassembling I've discovered there is quite a lot of functions which mix both SSE and AVX128 instructions. The ff_deblock_v_luma_8_avx, I mentioned before, is one of them. It is not a problem unless there is a non-zero in the upper half of YMM registers or the registers are in the state C (using Agner's terminology). In such case the registers would oscillate between states B and C and both of these transitions are time consuming. So here is my theory: Somewhere in the JVM or its dependent libraries is an AVX 256 code with missing vzeroupper instruction. This code leaves YMM registers in the state B. After that there could be also some SSE code which would leave YMM registers in the state C. Everything is fine until the ff_deblock_v_luma_8_avx (or similar function) is called. After that the registers oscillate between state B and C as the ff_deblock_v_luma_8_avx transits between SSE and AVX 128 instructions. It also corresponds to my previous profiling results. Knowing all this I tried to solve it. I added following macro into the libavutil/x86inc.asm: %macro vzeroupper 0 %if avx_enabled vzeroupper %endif %endmacro and I also updated all macros used after INIT_XMM avx in all files so there is vzeroupper after each cglobal in these macros. Everything works Wrong solution. vzeroupper is not free, so don't just sprinkle it everywhere. You should either fix JVM, or put a single vzeroupper on entry to the whole libav. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/1] x86: consistently use unaligned movs in the unaligned bswap
On Sun, 24 Mar 2013, Janne Grunau wrote: Fixes fate errors in asv1, ffvhuff and huffyuv on x86_32. --- libavcodec/x86/dsputil.asm | 18 +- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/libavcodec/x86/dsputil.asm b/libavcodec/x86/dsputil.asm index a1aaaf5..686cec8 100644 --- a/libavcodec/x86/dsputil.asm +++ b/libavcodec/x86/dsputil.asm @@ -552,23 +552,23 @@ VECTOR_CLIP_INT32 6, 1, 0, 0 %if cpuflag(ssse3) pshufb m0, m2 pshufb m1, m2 -mova [r0 + 0], m0 -mova [r0 + 16], m1 +mov%1[r0 + 0], m0 +mov%1[r0 + 16], m1 %else pshuflw m0, m0, 10110001b pshuflw m1, m1, 10110001b pshufhw m0, m0, 10110001b pshufhw m1, m1, 10110001b -mova m2, m0 -mova m3, m1 +mov%1m2, m0 +mov%1m3, m1 register-to-register move is always aligned. psllwm0, 8 psllwm1, 8 psrlwm2, 8 psrlwm3, 8 por m2, m0 por m3, m1 -mova [r0 + 0], m2 -mova [r0 + 16], m3 +mov%1[r0 + 0], m2 +mov%1[r0 + 16], m3 %endif add r0, 32 add r1, 32 @@ -581,15 +581,15 @@ VECTOR_CLIP_INT32 6, 1, 0, 0 mov%1m0, [r1] %if cpuflag(ssse3) pshufb m0, m2 -mova [r0], m0 +mov%1[r0], m0 %else pshuflw m0, m0, 10110001b pshufhw m0, m0, 10110001b -mova m2, m0 +mov%1m2, m0 same here. psllwm0, 8 psrlwm2, 8 por m2, m0 -mova [r0], m2 +mov%1[r0], m2 %endif add r1, 16 add r0, 16 --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [libav-api] High CPU load of the avcodec_decode_video2() function
On Sun, 17 Mar 2013, Jason Garrett-Glaser wrote: On Sun, Mar 17, 2013 at 5:48 PM, Ondrej Perutka perutka.ond...@gmail.com wrote: ... and according to the Intel reference: http://software.intel.com/sites/default/files/319433-014.pdf the VZEROUPPER or the VZEROALL instruction should be executed before and after any usage of VEX-encoded instructions (i.e. 256-bit AVX). It means both transitions SSE - AVX and AVX - SSE (not only before return from AVX function). See section 2.8.1. That's not what it says, and is not correct. x264 does not use vzeroupper before any of its AVX functions and the official Intel emulator records no false SSE/AVX transition problems. And if you (Ondrej) want to know how a both transitions can have performance penalties (as the Intel doc says), while still needing only a single vzeroupper to fix both, then see Agner's optimization guide http://agner.org/optimize/optimizing_assembly.pdf section 13.6. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] hpeldsp: x86: Convert dsputil_rnd_template to yasm
On Mon, 18 Feb 2013, Daniel Kang wrote: +%macro PAVGBP_MMX 6 +mova %3, %1 +mova %6, %4 +por%3, %2 +por%6, %5 +pxor %2, %1 +pxor %5, %4 +pand %2, m6 +pand %5, m6 +psrlq %2, 1 +psrlq %5, 1 +psubb %3, %2 +psubb %6, %5 +%endmacro + +%macro PAVGB_NRND_OP_MMX 4 +mova %3, %1 +pand %3, %2 +pxor %2, %1 +pand %2, %4 +psrlq %2, 1 +paddb %3, %2 +%endmacro + +%macro PAVGBP_NO_RND_MMX 6 +PAVGB_NRND_OP_MMX %1, %2, %3, m6 +PAVGB_NRND_OP_MMX %4, %5, %6, m6 +%endmacro + +%macro PAVGB_OP_MMX 4 +mova %3, %1 +por %3, %2 +pxor %2, %1 +pand %2, %4 +psrlq%2, 1 +psubb%3, %2 +%endmacro I meant eliminate PAVGBP_MMX and PAVGBP_NO_RND_MMX entirely, and instead call PAVGB_OP_MMX or PAVGB_NRND_OP_MMX twice from the functions that used to use them. +; put_pixels8_x2(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h) +%macro PUT_PIXELS8_X2_MMX 0-1 +cglobal put%1_pixels8_x2, 4,4 +pcmpeqd m6, m6 +paddbm6, m6 +.loop: +mova m0, [r1] +mova m1, [r1+1] +mova m2, [r1+r2] +mova m3, [r1+r2+1] +PAVGBP m0, m1, m4, m2, m3, m5 +mova [r0], m4 +mova [r0+r2*1], m5 +lea r1, [r1+r2*2] +lea r0, [r0+r2*2] +mova m0, [r1] +mova m1, [r1+1] +mova m2, [r1+r2] +mova m3, [r1+r2+1] +PAVGBP m0, m1, m4, m2, m3, m5 +mova [r0], m4 +mova [r0+r2*1], m5 +lea r1, [r1+r2*2] +lea r0, [r0+r2*2] +sub r3d, 4 +jne .loop +RET +%endmacro %rep. That said, I would have guessed that the original purpose of the unrolling in most of the functions in this patch was to allow manual register renaming to eliminate a few moves. In which case the functions where %rep works are precisely the ones that shouldn't have been unrolled. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] hpeldsp: x86: Convert dsputil_rnd_template to yasm
On Fri, 15 Feb 2013, Daniel Kang wrote: +%macro PAVGBP_MMX 6 +mova %3, %1 +mova %6, %4 +por%3, %2 +por%6, %5 +pxor %2, %1 +pxor %5, %4 +pand %2, m6 +pand %5, m6 +psrlq %2, 1 +psrlq %5, 1 +psubb %3, %2 +psubb %6, %5 +%endmacro + +%macro PAVGBP_NO_RND_MMX 6 +mova %3, %1 +mova %6, %4 +pand %3, %2 +pand %6, %5 +pxor %2, %1 +pxor %5, %4 +pand %2, m6 +pand %5, m6 +psrlq%2, 1 +psrlq%5, 1 +paddb%3, %2 +paddb%6, %5 +%endmacro Does this need to be interleaved, not just two calls to PAVGB_OP_MMX? +; put_pixels8_xy2(uint8_t *block, const uint8_t *pixels, int line_size, int h) +%macro PUT_PIXELS8_XY2_MMX 0-1 +cglobal put%1_pixels8_xy2, 4,5 +pxor m7, m7 +SET_RND(m6) +mova m0, [r1] +mova m4, [r1+1] +mova m1, m0 +mova m5, m4 +punpcklbwm0, m7 +punpcklbwm4, m7 +punpckhbwm1, m7 +punpckhbwm5, m7 +paddusw m4, m0 +paddusw m5, m1 +xor r4, r4 +add r1, r2 +.loop: +mova m0, [r1+r4] +mova m2, [r1+r4+1] +mova m1, m0 +mova m3, m2 +punpcklbwm0, m7 +punpcklbwm2, m7 +punpckhbwm1, m7 +punpckhbwm3, m7 +paddusw m0, m2 +paddusw m1, m3 +paddusw m4, m6 +paddusw m5, m6 +paddusw m4, m0 +paddusw m5, m1 +psrlwm4, 2 +psrlwm5, 2 +packuswb m4, m5 +mova[r0+r4], m4 +add r4, r2 +mova m2, [r1+r4] +mova m3, [r1+r4+1] +mova m3, m2 +mova m5, m4 +punpcklbwm2, m7 +punpcklbwm4, m7 +punpckhbwm3, m7 +punpckhbwm5, m7 +paddusw m4, m2 +paddusw m5, m3 +paddusw m0, m6 +paddusw m1, m6 +paddusw m0, m4 +paddusw m1, m5 +psrlwm0, 2 +psrlwm1, 2 +packuswb m0, m1 +mova[r0+r4], m0 +add r4, r2 +sub r3d, 2 +jne .loop +RET +%endmacro Does this and similar functions really need to be unrolled? If so, use %rep. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] hpeldsp: x86: Convert dsputil_rnd_template to yasm
On Thu, 14 Feb 2013, Ronald S. Bultje wrote:
On Feb 14, 2013 4:59 AM, Diego Biurrun di...@biurrun.de wrote:
On Wed, Feb 13, 2013 at 05:53:36PM -0500, Daniel Kang wrote:
@@ -56,6 +107,44 @@ PUT_PIXELS8_X2
+%macro PUT_PIXELS8_X2_MMX 0-1
+%if %0 == 1
+cglobal put%1_pixels8_x2, 4,4
+%else
+cglobal put_pixels8_x2, 4,4
+%endif
IIRC you don't need the %if, but you can just pass an empty
first parameter and it should do the right thing.
.. more below ..
MACRO 0-1 sets an empty string by default.
True, but strings aren't good for much in asm.
MACRO 0-1 sets an empty *token* by default, which is what you want if
you're going to concatenate it into a function name.
MACRO 0-1 {} works too if you want to explicitly specify the empty token.
--Loren Merritt
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] dsputil: x86: Convert h263 loop filter to yasm
On Tue, 5 Feb 2013, Luca Barbato wrote: On 02/02/13 00:28, Daniel Kang wrote: --- I am very skeptical when assembly works on the first time. More testing would be appreciated. --- libavcodec/x86/dsputil.asm | 159 libavcodec/x86/dsputil_mmx.c | 185 ++ 2 files changed, 167 insertions(+), 177 deletions(-) /usr/lib/gcc/x86_64-pc-linux-gnu/4.6.3/../../../../x86_64-pc-linux-gnu/bin/ld: libavcodec/libavcodec.a(dsputil.o): relocation R_X86_64_32 against `ff_h263_loop_filter_strength' can not be used when making a shared object; recompile with -fPIC libavcodec/libavcodec.a: could not read symbols: Bad value Looks like something is broken for x86_64. The offending line is movzx r3d, BYTE [ff_h263_loop_filter_strength+r2] since you can't have both a register offset and a PIC offset in the same address. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] vorbisdsp: convert x86 simd functions from inline asm to yasm.
On Mon, 21 Jan 2013, Ronald S. Bultje wrote: From: Ronald S. Bultje rsbul...@gmail.com --- libavcodec/x86/Makefile | 1 + libavcodec/x86/dsputil_mmx.c| 3 -- libavcodec/x86/dsputil_mmx.h| 2 - libavcodec/x86/vorbisdsp.asm| 83 + libavcodec/x86/vorbisdsp_init.c | 77 -- 5 files changed, 92 insertions(+), 74 deletions(-) create mode 100644 libavcodec/x86/vorbisdsp.asm LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH v2] dsputilenc: x86: Convert pixel inline asm to yasm
On Wed, 16 Jan 2013, Daniel Kang wrote: --- Fixed movu - mova comment from Loren --- libavcodec/x86/dsputilenc.asm | 152 + libavcodec/x86/dsputilenc_mmx.c | 201 --- 2 files changed, 172 insertions(+), 181 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] dsputilenc: x86: Convert pixel inline asm to yasm
On Mon, 14 Jan 2013, Daniel Kang wrote: +INIT_MMX mmx +; get_pixels_mmx(DCTELEM *block, const uint8_t *pixels, int line_size) +cglobal get_pixels, 3,4 +movsxdifnidn r2, r2d +add r0, 128 +mov r3, -128 +pxor m7, m7 +.loop: +movu m0, [r1] +movu m2, [r1+r2] +movu m1, m0 +movu m3, m2 +punpcklbwm0, m7 +punpckhbwm1, m7 +punpcklbwm2, m7 +punpckhbwm3, m7 +movu [r0+r3+ 0], m0 +movu [r0+r3+ 8], m1 +movu [r0+r3+16], m2 +movu [r0+r3+24], m3 +lea r1, [r1+r2*2] +add r3, 32 +js .loop +REP_RET Is this really unaligned? --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: Add cvisible macro for C functions with public prefix
On Sun, 13 Jan 2013, Diego Biurrun wrote: --- Now using local variables for HIDDEN and FUNCTION_PREFIX, as suggested by Loren. libavutil/x86/x86inc.asm | 40 libavutil/x86/x86util.asm |1 + 2 files changed, 29 insertions(+), 12 deletions(-) Patch OK. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] yadif: Fix bug with x86_32 MSVC
On Sat, 12 Jan 2013, Daniel Kang wrote: -mova [rsp+ 0], m0 -mova [rsp+16], m3 -mova [rsp+32], m1 +mova [rsp- 0], m0 +mova [rsp-16], m3 +mova [rsp-32], m1 You can't do that on x86_32. +%assign PAD -1*80 Unused? %macro YADIF 0 -cglobal yadif_filter_line, 7, 7, 8, 16*5, dst, prev, cur, next, w, prefs, \ - mrefs, parity, mode -test wq, wq +cglobal yadif_filter_line, 4, 7, 8, PAD Do you have a reason for removing all the named args? +cmp DWORD r4m, 0 jle .ret -movsxdifnidn prefsq, prefsd -movsxdifnidn mrefsq, mrefsd +%if ARCH_X86_32 +movifnidn r4, r5mp +movifnidn r5, r6mp +DECLARE_REG_TMP 4,5 +%else +movsxdifnidn r5, DWORD r5m +movsxdifnidn r6, DWORD r6m +DECLARE_REG_TMP 5,6 +%endif No ifnidn. After your change, they will not in fact be identical. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] yadif: Fix bug with x86_32 MSVC
On Sat, 12 Jan 2013, Daniel Kang wrote: On Sat, Jan 12, 2013 at 10:36 PM, Loren Merritt lor...@u.washington.edu wrote: Do you have a reason for removing all the named args? I can't use half of the named args, and I thought it was less confusing if I didn't use them at all. Half? The only ones you need to avoid are the 2 args you're moving. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 2/3] x86inc: Add cvisible macro for C functions with public prefix
On Wed, 9 Jan 2013, Diego Biurrun wrote: +%ifidn __OUTPUT_FORMAT__,elf HIDDEN That compares __OUTPUT_FORMAT__ against the string elf HIDDEN. can only be used in %if, not %ifidn or any of the other conditionals. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH v4] yadif: Port inline assembly to YASM
On Tue, 8 Jan 2013, Daniel Kang wrote: +cglobal yadif_filter_line, 9, 9, 8, 16*5, dst, prev, cur, next, w, prefs, \ + mrefs, parity, mode If you never use parity and mode as regs, then reduce register allocation. +test wq, wq +jle .ret +movsxdifnidn prefsq, prefsd +movsxdifnidn mrefsq, mrefsd + +cmp DWORD paritym, 0 +je .parity0 +FILTER 1, prevq, curq +jmp .ret + +.parity0: +FILTER 0, curq, nextq + +.ret: +RET --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH] x86inc: activate REP_RET automatically
Now RET checks whether it immediately follows a branch, so the programmer
dosen't have to keep track of that condition.
REP_RET is still needed manually when it's a branch target, but that's much
rarer.
The implementation involves lots of spurious labels, so get rid of them with
`strip`.
---
configure | 3 +++
library.mak | 1 +
libavutil/x86/x86inc.asm| 36 +++
libavcodec/x86/ac3dsp.asm | 6 ++---
libavcodec/x86/deinterlace.asm | 2 +-
libavcodec/x86/dsputil.asm | 22 -
libavcodec/x86/fft.asm | 8 +++---
libavcodec/x86/fmtconvert.asm | 12 -
libavcodec/x86/h264_chromamc.asm| 18 +++---
libavcodec/x86/h264_chromamc_10bit.asm | 10
libavcodec/x86/h264_deblock.asm | 2 +-
libavcodec/x86/h264_deblock_10bit.asm | 10
libavcodec/x86/h264_idct.asm| 16 ++--
libavcodec/x86/h264_intrapred.asm | 28 ++---
libavcodec/x86/h264_intrapred_10bit.asm | 16 ++--
libavcodec/x86/h264_qpel_10bit.asm | 2 +-
libavcodec/x86/h264_qpel_8bit.asm | 20 +++
libavcodec/x86/h264_weight.asm | 16 ++--
libavcodec/x86/h264_weight_10bit.asm| 12 -
libavcodec/x86/pngdsp.asm | 2 +-
libavcodec/x86/rv34dsp.asm | 2 +-
libavcodec/x86/rv40dsp.asm | 10
libavcodec/x86/sbrdsp.asm | 2 +-
libavcodec/x86/vp8dsp.asm | 40 +++---
libavfilter/x86/af_volume.asm | 6 ++---
libavfilter/x86/hqdn3d.asm | 2 +-
libavresample/x86/audio_convert.asm | 44 -
libavresample/x86/audio_mix.asm | 10
libavutil/x86/float_dsp.asm | 8 +++---
libswscale/x86/input.asm| 14 +--
libswscale/x86/output.asm | 8 +++---
libswscale/x86/scale.asm| 2 +-
32 files changed, 210 insertions(+), 180 deletions(-)
diff --git a/configure b/configure
index 08f1d82..ac056b3 100755
--- a/configure
+++ b/configure
@@ -1802,6 +1802,7 @@ nm_default=nm -g
objformat=elf
pkg_config_default=pkg-config
ranlib=ranlib
+strip=strip
yasmexe=yasm
nogas=:
@@ -2051,6 +2052,7 @@ cc_default=${cross_prefix}${cc_default}
nm_default=${cross_prefix}${nm_default}
pkg_config_default=${cross_prefix}${pkg_config_default}
ranlib=${cross_prefix}${ranlib}
+strip=${cross_prefix}${strip}
sysinclude_default=${sysroot}/usr/include
@@ -3842,6 +3844,7 @@ ARCH=$arch
CC=$cc
AS=$as
LD=$ld
+STRIP=$strip
DEPCC=$dep_cc
DEPCCFLAGS=$DEPCCFLAGS \$(CPPFLAGS)
DEPAS=$as
diff --git a/library.mak b/library.mak
index 3b4bd2d..23ee5fa 100644
--- a/library.mak
+++ b/library.mak
@@ -26,6 +26,7 @@ $(SUBDIR)%-test.i: $(SUBDIR)%.c
$(SUBDIR)x86/%.o: $(SUBDIR)x86/%.asm
$(DEPYASM) $(YASMFLAGS) -I $(D)/ -M -o $@ $ $(@:.o=.d)
$(YASM) $(YASMFLAGS) -I $(D)/ -o $@ $
+ -@ $(if $(STRIP), $(STRIP) -wN '..@*' $@)
LIBOBJS := $(OBJS) $(SUBDIR)%.h.o $(TESTOBJS)
$(LIBOBJS) $(LIBOBJS:.o=.i): CPPFLAGS += -DHAVE_AV_CONFIG_H
diff --git a/libavutil/x86/x86inc.asm b/libavutil/x86/x86inc.asm
index 2617cdf..5594db4 100644
--- a/libavutil/x86/x86inc.asm
+++ b/libavutil/x86/x86inc.asm
@@ -131,8 +131,7 @@ CPUNOP amdnop
; Pops anything that was pushed by PROLOGUE, and returns.
; REP_RET:
-; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
-; which are slow when a normal ret follows a branch.
+; Use this instead of RET if it's a branch target.
; registers:
; rN and rNq are the native-size register holding function argument N
@@ -480,7 +479,7 @@ DECLARE_REG 14, R15, 120
%if mmsize == 32
vzeroupper
%endif
-ret
+AUTO_REP_RET
%endmacro
%elif ARCH_X86_64 ; *nix x64 ;=
@@ -527,7 +526,7 @@ DECLARE_REG 14, R15, 72
%if mmsize == 32
vzeroupper
%endif
-ret
+AUTO_REP_RET
%endmacro
%else ; X86_32 ;==
@@ -583,7 +582,7 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
%if mmsize == 32
vzeroupper
%endif
-ret
+AUTO_REP_RET
%endmacro
%endif ;==
@@ -597,6 +596,10 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
%endmacro
%endif
+; On AMD cpus =K10, an ordinary ret is slow if it immediately follows either
+; a branch or a branch target. So switch to a 2-byte form of ret in that case.
+; We can automatically detect follows a branch, but not a branch target.
+; (SSSE3 is a sufficient condition to know that your cpu doesn't have this
problem.)
%macro REP_RET 0
%if has_epilogue
RET
@@ -605,6 +608,29 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
%endif
%endmacro
+%define last_branch_adr $$
Re: [libav-devel] [PATCH 2/4] lavr: x86: add SSE2 and FMA4 quantization for non-filtering dither methods
On Sat, 15 Dec 2012, Justin Ruggles wrote: +;-- +; void ff_quantize_int_rectangular/triangular(int16_t *dst, float *src, +; int *dither, int len); +;-- + +%macro QUANTIZE_INT 1 +cglobal quantize_int_%1, 4,5,6, dst, src, dither0, len, dither1 +lea lenq, [2*lend] +add dstq, lenq +lea srcq, [srcq+2*lenq] +%ifidn %1, triangular +lea dither1q, [dither0q+4*lenq] +%endif +lea dither0q, [dither0q+2*lenq] +neg lenq +mova m4, [pf_dither_scale] +mova m5, [pf_s16_scale] +.loop: +cvtdq2ps m0, [dither0q+2*lenq] +cvtdq2ps m1, [dither0q+2*lenq+mmsize] Do you have to store dither in a format different than the asm works with? +%ifidn %1, triangular +cvtdq2ps m2, [dither1q+2*lenq] +cvtdq2ps m3, [dither1q+2*lenq+mmsize] +addps m0, m2 +addps m1, m3 Can you add before converting, or does that overflow? +%endif +mulps m2, m5, [srcq+2*lenq] +mulps m3, m5, [srcq+2*lenq+mmsize] +fmaddps m0, m0, m4, m2, m3 +fmaddps m1, m1, m4, m3, m2 +cvtps2dq m0, m0 +cvtps2dq m1, m1 +packssdw m0, m1 +mova [dstq+lenq], m0 +add lenq, mmsize +jl .loop +REP_RET +%endmacro --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 2/4] lavr: x86: add SSE2 and FMA4 quantization for non-filtering dither methods
On Sun, 16 Dec 2012, Justin Ruggles wrote: On 12/16/2012 06:27 AM, Loren Merritt wrote: Do you have to store dither in a format different than the asm works with? That's the whole point of these functions. av_get_lfg() returns an unsigned int, so instead of converting each value to float and scaling individually, we can use SSE2 to do the conversion/scaling of multiple values along with adding the noise to the samples. For the triangular dither with high-pass it actually does convert/scale the dither noise separately, filters it, then adds it into the samples. But for the dither methods that don't require filtering, we can do it all at once. Why is av_get_lfg in the inner loop? Is there an audible difference between that and repeating a constant array? --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 6/7] x86: af_volume: add SSE2/SSSE3/AVX-optimized s32 volume scaling
On Mon, 3 Dec 2012, Justin Ruggles wrote: On 12/03/2012 04:19 PM, Loren Merritt wrote: On Sun, 2 Dec 2012, Justin Ruggles wrote: +; NOTE: This is not bit-identical with the C version because it clips to +; [-INT_MAX, INT_MAX] instead of [INT_MIN, INT_MAX] + +INIT_XMM ssse3 INIT_XMM ssse3, atom Why is that needed? We don't have 2 different ssse3 versions of that function, and that cpuflag isn't used in the function itself. Not really needed, but I thought it was more aesthetic to have the function name reflect what it's optimized for. Normally I expect foo_ssse3 to be an improvemenet over foo_sse2. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 6/7] x86: af_volume: add SSE2/SSSE3/AVX-optimized s32 volume scaling
On Sun, 2 Dec 2012, Justin Ruggles wrote: +; NOTE: This is not bit-identical with the C version because it clips to +; [-INT_MAX, INT_MAX] instead of [INT_MIN, INT_MAX] + +INIT_XMM ssse3 INIT_XMM ssse3, atom +cglobal scale_samples_s32, 4,4,8, dst, src, len, volume +movdm4, volumem +pshufd m4, m4, 0 +movam5, [pq_128] +pxorm6, m6 +lea lenq, [lend*4-mmsize] +.loop: +; src[i] = av_clipl_int32((src[i] * volume + 128) 8); +movam7, [srcq+lenq] +pabsd m3, m7 +pshufd m0, m3, q0100 +pshufd m1, m3, q0302 +pmuludq m0, m4 +pmuludq m1, m4 +paddq m0, m5 +paddq m1, m5 +psrlq m0, 7 +psrlq m1, 7 +shufps m2, m0, m1, q3131 +shufps m0, m0, m1, q2020 +pcmpgtd m2, m6 +por m0, m2 +psrld m0, 1 +psignd m0, m7 +mova [dstq+lenq], m0 +sub lenq, mmsize +jge .loop +REP_RET --- a/libavutil/x86/x86inc.asm +++ b/libavutil/x86/x86inc.asm @@ -957,6 +957,7 @@ AVX_INSTR cmpsd, 1, 0, 0 AVX_INSTR cmpss, 1, 0, 0 AVX_INSTR cvtdq2ps, 1, 0, 0 AVX_INSTR cvtps2dq, 1, 0, 0 +AVX_INSTR cvtpd2dq, 1, 0, 0 AVX_INSTR divpd, 1, 0, 0 AVX_INSTR divps, 1, 0, 0 AVX_INSTR divsd, 1, 0, 0 Alphabetical order --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 10/10] SBR DSP x86: implement SSE hf_apply_noise
On Sat, 1 Dec 2012, Christophe Gisquet wrote: 2012/11/30 Loren Merritt lor...@u.washington.edu: +cglobal sbr_hf_apply_noise_0, 4,5,8, Y,s_m,q_filt,noise,kx,m_max + mova m0, [ps_noise0] + mov r4d, m_maxm + call hf_apply_noise_main + RET TAIL_CALL hf_apply_noise_main, 1 Which makes me think that every caller should have the same epilog (same stack offset etc). Is there a way I just do a jmp here and let the jumpee do the epilog. Move the callee after the callers in sourcecode order, and make it a plain label, not a cglobal. Then it doesn't reset the registers in use, so RET does the same thing it would have in the previous function. Also don't have to repeat DECLARE_ARGS that way. One of the callers can fallthrough rather than jump. Another thing I'm wondering (can't make sure for the next 4 days): mov r4d, m_maxm If I'm not mistaken, m_max should already be in r5 for linux x86_64/amd64 ABI (whatever I should call it). So I could save that mov and have instead hf_apply_noise_main use r5 under that condition. Yes. But probably negligible. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 01/10] SBR DSP x86: implement SSE sbr_hf_gen
On Fri, 30 Nov 2012, Christophe Gisquet wrote:
+movam0, [X_lowq + start]
+movlhps m1, m1 ; (a2 a3 a2 a3)
+movlhps m2, m2 ; (a0 a1 a0 a1)
+shufps m3, m3, 00010001b ; (a3 a2 a3 a2)
+shufps m4, m4, 00010001b ; (a1 a0 a1 a0)
+xorps m3, m7 ; (-a3 a2 -a3 a2)
+xorps m4, m7 ; (-a1 a0 -a1 a0)
+.loop2:
+movam5, m0
+movam6, m0
+shufps m0, m0, 1010b ; {Xl[-2][0],,Xl[-1][0],}
+shufps m5, m5, 0101b ; {Xl[-2][1],,Xl[-1][1],}
+mulps m0, m2
+mulps m5, m4
+movam7, m6
+addps m5, m0
+movam0, [X_lowq + start + 2*2*4]
+shufps m6, m0, 1010b ; {Xl[-1][0],,Xl[0][0],}
+shufps m7, m0, 0101b ; {Xl[-1][1],,Xl[1][1],}
Recommend using base-4 for shuffle constants.
--Loren Merritt
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 02/10] SBR DSP x86: implement SSE sum64x5
On Fri, 30 Nov 2012, Christophe Gisquet wrote: 698 to 174 cycles on penrynn. Unrolling is a 6 cycles gain. --- libavcodec/x86/sbrdsp.asm| 22 ++ libavcodec/x86/sbrdsp_init.c |2 ++ 2 files changed, 24 insertions(+), 0 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 03/10] SBR DSP x86: implement SSE qmf_post_shuffle
On Fri, 30 Nov 2012, Christophe Gisquet wrote: +cglobal sbr_qmf_post_shuffle, 2,3,3,W,z + lea r2q, [zq + (64-4)*4] +.loop: + mova m0, [r2q] + mova m1, [zq ] + xorps m0, [ps_neg] + shufps m0, m0, 0x1B + mova m2, m0 + unpcklps m0, m1 + unpckhps m2, m1 + mova [Wq + 0], m0 + mova [Wq + 16], m2 + addWq, 32 + sub r2q, 16 + addzq, 16 + cmpzq, r2q + jl .loop + REP_RET If you increment an index into W and z rather than the pointers themselves, then you can eliminate an add and a cmp. 4 space tabs. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 10/10] SBR DSP x86: implement SSE hf_apply_noise
On Fri, 30 Nov 2012, Christophe Gisquet wrote: 497 to 253 cycles under Win64. cpu is more relevant than os. +; r0q=Y r1q=s_m r2q=q_filt r3q=noise r4q=max_m +cglobal hf_apply_noise_main You can invoke DEFINE_ARGS even if not generating a prologue. + dec r3q + shl r4q, 2 + lea r0q, [r0q + 2*r4q] + add r1q, r4q + add r2q, r4q + shl r3q, 3 + xorps m5, m5 + neg r4q +.loop: + add r3q, 16 + and r3q, 0x1ff3 + movh m1, [r2q + r4q] + movu m3, [r3q + sbr_noise_table] + movh m2, [r2q + r4q + 8] + add r3q, 16 + and r3q, 0x1ff3 + movu m4, [r3q + sbr_noise_table] + unpcklps m1, m1 + unpcklps m2, m2 + mulps m1, m3 ; m2 = q_filt[m] * ff_sbr_noise_table[noise] + mulps m2, m4 ; m2 = q_filt[m] * ff_sbr_noise_table[noise] + movh m3, [r1q + r4q] + movh m4, [r1q + r4q + 8] Can these be a single aligned load? + unpcklps m3, m3 + unpcklps m4, m4 + mova m6, m3 + mova m7, m4 + mulps m3, m0 ; s_m[m] * phi_sign + mulps m4, m0 ; s_m[m] * phi_sign + cmpps m6, m5, 0 ; m1 == 0 + cmpps m7, m5, 0 ; m1 == 0 You mean m7 == 0? + andps m1, m6 + andps m2, m7 + movu m6, [r0q + 2*r4q] + movu m7, [r0q + 2*r4q + 16] + addps m6, m1 + addps m7, m2 + addps m6, m3 + addps m7, m4 + movu[r0q + 2*r4q], m6 + movu[r0q + 2*r4q + 16], m7 + add r4q, 16 + jl .loop + ret + +; sbr_hf_apply_noise_0(float (*Y)[2], const float *s_m, +; const float *q_filt, int noise, +; int kx, int m_max) +cglobal sbr_hf_apply_noise_0, 4,5,8, Y,s_m,q_filt,noise,kx,m_max + mova m0, [ps_noise0] + mov r4d, m_maxm + call hf_apply_noise_main + RET TAIL_CALL hf_apply_noise_main, 1 --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/1] h264: reset has_b_frames after enabling low_delay from SPS
On Wed, 21 Nov 2012, Janne Grunau wrote:
On 2012-11-16 18:14:29 -0800, Ronald S. Bultje wrote:
On Fri, Nov 16, 2012 at 8:43 AM, Janne Grunau janne-li...@jannau.netwrote:
Fixes a crash in fuzzed file nasa-8s2.ts_s20033 caused by a too large
has_b_frames value. low_delay keeps getting re-enabled from the the
presumely broken SPS.
---
libavcodec/h264.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/libavcodec/h264.c b/libavcodec/h264.c
index c30c478..7935fe6 100644
--- a/libavcodec/h264.c
+++ b/libavcodec/h264.c
@@ -2346,8 +2346,10 @@ static int h264_set_parameter_from_sps(H264Context
*h)
if (s-flags CODEC_FLAG_LOW_DELAY ||
(h-sps.bitstream_restriction_flag
- !h-sps.num_reorder_frames))
+ !h-sps.num_reorder_frames)) {
s-low_delay = 1;
+s-avctx-has_b_frames = 0;
+}
So I'm going to have to wonder what happens to the delayed frames already
cached in h-delayed_pics[]? Are they still output?
yes, delayed pictures are returned as long as h-delayed_pics[0] is not
NULL.
And when you return something from delayed_pics[] instead of the new
frame that just got decoded, the new frame gets delayed, so that you
don't ever actually switch to low delay mode?
--Loren Merritt
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 35/45] x86: PSIGNW: port to cpuflags
On Sun, 18 Nov 2012, Justin Ruggles wrote: On 07/31/2012 06:17 PM, Diego Biurrun wrote: --- libavutil/x86/x86util.asm | 10 +- 1 files changed, 5 insertions(+), 5 deletions(-) diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm index b844131..db8899b 100644 --- a/libavutil/x86/x86util.asm +++ b/libavutil/x86/x86util.asm @@ -160,13 +160,13 @@ %endif %endmacro -%macro PSIGNW_MMX 2 +%macro PSIGNW 2 +%if cpuflag(ssse3) +psignw %1, %2 +%else %elif mmsize == 8 || cpuflag(sse2) pxor %1, %2 psubw %1, %2 -%endmacro - -%macro PSIGNW_SSSE3 2 -psignw %1, %2 +%endif %else %error PSIGNW with XMM requires SSE2 %endif Do you have a use-case for this error message? None of our other macros have one, because I assume that you just don't try to instantiate sse1 verions of functions that do integer simd. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [RFC PATCH 1/2] h264: set parameters from SPS whenever it changes
On Fri, 16 Nov 2012, Ronald S. Bultje wrote: So ... I originally added code to not support changing bitdepth at all, basically to prevent this and alike issues. E.g., what if the reference is 8bit but the next frame is 10bit? The standard allows SPS to change only at IDR-frames, i.e. where there aren't any old references remaining accessible. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86: dsputil: port to cpuflags
On Wed, 14 Nov 2012, Diego Biurrun wrote: --- Now addressing the points from Loren's review and more refactoring around bswap_buf. libavcodec/x86/dsputil.asm | 220 -- libavcodec/x86/dsputil_mmx.c | 26 +++--- 2 files changed, 119 insertions(+), 127 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 25/45] x86: h264_weight_10bit: port to cpuflags
On Wed, 1 Aug 2012, Diego Biurrun wrote: libavcodec/x86/h264_weight_10bit.asm | 132 ++ 1 files changed, 69 insertions(+), 63 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 12/45] x86: dsputil_yasm: port to cpuflags
On Wed, 1 Aug 2012, Diego Biurrun wrote: -%macro APPLY_WINDOW_INT16 3 ; %1=instruction set, %2=mmxext/sse2 bit exact version, %3=has_ssse3 -cglobal apply_window_int16_%1, 4,5,6, output, input, window, offset, offset2 +%macro APPLY_WINDOW_INT16 1 ; %1=mmxext/sse2 bitexact version Comment is no longer true, you changed it to all of the bitexact versions. +%if %1 +cglobal apply_window_int16, 4, 5, 6, output, input, window, offset, offset2 +%else +cglobal apply_window_int16_round, 4, 5, 6, output, input, window, offset, offset2 +%endif lea offset2q, [offsetq-mmsize] -%if %2 -mova m5, [pd_16384] -%elifidn %1, ssse3 +%if cpuflag(ssse3) mova m5, [pb_revwords] ALIGN 16 +%elif %1 +mova m5, [pd_16384] %endif Atom version uses neither of these constants. @@ -426,17 +431,9 @@ cglobal add_hfyu_median_prediction_mmxext, 6,6,0, dst, top, diff, w, left, left_ %endmacro ; int add_hfyu_left_prediction(uint8_t *dst, const uint8_t *src, int w, int left) -INIT_MMX -cglobal add_hfyu_left_prediction_ssse3, 3,3,7, dst, src, w, left -.skip_prologue: -movam5, [pb_7] -movam4, [pb_] -movam3, [pb_zz11zz55zz99zzdd] -movdm0, leftm -psllq m0, 56 -ADD_HFYU_LEFT_LOOP 1 - -INIT_XMM +%macro ADD_HFYU_LEFT_PREDICTION 0 +cglobal add_hfyu_left_prediction, 3,3,7, dst, src, w, left +%if cpuflag(sse4) cglobal add_hfyu_left_prediction_sse4, 3,3,7, dst, src, w, left movam5, [pb_f] movam6, [pb_] @@ -451,10 +448,25 @@ cglobal add_hfyu_left_prediction_sse4, 3,3,7, dst, src, w, left ADD_HFYU_LEFT_LOOP 1 .unaligned: ADD_HFYU_LEFT_LOOP 0 +%else +.skip_prologue: +movam5, [pb_7] +movam4, [pb_] +movam3, [pb_zz11zz55zz99zzdd] +movdm0, leftm +psllq m0, 56 +ADD_HFYU_LEFT_LOOP 1 +%endif +%endmacro +INIT_MMX ssse3 +ADD_HFYU_LEFT_PREDICTION +INIT_XMM sse4 +ADD_HFYU_LEFT_PREDICTION There's no code being shared here, why add a superfluous level of indirection? Likewise with bswap_buf. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 24/45] x86: h264_qpel_10bit: drop unused parameter from MC10/MC20/MC30 macros
On Fri, 9 Nov 2012, Diego Biurrun wrote: On Wed, Aug 01, 2012 at 12:17:48AM +0200, Diego Biurrun wrote: --- libavcodec/x86/h264_qpel_10bit.asm | 10 +- 1 files changed, 5 insertions(+), 5 deletions(-) .. ping .. (this is rather simple and trivial) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 08/10] x86: af_volume: add SSE2/SSSE3/AVX-optimized s32 volume scaling
On Fri, 9 Nov 2012, Christophe Gisquet wrote: 2012/11/8 Loren Merritt lor...@u.washington.edu: This code shouldn't be latency-bound anyway. Or if it is, all you need is more unrolling. But you do notice the discrepancy in timings? Atom's SSSE3 version is twice as fast as SSE2, but SB SSSE3 is slower than SSE2. Using opannotate may help lay the blame if that is the issue. You mean Atom's SSE2 is twice as slow as SSSE3. Of course there's a discrepancy when mulpd is 1 uop on SB and 9 cycles on Atom; and similarly for all the other double instructions. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86: h264_qpel_10bit: port to cpuflags
On Thu, 8 Nov 2012, Diego Biurrun wrote: --- This version incorporates the changes suggested by Loren. libavcodec/x86/h264_qpel_10bit.asm | 314 ++-- 1 files changed, 155 insertions(+), 159 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 08/10] x86: af_volume: add SSE2/SSSE3/AVX-optimized s32 volume scaling
On Wed, 7 Nov 2012, Christophe Gisquet wrote: err, mail sent before I finished editing it... 2012/11/7 Christophe Gisquet christophe.gisq...@gmail.com: Nehalem Nehalem has 2 cycles delay, SB 0 or 1, Atom 0 when mixing such instruction. So that mix of instructions may cause issues. This code shouldn't be latency-bound anyway. Or if it is, all you need is more unrolling. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 23/45] x86: h264_qpel_10bit: port to cpuflags
On Wed, 1 Aug 2012, Diego Biurrun wrote: -%macro MCAxA 8 +%macro MCAxA 7 %if ARCH_X86_64 -%ifnidn %1,mmxext -MCAxA_OP %1,%2,%3,%4,%5,%6,%7,%8 +%ifnidn SUFFIX, mmxext +MCAxA_OP %1, %2, %3, %4, %5, %6, %7 %endif %else -MCAxA_OP %1,%2,%3,%4,%5,%6,%7,%8 +MCAxA_OP %1, %2, %3, %4, %5, %6, %7 %endif %endmacro %macro MCAxA 7 %if ARCH_X86_32 || cpuflag(sse2) MCAxA_OP %1, %2, %3, %4, %5, %6, %7 %endif %endmacro and could be inlined into cglobal_mc. -%macro MCAxA_OP 8 +%macro MCAxA_OP 7 %if ARCH_X86_32 -cglobal %2_h264_qpel%5_%3_10_%1, %6,%7,%8 -call stub_%2_h264_qpel%4_%3_10_%1 +cglobal %1_h264_qpel%4_%2_10, %5, %6, %7 +call stub_%1_h264_qpel%3_%2_10 %+ cpuname mov r0, r0m mov r1, r1m -add r0, %4*2 +add r0, %3*2 add r1, %4*2 -call stub_%2_h264_qpel%4_%3_10_%1 +call stub_%1_h264_qpel%3_%2_10 %+ cpuname mov r0, r0m mov r1, r1m -lea r0, [r0+r2*%4] -lea r1, [r1+r2*%4] -call stub_%2_h264_qpel%4_%3_10_%1 +lea r0, [r0+r2*%3] +lea r1, [r1+r2*%3] +call stub_%1_h264_qpel%3_%2_10 %+ cpuname mov r0, r0m mov r1, r1m -lea r0, [r0+r2*%4+%4*2] -lea r1, [r1+r2*%4+%4*2] -call stub_%2_h264_qpel%4_%3_10_%1 +lea r0, [r0+r2*%3+%3*2] +lea r1, [r1+r2*%3+%3*2] +call stub_%1_h264_qpel%3_%2_10 %+ cpuname RET All your uses of cpuname should instead be SUFFIX. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 6/6] x86: af_volume: add SSSE3/AVX-optimized s32 volume scaling
On Mon, 5 Nov 2012, Justin Ruggles wrote: Yeah, it works for SSE2 and AVX. Unfortunately it's not too fast. Not sure how I should handle this in terms of checking flags to determine when to use the SSE2 version. I suppose it depends on how well it performs on Intel cpus that have SSE2 but not SSSE3. Athlon64: C- 6267 SSE2 - 4332 Sandy Bridge: C- 4001 SSE2 - 4560 SSSE3- 2033 AVX (int)- 1937 AVX (double) - 44619 Here is the code I tested for SSE2 and AVX: INIT_XMM sse2 cglobal scale_samples_s32, 4,4,8, dst, src, len, volume movdm2, volumem pshufd m2, m2, 0 cvtdq2pdm2, m2 mulpd m2, [pd_1_256] movam3, [pd_int32_max] lea lenq, [lend*4-mmsize] .loop: movam0, [srcq+lenq] movhlps m1, m0 cvtdq2pdm0, m0 cvtdq2pdm1, m1 mulpd m0, m2 mulpd m1, m2 minpd m0, m3 minpd m1, m3 cvtpd2dqm0, m0 cvtpd2dqm1, m1 punpcklqdq m0, m1 mova [dstq+lenq], m0 sub lenq, mmsize jge .loop REP_RET INIT_XMM avx cglobal scale_samples_s32, 4,4,8, dst, src, len, volume movd xmm1, volumem pshufdxmm1, xmm1, 0 vcvtdq2pd ymm1, xmm1 vmulpdymm1, ymm1, [pd_1_256] vmovapd ymm2, [pd_int32_max] lea lenq, [lend*4-mmsize] .loop: vcvtdq2pd ymm0, [srcq+lenq] vmulpdymm0, ymm0, ymm1 vminpdymm0, ymm0, ymm2 vcvtpd2dq xmm0, ymm0 mova [dstq+lenq], xmm0 vmovdqa. You can't mix ymm instructions with xmm instructions that don't begin with a v. Or rather, you can because intel didn't even have the decency to make that combination crash so that you'd notice it; instead it's very slow. sub lenq, mmsize jge .loop vzeroupper RET --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: Set program_name and cpuflags_3dnowext outside of x86inc.asm
On Sat, 3 Nov 2012, Diego Biurrun wrote: On Sat, Nov 03, 2012 at 06:13:31AM +, Loren Merritt wrote: On Fri, 2 Nov 2012, Diego Biurrun wrote: This reduces the local difference to the x264 upstream version. --- Now also sets cpuflags_3dnowext locally, further reducing the difference. x264 doesn't use cpuflags_3dnow* at all, it only exists because it's in libav. I was about to send you some x264 patches to sync those files further. Do you want me to drop the 3dnow cpuflags on the x264 side while at it now that we have a means of setting them on the libav side w/o patching x86inc.asm? No preference. Just saying it should be either synched or dropped. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: Set program_name and cpuflags_3dnowext outside of x86inc.asm
On Fri, 2 Nov 2012, Diego Biurrun wrote: This reduces the local difference to the x264 upstream version. --- Now also sets cpuflags_3dnowext locally, further reducing the difference. x264 doesn't use cpuflags_3dnow* at all, it only exists because it's in libav. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: Add cpuflags_mmxext alias for cpuflags_mmx2
On Mon, 29 Oct 2012, Ronald S. Bultje wrote: On Mon, Oct 29, 2012 at 5:15 PM, Diego Biurrun di...@biurrun.de wrote: This allows using mmxext as name in Libav while staying compatible with changes to the YASM macro infrastructure imported from x264. --- libavutil/x86/x86inc.asm |1 + 1 files changed, 1 insertions(+), 0 deletions(-) diff --git a/libavutil/x86/x86inc.asm b/libavutil/x86/x86inc.asm index 1fe9f55..2655154 100644 --- a/libavutil/x86/x86inc.asm +++ b/libavutil/x86/x86inc.asm @@ -35,6 +35,7 @@ ; to x264-de...@videolan.org . %define program_name ff +%define cpuflags_mmxext cpuflags_mmx2 Any such a change to x86inc.asm is OK if and only if x264 accepts this change upstream. Loren? I'll accept the addition of the synonym. (This doesn't change my objections to the name, I'm just allowing an inert line of code to live in the x264 repo if that simplifies merging.) --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 6/6] x86: af_volume: add SSSE3/AVX-optimized s32 volume scaling
On Sat, 29 Sep 2012, Justin Ruggles wrote: +cglobal scale_samples_s32, 4,4,8, dst, src, len, volume +movdm4, volumem +pshufd m4, m4, 0 +movam5, [pq_128] +pxorm6, m6 +lea lenq, [lend*4-mmsize] +.loop: +; src[i] = av_clipl_int32((src[i] * volume + 128) 8); +movam7, [srcq+lenq] +pabsd m3, m7 +pshufd m0, m3, q0100 +pshufd m1, m3, q0302 +pmuludq m0, m4 +pmuludq m1, m4 +paddq m0, m5 +paddq m1, m5 +psrlq m0, 7 +psrlq m1, 7 +shufps m2, m0, m1, q3131 +shufps m0, m0, m1, q2020 +pcmpgtd m2, m6 +por m0, m2 +psrld m0, 1 +psignd m0, m7 +mova [dstq+lenq], m0 +sub lenq, mmsize +jge .loop +REP_RET +%endmacro .loop: vcvtdq2pd ymm0, [srcq+lenq] vmulpdymm0, ymm1 vminpdymm0, ymm2 ; no max needed, since underflows convert to INT_MIN anyway vcvtpd2dq xmm0, ymm0 vmovdqa [dstq+lenq], xmm0 sub lenq, mmsize jge .loop Should also work for sse2; I don't know if that'll be faster than ssse3 int. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 6/8] x86: af_volume: add SSE2-optimized s16 volume scaling
On Fri, 28 Sep 2012, Justin Ruggles wrote:
--- /dev/null
+++ b/libavfilter/x86/af_volume_init.c
+void ff_volume_init_x86(VolumeContext *vol)
+{
+int mm_flags = av_get_cpu_flags();
+enum AVSampleFormat sample_fmt =
av_get_packed_sample_fmt(vol-sample_fmt);
+
+if (sample_fmt == AV_SAMPLE_FMT_S16) {
+if (EXTERNAL_SSE2(mm_flags) vol-volume_i 65536) {
32768
+vol-scale_samples_int = ff_scale_samples_s16_sse2;
+vol-samples_align = 8;
+}
+}
+}
--Loren Merritt
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 7/9] x86: af_volume: add SSE4-optimized s16 volume scaling
On Tue, 25 Sep 2012, Justin Ruggles wrote: +;-- +; void ff_scale_samples_s16(VolumeContext *ctx, uint8_t *src, int len, +; int volume) +;-- + +INIT_XMM sse4 +cglobal scale_samples_s16, 4,4,4, ctx, src, len, volume +movdm0, volumed +pshufd m0, m0, 0 +movam1, [pd_128] +lea lenq, [lend*2-mmsize] +.loop: +; src[i] = av_clip_int16((src[i] * volume + 128) 8); +movam2, [srcq+lenq] +pmovsxwdm3, m2 +psrldq m2, 8 +pmovsxwdm2, m2 +pmulld m2, m0 +pmulld m3, m0 +paddd m2, m1 +paddd m3, m1 +psrad m2, 8 +psrad m3, 8 +packssdwm3, m2 +mova [srcq+lenq], m3 +sublenq, mmsize +jge .loop +REP_RET INIT_XMM sse2 cglobal scale_samples_s16, 3,3,4, ctx, src, len, volume movdm0, volumem pshuflw m0, m0, 0 punpcklwd m0, [pw_1] movam1, [pw_128] lea lenq, [lend*2-mmsize] .loop: ; src[i] = av_clip_int16((src[i] * volume + 128) 8); movam2, [srcq+lenq] punpcklwd m3, m2, m1 punpckhwd m2, m1 pmaddwd m3, m0 pmaddwd m2, m0 psrad m3, 8 psrad m2, 8 packssdwm3, m2 mova [srcq+lenq], m3 sublenq, mmsize jge .loop REP_RET (untested) --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 9/9] x86: af_volume: add SSSE3/AVX-optimized s32 volume scaling
On Tue, 25 Sep 2012, Justin Ruggles wrote: SECTION_RODATA 32 pd_128: times 4 dd 128 +pq_128: times 2 dq 128 +pq_maxsq: times 2 dq 549755813632 ; INT_MAX 8 +pb_packqd_rsh8: db 1, 2, 3, 4, 9, 10, 11, 12, -1, -1, -1, -1, -1, -1, -1, -1 SECTION_TEXT @@ -55,3 +58,58 @@ cglobal scale_samples_s16, 4,4,4, ctx, src, len, volume sublenq, mmsize jge .loop REP_RET + +;-- +; void ff_scale_samples_s32(VolumeContext *ctx, uint8_t *src, int len, +; int volume) +;-- + +; NOTE: This is not bit-identical with the C version because it clips to +; [-INT_MAX, INT_MAX] instead of [INT_MIN, INT_MAX] + +%macro PMINSQ 3 ; dst, src, tmp +pcmpgtq %3, %2, %1 +pxor %1, %2 +pand %1, %3 +pxor %1, %2 +%endmacro + +%macro SCALE_SAMPLES_S32 0 +cglobal scale_samples_s32, 4,4,8, ctx, src, len, volume +movdm0, volumed +pshufd m0, m0, 0 +movam3, [pq_128] +movam4, [pq_maxsq] +movam5, [pb_packqd_rsh8] +lea lenq, [lend*4-mmsize] +.loop: +; src[i] = av_clipl_int32((src[i] * volume + 128) 8); +movam1, [srcq+lenq] +movam6, m1 +movam2, m1 +punpckhdq m2, m1, m1 +punpckldq m1, m1, m1 +pabsd m1, m1 +pabsd m2, m2 +pmuludq m1, m0 +pmuludq m2, m0 +paddq m1, m3 +paddq m2, m3 +PMINSQ m1, m4, m7 +PMINSQ m2, m4, m7 +pshufb m1, m5 +pshufb m2, m5 +punpcklqdq m1, m2 +psignd m1, m6 +mova [srcq+lenq], m1 +sub lenq, mmsize +jge .loop +REP_RET +%endmacro cglobal scale_samples_s32, 3,3,7, ctx, src, len, volume movdm4, volumem pshufd m4, m4, 0 movam5, [pq_128] pxorm6, m6 lea lenq, [lend*4-mmsize] .loop: ; src[i] = av_clipl_int32((src[i] * volume + 128) 8); pabsd m3, [srcq+lenq] pshufd m0, m3, q0100 pshufd m1, m3, q0302 pmuludq m0, m4 pmuludq m1, m4 paddq m0, m5 paddq m1, m5 psrlq m0, 7 psrlq m1, 7 shufps m2, m0, m1, q3131 shufps m0, m0, m1, q2020 pcmpgtd m2, m6 por m0, m2 psrld m0, 1 psignd m0, m3 mova [srcq+lenq], m0 sub lenq, mmsize jge .loop REP_RET (untested) --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] libx264: change i_qfactor to use x264cli's default
On Tue, 25 Sep 2012, John Van Sickle wrote: On Tue, Sep 25, 2012 at 2:12 PM, Luca Barbato lu_z...@gentoo.org wrote: On 09/25/2012 06:29 PM, John Van Sickle wrote: -x4-params.rc.f_ip_factor = 1 / fabs(avctx-i_quant_factor); +if (avctx-i_quant_factor = 0) +x4-params.rc.f_ip_factor = avctx-i_quant_factor; Looks strange to me. before it was 1/qf... As discussed on irc, I have no idea why it was previously: 1 / fabs(avctx-i_quant_factor); With my patch the option now behaves the same way as x264cli. The default is 1.40, and using -tune grain correctly sets ipratio to 1.10. If you set --ipratio .80 using x264cli, that's what you get. Not 1/.80 like avconv is currently doing. Do this make sense? The meaning that x264 assigns to f_ip_factor is the reciprocal of the meaning that libav codecs assign to i_quant_factor. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] vorbisdec: ensure FASTDIV denominator is never 1
On Thu, 13 Sep 2012, Luca Barbato wrote: Both in usage of FASTDIV the denominator might be 1. Using a branch would make the function slower than using a normal division. Which CPU do you have in mind where a division is faster than a branch miss, let alone a predictable branch? (No objection to the patch, it's only the commit message that's wrong.) --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/2] x86: dsputil: Move SSE4 function pointer setting to SSE4 section
On Thu, 6 Sep 2012, Diego Elio Pettenò wrote: On 06/09/2012 13:20, Diego Biurrun wrote: -if (mm_flags AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe -c-add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4; This looks like it might have to be tied to something else -- AMD also has ssse3 don't they? Semantically it should check SSSE3 SHUFFLE_IS_FAST rather than SSE4. If anyone wants to update the set of cpuflags to include the things (such as SHUFFLE_IS_FAST) that I invented for x264 to represent performance characteristics that aren't instructions sets. But the extensional equivalence of SHUFFLE_IS_FAST and SSE4 remains true on AMD too. --Loren Merritt___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86: allow using add_hfyu_median_prediction_cmov on any cpu with cmov
On Sun, 2 Sep 2012, Mans Rullgard wrote: For some reasion add_hfyu_median_prediction_cmov is only selected on 3Dnow-capable CPUs, even though it uses no 3Dnow instructions. This patch allows it to be selected on any cpu with cmov with the possibility of being overridden by the mmxext version. As long as mmxext doesn't override it on AMD, where cmov is faster. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 7/7] vf_hqdn3d: x86 asm
On Thu, 9 Aug 2012, Kieran Kunhya wrote: On Sat, Jul 28, 2012 at 8:55 PM, Loren Merritt lor...@u.washington.edu wrote: 13% faster on penryn, 16% on sandybridge, 15% on bulldozer Not simd; a compiler should have generated this, but gcc didn't. --- libavfilter/vf_hqdn3d.c| 27 ++-- libavfilter/x86/Makefile |1 + libavfilter/x86/hqdn3d.asm | 106 libavutil/x86/x86inc.asm |1 + 4 files changed, 131 insertions(+), 4 deletions(-) create mode 100644 libavfilter/x86/hqdn3d.asm Looks ok to me. Pushed. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] H.264: Convert 8-bit qpel inlined assembly to yasm
On Thu, 23 Aug 2012, Daniel Kang wrote: On Wed, Aug 22, 2012 at 11:30 PM, Loren Merritt lor...@u.washington.edu wrote: On Wed, 22 Aug 2012, daniel.d.k...@gmail.com wrote: +%macro QPEL4_H_LOWPASS_OP 1 +cglobal %1_h264_qpel4_h_lowpass, 4,5 ; dst, src, dstStride, srcStride +%define OP op_%1h I don't think this define clarifies anything, and it's only used once or twice each function. What do you suggest I do instead? Write op_%1h in the place where it's used. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] H.264: Convert 8-bit qpel inlined assembly to yasm
On Wed, 22 Aug 2012, daniel.d.k...@gmail.com wrote: +; void pixels8_l2_mmx2(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h) +%macro PIXELS8_L2 1 +%define OP op_%1 +cglobal %1_pixels8_l2, 6,6 +test r5d, 1 +je .loop +movam0, [r1] +movam1, [r2] +add r1, r4 +add r2, 8 +pavgb m0, m1 +OP m0, [r0] +add r0, r3 +decr5d +.loop: +movam0, [r1] +add r1, r4 +movam1, [r1] +add r1, r4 +pavgb m0, [r2] +pavgb m1, [r2+8] +OP m0, [r0] +add r0, r3 +OP m1, [r0] +add r0, r3 +movam0, [r1] +add r1, r4 +movam1, [r1] +add r1, r4 +pavgb m0, [r2+16] +pavgb m1, [r2+24] +OP m0, [r0] +add r0, r3 +OP m1, [r0] +add r0, r3 +add r2, 32 +subr5d, 4 +jne .loop +REP_RET +%endmacro More adds than necessary. Use [r1+r4]. +%macro QPEL4_H_LOWPASS_OP 1 +cglobal %1_h264_qpel4_h_lowpass, 4,5 ; dst, src, dstStride, srcStride +%define OP op_%1h I don't think this define clarifies anything, and it's only used once or twice each function. +%macro QPEL8_H_LOWPASS_OP_XMM 1 +%define OP op_%1h +cglobal %1_h264_qpel8_h_lowpass, 4,5,7 ; dst, src, dstStride, srcStride +mov r4d, 8 +pxor m7, m7 +mova m6, [pw_5] +.loop: +lddqu m1, [r1-2] +mova m0, m1 +punpckhbw m1, m7 +punpcklbw m0, m7 +mova m2, m1 +mova m3, m1 +mova m4, m1 +mova m5, m1 +palignr m4, m0, 2 +palignr m3, m0, 4 +palignr m2, m0, 6 +palignr m1, m0, 8 +palignr m5, m0, 10 +paddw m0, m5 +paddw m2, m3 +paddw m1, m4 +psllw m2, 2 +psubw m2, m1 +paddw m0, [pw_16] +pmullwm2, m6 +paddw m2, m0 +psraw m2, 5 +packuswb m2, m2 +OPm2, [r0], m4 +add r1, r3 +add r0, r2 +dec r4d +jne .loop +REP_RET +%endmacro + +INIT_XMM ssse3 +QPEL8_H_LOWPASS_OP_XMM put +QPEL8_H_LOWPASS_OP_XMM avg There aren't any cpus that have both lddqu and ssse3. Use movu instead, since that's what lddqu actually does on everything other than pentium4. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86: Drop silly _yasm suffixes from filenames
On Wed, 8 Aug 2012, M�ns Rullg�rd wrote:
Diego Biurrun di...@biurrun.de writes:
libavcodec/x86/Makefile|6 +++---
libavcodec/x86/{dsputil_yasm.asm = dsputil.asm} |0
.../x86/{dsputilenc_yasm.asm = dsputilenc.asm}|0
libavcodec/x86/{vc1dsp_yasm.asm = vc1dsp.asm} |0
4 files changed, 3 insertions(+), 3 deletions(-)
rename libavcodec/x86/{dsputil_yasm.asm = dsputil.asm} (100%)
rename libavcodec/x86/{dsputilenc_yasm.asm = dsputilenc.asm} (100%)
rename libavcodec/x86/{vc1dsp_yasm.asm = vc1dsp.asm} (100%)
Makes sense. _yasm doesn't convey any information not already given by
the .asm suffix, given that all the x86 .asm files use yasm syntax.
Same for _mmx.
Do we have a standard for what to do when we have both a dsputil.c and a
dsputil.asm, where the .o's would collide if we gave both their natural
basename?
--Loren Merritt___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 06/15] lavr: x86: optimized 6-channel fltp to s16 conversion
On Mon, 6 Aug 2012, Justin Ruggles wrote: --- a/libavresample/x86/audio_convert.asm +++ b/libavresample/x86/audio_convert.asm @@ -581,6 +581,120 @@ CONV_FLTP_TO_S16_2CH INIT_XMM ssse3 CONV_FLTP_TO_S16_2CH +;-- +; void ff_conv_fltp_to_s16_6ch(int16_t *dst, float *const *src, int len, +; int channels); +;-- + +%macro CONV_FLTP_TO_S16_6CH 0 +%if ARCH_X86_64 +cglobal conv_fltp_to_s16_6ch, 3,8,7, dst, src, len, src1, src2, src3, src4, src5 +%else +cglobal conv_fltp_to_s16_6ch, 2,7,7, dst, src, src1, src2, src3, src4, src5 +%define lend dword r2m +%endif +movsrc1q, [srcq+1*gprsize] +movsrc2q, [srcq+2*gprsize] +movsrc3q, [srcq+3*gprsize] +movsrc4q, [srcq+4*gprsize] +movsrc5q, [srcq+5*gprsize] +mov srcq, [srcq] +subsrc1q, srcq +subsrc2q, srcq +subsrc3q, srcq +subsrc4q, srcq +subsrc5q, srcq +movaps xmm6, [pf_s16_scale] +.loop: +%if cpuflag(sse2) +mulps m0, m6, [srcq ] +mulps m1, m6, [srcq+src1q] +mulps m2, m6, [srcq+src2q] +mulps m3, m6, [srcq+src3q] +mulps m4, m6, [srcq+src4q] +mulps m5, m6, [srcq+src5q] +cvtps2dq m0, m0 +cvtps2dq m1, m1 +cvtps2dq m2, m2 +cvtps2dq m3, m3 +cvtps2dq m4, m4 +cvtps2dq m5, m5 +packssdw m0, m3; m0 = 0, 6, 12, 18, 3, 9, 15, 21 +packssdw m1, m4; m1 = 1, 7, 13, 19, 4, 10, 16, 22 +packssdw m2, m5; m2 = 2, 8, 14, 20, 5, 11, 17, 23 +; unpack words: +movhlps m3, m0; m3 = 3, 9, 15, 21, x, x, x, x +punpcklwd m0, m1; m0 = 0, 1, 6, 7, 12, 13, 18, 19 +punpckhwd m1, m2; m1 = 4, 5, 10, 11, 16, 17, 22, 23 +punpcklwd m2, m3; m2 = 2, 3, 8, 9, 14, 15, 20, 21 +; blend dwords: +shufpsm3, m0, m2, q2020 ; m3 = 0, 1, 12, 13, 2, 3, 14, 15 +shufpsm0, m1, q2031 ; m0 = 6, 7, 18, 19, 4, 5, 16, 17 +shufpsm2, m1, q3131 ; m2 = 8, 9, 20, 21, 10, 11, 22, 23 +; shuffle dwords: +shufpsm1, m2, m3, q3120 ; m1 = 8, 9, 10, 11, 12, 13, 14, 15 +shufpsm3, m0, q0220 ; m3 = 0, 1, 2, 3, 4, 5, 6, 7 +shufpsm0, m2, q3113 ; m0 = 16, 17, 18, 19, 20, 21, 22, 23 +mova [dstq+0*mmsize], m3 +mova [dstq+1*mmsize], m1 +mova [dstq+2*mmsize], m0 +%else ; sse +movaxmm0, [srcq ] +movaxmm1, [srcq+src1q] +movaxmm2, [srcq+src2q] +movaxmm3, [srcq+src3q] +movaxmm4, [srcq+src4q] +movaxmm5, [srcq+src5q] Even aside from the weirdness of seeing a mova that doesn't move the whole register, movq xmm, mem is a sse2 instruction. You need movlps. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 10/15] lavr: x86: optimized 2-channel s16 to fltp conversion
On Mon, 6 Aug 2012, Justin Ruggles wrote: --- libavresample/x86/audio_convert.asm| 39 libavresample/x86/audio_convert_init.c | 13 ++ 2 files changed, 52 insertions(+), 0 deletions(-) diff --git a/libavresample/x86/audio_convert.asm b/libavresample/x86/audio_convert.asm index 09c4e7f..ea0debf 100644 --- a/libavresample/x86/audio_convert.asm +++ b/libavresample/x86/audio_convert.asm @@ -923,3 +923,42 @@ CONV_S16_TO_S16P_6CH INIT_XMM avx CONV_S16_TO_S16P_6CH %endif + +;-- +; void ff_conv_s16_to_fltp_2ch(float *const *dst, int16_t *src, int len, +; int channels); +;-- + +%macro CONV_S16_TO_FLTP_2CH 0 +cglobal conv_s16_to_fltp_2ch, 3,4,4, dst0, src, len, dst1 +lea lenq, [4*lend] +mov dst1q, [dst0q+gprsize] +mov dst0q, [dst0q] +add srcq, lenq +add dst0q, lenq +add dst1q, lenq +neg lenq +movam3, [pf_s16_inv_scale] +.loop: +mova m0, [srcq+lenq] +S16_TO_S32_SX 0, 1 +cvtdq2ps m0, m0 +cvtdq2ps m1, m1 +mulps m0, m0, m3 +mulps m1, m1, m3 +DEINT2_PS 0, 1, 2 pslld m1, m0, 16 psrad m0, m0, 16 psrad m1, m1, 16 cvtdq2ps m0, m0 cvtdq2ps m1, m1 mulps m0, m0, m3 mulps m1, m1, m3 +mova [dst0q+lenq], m0 +mova [dst1q+lenq], m1 +add lenq, mmsize +jl .loop +REP_RET +%endmacro --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 07/45] x86: mmx2 --- mmxext in asm constructs
On Sat, 4 Aug 2012, Diego Biurrun wrote: On Sat, Aug 04, 2012 at 03:11:50PM -0400, Justin Ruggles wrote: On 07/31/2012 06:17 PM, Diego Biurrun wrote: --- libavcodec/x86/ac3dsp.asm |4 +- libavcodec/x86/ac3dsp_mmx.c|4 +- libavcodec/x86/dsputil_yasm.asm|8 ++-- libavcodec/x86/dsputilenc_mmx.c|6 +- libavcodec/x86/dsputilenc_yasm.asm |2 +- libavcodec/x86/h264_chromamc.asm | 14 +++--- libavcodec/x86/h264_chromamc_10bit.asm |4 +- libavcodec/x86/h264_deblock.asm| 20 libavcodec/x86/h264_deblock_10bit.asm |4 +- libavcodec/x86/h264_idct.asm | 46 + libavcodec/x86/h264_idct_10bit.asm |2 +- libavcodec/x86/h264_intrapred.asm | 26 +- libavcodec/x86/h264_intrapred_init.c | 40 +++--- libavcodec/x86/h264_weight.asm | 12 ++-- libavcodec/x86/pngdsp-init.c |6 +- libavcodec/x86/pngdsp.asm |2 +- libavcodec/x86/rv34dsp.asm |6 +- libavcodec/x86/rv34dsp_init.c | 10 ++-- libavcodec/x86/rv40dsp.asm |4 +- libavcodec/x86/rv40dsp_init.c | 30 ++-- libavcodec/x86/vc1dsp_mmx.c| 10 ++-- libavcodec/x86/vc1dsp_yasm.asm | 10 ++-- libavcodec/x86/vp3dsp.asm |4 +- libavcodec/x86/vp3dsp_init.c | 16 +++--- libavcodec/x86/vp8dsp-init.c | 86 libavcodec/x86/vp8dsp.asm | 30 ++-- libavutil/x86/x86inc.asm |6 +- libavutil/x86/x86util.asm |2 +- libswscale/x86/output.asm |4 +- libswscale/x86/swscale.c |8 ++-- 30 files changed, 215 insertions(+), 211 deletions(-) Looks ok, but probably should get other opinions on this as well. I know Ronald was trying to keep x86inc.asm sychronized with x264, and trying to do so after this change would likely require similar extensive cpuflag modifications in x264. I volunteer to patch x264 if such a change would be accepted on their side. Rejected. I like mmx2 better. However, I wouldn't be opposed to dropping mmx1 entirely and using the name mmx to refer to mmx2. (x264 doesn't actually support mmx1 anyway; we use mmx2 inline asm that's actually inlined in places where runtime cpu detection is impossible.) But that wouldn't help synchronization if libav doesn't do so. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] dsputil: make add_hfyu_left_prediction_sse4() support unaligned src.
On Fri, 3 Aug 2012, Hendrik Leppkes wrote: From: Ronald S. Bultje rsbul...@gmail.com This makes add_hfyu_left_prediction_sse4() handle sources that are not 16-byte aligned in its own function rather than by proxying the call to add_hfyu_left_prediction_ssse3(). This fixes a crash on Win64, since the sse4 version clobberes xmm6, but the ssse3 version (which uses MMX regs) does not restore it, thus leading to XMM clobbering and RSP being off. Fixes bug 342. --- The previous patch assumed that dst would be aligned if src was not, causing another crash. Instead, assume both are not aligned. libavcodec/x86/dsputil_yasm.asm | 14 ++ 1 file changed, 10 insertions(+), 4 deletions(-) diff --git a/libavcodec/x86/dsputil_yasm.asm b/libavcodec/x86/dsputil_yasm.asm index 70a0aa1..58e4ca0 100644 --- a/libavcodec/x86/dsputil_yasm.asm +++ b/libavcodec/x86/dsputil_yasm.asm @@ -388,12 +388,16 @@ cglobal add_hfyu_median_prediction_mmx2, 6,6,0, dst, top, diff, w, left, left_to RET -%macro ADD_HFYU_LEFT_LOOP 1 ; %1 = is_aligned +%macro ADD_HFYU_LEFT_LOOP 1-2 ; %1 = is_aligned, %2 = src_is_unaligned The two have opposite semantics? And there's no reason to use variable number of arguments here. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] dsputil: make add_hfyu_left_prediction_sse4() support unaligned src.
On Fri, 3 Aug 2012, Ronald S. Bultje wrote: From: Ronald S. Bultje rsbul...@gmail.com This makes add_hfyu_left_prediction_sse4() handle sources that are not 16-byte aligned in its own function rather than by proxying the call to add_hfyu_left_prediction_ssse3(). This fixes a crash on Win64, since the sse4 version clobberes xmm6, but the ssse3 version (which uses MMX regs) does not restore it, thus leading to XMM clobbering and RSP being off. Fixes bug 342. --- libavcodec/x86/dsputil_yasm.asm | 20 +--- 1 file changed, 13 insertions(+), 7 deletions(-) LGTM --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86: Refactor PSWAPD fallback implementations and port to cpuflags
On Thu, 2 Aug 2012, Diego Biurrun wrote:
On Wed, Aug 01, 2012 at 07:41:01AM -0700, Ronald S. Bultje wrote:
On Wed, Aug 1, 2012 at 5:30 AM, Diego Biurrun di...@biurrun.de wrote:
--- a/libavcodec/x86/fft_mmx.asm
+++ b/libavcodec/x86/fft_mmx.asm
@@ -105,7 +105,8 @@ SECTION_TEXT
pxor %3, [ps_m1p1] ; {t8,t7}
mova %6, %1
-pswapd %3, %3
+movd [r0+12], %3
+punpckhdq %3, [r0+8]
Needs rebase?
No, it's the only caller with two identical arguments, so I moved the
macro branch for identical arguments out of the macro and into the caller.
But then it doesn't generate pswapd in the 3dnow2 instantiation.
Or you could make it not have two identical arguments; there's a spare
mmreg (%4) at that point.
--Loren Merritt
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 07/15] lavr: x86: optimized 2-channel fltp to flt conversion
On Mon, 30 Jul 2012, Justin Ruggles wrote: --- a/libavutil/x86/x86util.asm +++ b/libavutil/x86/x86util.asm @@ -47,6 +47,12 @@ SWAP %1, %3, %2 %endmacro +%macro SBUTTERFLYPS2 3 +unpcklps m%3, m%1, m%2 +unpckhps m%1, m%1, m%2 +SWAP %1, %3, %2 +%endmacro What's the difference between this and SBUTTERFLYPS? --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 43/45] x86: Refactor PSWAPD fallback implementations and port to cpuflags
On Wed, 1 Aug 2012, Diego Biurrun wrote: diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm index 6911a9c..f358257 100644 --- a/libavutil/x86/x86util.asm +++ b/libavutil/x86/x86util.asm @@ -306,6 +306,23 @@ %endif %endmacro +%macro PSWAPD 2 +%if cpuflag(sse) +pshufw %1, %2, 0x4e That's mmx2. The macro was called sse because it was in a float function which thus didn't have a mmx2 version. +%elif cpuflag(3dnowext) +pswapd %1, %2 +%else +%ifidn %1, %2 +movd [r0+12], %1 +punpckhdq %1, [r0+8] +%else +movq %1, %2 +psrlq %1, 32 +punpckldq %1, %2 +%endif +%endif +%endmacro --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/2] h264: remove 3dnow qpel code
On Wed, 1 Aug 2012, Luca Barbato wrote: From: Daniel Kang daniel.d.k...@gmail.com Remove the code to eases porting the other qpel optimizations to yasm. AMD has deprecated 3dnow and the only CPUs that have 3dnow and do not have mmxext are 12 years old. libavcodec/x86/dsputil_mmx_avg_template.c:58:1: warning: `put_pixels4_l2_3dnow' defined but not used libavcodec/x86/dsputil_mmx_avg_template.c:229:1: warning: `avg_pixels4_l2_3dnow' defined but not used libavcodec/x86/dsputil_mmx_avg_template.c:875:1: warning: `avg_pixels4_3dnow' defined but not used --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 2/2] h264: convert 8-bit qpel inlined assembly to yasm
On Wed, 1 Aug 2012, Luca Barbato wrote: +%macro OP_MOVH_MMX 3 +movh %3, %2 +pavgb %1, %3 +movh %2, %1 +%endmacro + +%macro MOVH_MMX 3 +movh %2, %1 +%endmacro + +%macro OP_MOV_MMX 3 +mova %3, %2 +pavgb %1, %3 +mova %2, %1 pavgb %1, %2 mova %2, %1 (Just for the full width one) +%endmacro + +%macro MOV_MMX 3 +mova %2, %1 +%endmacro It's op_put vs op_avg (or mov vs avg), not mov vs op_mov. Plus, naming them put vs avg would allow you to exploit the same put vs avg that's already in all the function names, rather than a separate %define OP. +%macro QPEL8OR16_V_LOWPASS_OP 1 +cglobal %1_h264_qpel8or16_v_lowpass, 5,5,7 ; dst, src, dstStride, srcStride, h +%if cpuflag(sse2) +sub r1, r3 +sub r1, r3 +%endif +pxor m7, m7 +movh m0, [r1] +movh m1, [r1+r3] +lea r1, [r1+2*r3] +movh m2, [r1] +movh m3, [r1+r3] +lea r1, [r1+2*r3] +movh m4, [r1] +add r1, r3 +punpcklbw m0, m7 +punpcklbw m1, m7 +punpcklbw m2, m7 +punpcklbw m3, m7 +punpcklbw m4, m7 +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +cmp r4d, 16 +jne .end +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +FILT_V +.end: +RET +%endmacro (and other cases of this) REP_RET ... and I'll skip the suggestions for improvement, since Daniel Kang has a separate branch for that. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 1/2] fft: port FFT and IMDCT functions for 3dnow to yasm.
On Sun, 29 Jul 2012, Ronald S. Bultje wrote: +%macro PSWAPD 2 +%if cpuflag(3dnow2) +pswapd%1, %2 +%else +mova %1, %2 +psrlq %1, 32 +punpckldq %1, %2 +%endif +%endmacro Merge with the macro named pswapd. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] h264: convert loop filter strength dsp function to yasm.
On Sat, 28 Jul 2012, Ronald S. Bultje wrote: On Fri, Jul 27, 2012 at 1:01 PM, Loren Merritt lor...@u.washington.edu wrote: %%.b_idx_loop: Automatically generates a different label for each instantiation of the macro. My disassembly now looks like this: 0x0001004c43b4 ff_h264_loop_filter_strength_mmx2.nofield+25:jne 0x1004c44d1 ff_h264_loop_filter_strength_mmx2.bidir 0x0001004c43ba ff_h264_loop_filter_strength_mmx2.nofield+31:xor %r8d,%r8d 0x0001004c43bd ..@5001..b_idx_loop+0: pxor %mm0,%mm0 0x0001004c43c0 ..@5001..b_idx_loop+3: test %r11d,%r8d Can I somehow keep the function name in it? I find that somewhat useful when debugging. Nope. And I've patched my objdump to remove function names prefixed to local labels. The bigger problem is that gdb doesn't know the difference between local labels and functions, so disassembly will fail to encompass a whole function if it contains any labels, regardless of what they're named. For this reason, I sometimes find it easier to `strip -x` before using gdb. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 7/7] vf_hqdn3d: x86 asm
13% faster on penryn, 16% on sandybridge, 15% on bulldozer
Not simd; a compiler should have generated this, but gcc didn't.
---
libavfilter/vf_hqdn3d.c| 27 ++--
libavfilter/x86/Makefile |1 +
libavfilter/x86/hqdn3d.asm | 106
libavutil/x86/x86inc.asm |1 +
4 files changed, 131 insertions(+), 4 deletions(-)
create mode 100644 libavfilter/x86/hqdn3d.asm
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index d263cff..ef59691 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -40,8 +40,14 @@ typedef struct {
double strength[4];
int hsub, vsub;
int depth;
+void (*denoise_row[17])(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
} HQDN3DContext;
+void ff_hqdn3d_row_8_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+void ff_hqdn3d_row_9_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+void ff_hqdn3d_row_10_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+void ff_hqdn3d_row_16_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+
#define LUT_BITS (depth==16 ? 8 : 4)
#define RIGHTSHIFT(a,b) (((a)+(((1(b))-1)1))(b))
#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) (16-depth))
@@ -78,7 +84,8 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
av_always_inline
-static void denoise_spatial(uint8_t *src, uint8_t *dst,
+static void denoise_spatial(HQDN3DContext *hqdn3d,
+uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
int16_t *spatial, int16_t *temporal, int depth)
@@ -103,6 +110,10 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
src += sstride;
dst += dstride;
frame_ant += w;
+if (hqdn3d-denoise_row[depth]) {
+hqdn3d-denoise_row[depth](src, dst, line_ant, frame_ant, w,
spatial, temporal);
+continue;
+}
pixel_ant = LOAD(0);
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial,
depth);
@@ -117,7 +128,8 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
}
av_always_inline
-static void denoise_depth(uint8_t *src, uint8_t *dst,
+static void denoise_depth(HQDN3DContext *hqdn3d,
+ uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
int16_t *spatial, int16_t *temporal, int depth)
@@ -137,7 +149,7 @@ static void denoise_depth(uint8_t *src, uint8_t *dst,
}
if (spatial[0])
-denoise_spatial(src, dst, line_ant, frame_ant,
+denoise_spatial(hqdn3d, src, dst, line_ant, frame_ant,
w, h, sstride, dstride, spatial, temporal, depth);
else
denoise_temporal(src, dst, frame_ant,
@@ -297,6 +309,13 @@ static int config_input(AVFilterLink *inlink)
}
}
+#if HAVE_YASM
+hqdn3d-denoise_row[ 8] = ff_hqdn3d_row_8_x86;
+hqdn3d-denoise_row[ 9] = ff_hqdn3d_row_9_x86;
+hqdn3d-denoise_row[10] = ff_hqdn3d_row_10_x86;
+hqdn3d-denoise_row[16] = ff_hqdn3d_row_16_x86;
+#endif
+
return 0;
}
@@ -314,7 +333,7 @@ static int end_frame(AVFilterLink *inlink)
int ret, c;
for (c = 0; c 3; c++) {
-denoise(inpic-data[c], outpic-data[c],
+denoise(hqdn3d, inpic-data[c], outpic-data[c],
hqdn3d-line, hqdn3d-frame_prev[c],
inpic-video-w (!!c * hqdn3d-hsub),
inpic-video-h (!!c * hqdn3d-vsub),
diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
index e98693d..46fc84f 100644
--- a/libavfilter/x86/Makefile
+++ b/libavfilter/x86/Makefile
@@ -1,2 +1,3 @@
MMX-OBJS-$(CONFIG_YADIF_FILTER) += x86/yadif.o
MMX-OBJS-$(CONFIG_GRADFUN_FILTER)+= x86/gradfun.o
+YASM-OBJS-$(CONFIG_HQDN3D_FILTER)+= x86/hqdn3d.o
diff --git a/libavfilter/x86/hqdn3d.asm b/libavfilter/x86/hqdn3d.asm
new file mode 100644
index 000..7254194
--- /dev/null
+++ b/libavfilter/x86/hqdn3d.asm
@@ -0,0 +1,106 @@
+;**
+;* Copyright (c) 2012 Loren Merritt
+;*
+;* This file is part of Libav.
+;*
+;* Libav is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1
Re: [libav-devel] [PATCH] h264_loopfilter: port x86 simd to cpuflags.
On Thu, 26 Jul 2012, Ronald S. Bultje wrote: From: Ronald S. Bultje rsbul...@gmail.com --- libavcodec/x86/h264_deblock.asm | 124 +++-- libavcodec/x86/h264_deblock_10bit.asm | 77 ++-- libavcodec/x86/h264dsp_mmx.c | 60 3 files changed, 139 insertions(+), 122 deletions(-) diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm index 1982dc4..b5e81e7 100644 --- a/libavcodec/x86/h264_deblock.asm +++ b/libavcodec/x86/h264_deblock.asm @@ -282,8 +282,8 @@ cextern pb_A1 ;- ; void deblock_v_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) ;- -%macro DEBLOCK_LUMA 1 -cglobal deblock_v_luma_8_%1, 5,5,10 +%macro DEBLOCK_V_LUMA 0 +cglobal deblock_v_luma_8, 5,5,10 movdm8, [r4] ; tc0 lea r4, [r1*3] dec r2d; alpha-1 @@ -323,12 +323,13 @@ cglobal deblock_v_luma_8_%1, 5,5,10 mova[r4+2*r1], m1 mova[r0], m2 RET +%endmacro No indent. ;- ; void deblock_h_luma( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 ) ;- -INIT_MMX -cglobal deblock_h_luma_8_%1, 5,9 +%macro DEBLOCK_H_LUMA 0 +cglobal deblock_h_luma_8, 5,9 movsxd r7, r1d lear8, [r7+r7*2] lear6, [r0-4] @@ -355,7 +356,7 @@ cglobal deblock_h_luma_8_%1, 5,9 %if WIN64 mov[rsp+0x20], r4 %endif -call deblock_v_luma_8_%1 +call deblock_v_luma_8 %+ SUFFIX call automatically appends SUFFIX if there is a function by that name. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] h264: convert loop filter strength dsp function to yasm.
On Fri, 27 Jul 2012, Ronald S. Bultje wrote: From: Ronald S. Bultje rsbul...@gmail.com This completes the conversion of h264dsp to yasm; note that h264 also uses some dsputil functions, most notably qpel. Performance-wise, the yasm-version is ~10 cycles faster (182-172) on x86-64, and ~8 cycles faster (201-193) on x86-32. --- libavcodec/x86/h264_deblock.asm | 168 +++ libavcodec/x86/h264dsp_mmx.c| 162 ++--- 2 files changed, 175 insertions(+), 155 deletions(-) diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm index 1982dc4..77b25d2 100644 --- a/libavcodec/x86/h264_deblock.asm +++ b/libavcodec/x86/h264_deblock.asm @@ -27,6 +27,10 @@ %include x86inc.asm %include x86util.asm +SECTION_RODATA + +pb_3_1: times 4 db 3, 1 + SECTION .text cextern pb_0 @@ -911,3 +915,167 @@ ff_chroma_intra_body_mmxext: paddb m1, m5 paddb m2, m6 ret + +;- +; void h264_loop_filter_strength(int16_t bs[2][4][4], uint8_t nnz[40], +;int8_t ref[2][40], int16_t mv[2][40][2], +;int bidir,int edges,int step, +;int mask_mv0, int mask_mv1, int field); +; +; bidiris 0 or 1 +; edgesis 1 or 4 +; step is 1 or 2 +; mask_mv0 is 0 or 3 +; mask_mv1 is 0 or 1 +; fieldis 0 or 1 +;- +%macro loop_filter_strength_iteration 7 ; edges, step, mask_mv, +; dir, d_idx, mask_dir, bidir +%define edgesd%1 +%define stepd %2 +%define mask_mvd %3 +%define dir %4 +%define d_idx %5 +%define mask_dir %6 +%define bidir %7 +xor b_idxd, b_idxd ; for (b_idx = 0; b_idx edges; b_idx += step) +.b_idx_loop_ %+ dir %+ _ %+ bidir: %%.b_idx_loop: Automatically generates a different label for each instantiation of the macro. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: automatically insert vzeroupper for YMM functions.
On Wed, 25 Jul 2012, Ronald S. Bultje wrote: diff --git a/libavcodec/x86/dct32_sse.asm b/libavcodec/x86/dct32_sse.asm index e3c8a45..351c88d 100644 --- a/libavcodec/x86/dct32_sse.asm +++ b/libavcodec/x86/dct32_sse.asm @@ -278,8 +278,6 @@ cglobal dct32_float_avx, 2,3,8, out, in, tmp vperm2f128 m0, m1, m1, 0x31 vmovaps [outq+96], m1 -vzeroupper - ;pass 6, no SIMD... INIT_XMM PASS6_AND_PERMUTE This one is followed by xmm instructions in the same function, so moving it to RET doesn't work. And the INIT_XMM means it doesn't happen at RET either. So just don't remove it. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 1/7] factor identical ff_inplace_start_frame out of two filters
---
libavfilter/vf_delogo.c | 38 +-
libavfilter/vf_gradfun.c | 37 +
libavfilter/video.c | 36
libavfilter/video.h |1 +
4 files changed, 39 insertions(+), 73 deletions(-)
diff --git a/libavfilter/vf_delogo.c b/libavfilter/vf_delogo.c
index 343585f..e8b5431 100644
--- a/libavfilter/vf_delogo.c
+++ b/libavfilter/vf_delogo.c
@@ -214,42 +214,6 @@ static av_cold int init(AVFilterContext *ctx, const char
*args)
return 0;
}
-static int start_frame(AVFilterLink *inlink, AVFilterBufferRef *inpicref)
-{
-AVFilterLink *outlink = inlink-dst-outputs[0];
-AVFilterBufferRef *outpicref = NULL, *for_next_filter;
-int ret = 0;
-
-if (inpicref-perms AV_PERM_PRESERVE) {
-outpicref = ff_get_video_buffer(outlink, AV_PERM_WRITE,
-outlink-w, outlink-h);
-if (!outpicref)
-return AVERROR(ENOMEM);
-
-avfilter_copy_buffer_ref_props(outpicref, inpicref);
-outpicref-video-w = outlink-w;
-outpicref-video-h = outlink-h;
-} else {
-outpicref = avfilter_ref_buffer(inpicref, ~0);
-if (!outpicref)
-return AVERROR(ENOMEM);
-}
-
-for_next_filter = avfilter_ref_buffer(outpicref, ~0);
-if (for_next_filter)
-ret = ff_start_frame(outlink, for_next_filter);
-else
-ret = AVERROR(ENOMEM);
-
-if (ret 0) {
-avfilter_unref_bufferp(outpicref);
-return ret;
-}
-
-outlink-out_buf = outpicref;
-return 0;
-}
-
static int null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir)
{
return 0;
@@ -296,7 +260,7 @@ AVFilter avfilter_vf_delogo = {
.inputs= (const AVFilterPad[]) {{ .name = default,
.type =
AVMEDIA_TYPE_VIDEO,
.get_video_buffer =
ff_null_get_video_buffer,
- .start_frame = start_frame,
+ .start_frame =
ff_inplace_start_frame,
.draw_slice = null_draw_slice,
.end_frame= end_frame,
.min_perms= AV_PERM_WRITE |
AV_PERM_READ,
diff --git a/libavfilter/vf_gradfun.c b/libavfilter/vf_gradfun.c
index 52dcb70..f8896b5 100644
--- a/libavfilter/vf_gradfun.c
+++ b/libavfilter/vf_gradfun.c
@@ -180,41 +180,6 @@ static int config_input(AVFilterLink *inlink)
return 0;
}
-static int start_frame(AVFilterLink *inlink, AVFilterBufferRef *inpicref)
-{
-AVFilterLink *outlink = inlink-dst-outputs[0];
-AVFilterBufferRef *outpicref = NULL, *for_next_filter;
-int ret = 0;
-
-if (inpicref-perms AV_PERM_PRESERVE) {
-outpicref = ff_get_video_buffer(outlink, AV_PERM_WRITE, outlink-w,
outlink-h);
-if (!outpicref)
-return AVERROR(ENOMEM);
-
-avfilter_copy_buffer_ref_props(outpicref, inpicref);
-outpicref-video-w = outlink-w;
-outpicref-video-h = outlink-h;
-} else {
-outpicref = avfilter_ref_buffer(inpicref, ~0);
-if (!outpicref)
-return AVERROR(ENOMEM);
-}
-
-for_next_filter = avfilter_ref_buffer(outpicref, ~0);
-if (for_next_filter)
-ret = ff_start_frame(outlink, for_next_filter);
-else
-ret = AVERROR(ENOMEM);
-
-if (ret 0) {
-avfilter_unref_bufferp(outpicref);
-return ret;
-}
-
-outlink-out_buf = outpicref;
-return 0;
-}
-
static int null_draw_slice(AVFilterLink *link, int y, int h, int slice_dir)
{
return 0;
@@ -261,7 +226,7 @@ AVFilter avfilter_vf_gradfun = {
.inputs= (const AVFilterPad[]) {{ .name = default,
.type =
AVMEDIA_TYPE_VIDEO,
.config_props = config_input,
- .start_frame = start_frame,
+ .start_frame =
ff_inplace_start_frame,
.draw_slice = null_draw_slice,
.end_frame= end_frame,
.min_perms= AV_PERM_READ, },
diff --git a/libavfilter/video.c b/libavfilter/video.c
index 6e50637..ebbbc34 100644
--- a/libavfilter/video.c
+++ b/libavfilter/video.c
@@ -168,6 +168,42 @@ int ff_null_start_frame(AVFilterLink *link,
AVFilterBufferRef *picref)
return ff_start_frame(link-dst-outputs[0], buf_out);
}
+// for filters that support (but don't require) outpic==inpic
+int ff_inplace_start_frame(AVFilterLink *inlink, AVFilterBufferRef *inpicref)
+{
+AVFilterLink *outlink = inlink-dst-outputs[0];
+AVFilterBufferRef
[libav-devel] [PATCH 2/7] vf_hqdn3d: cosmetics
Change code style to match the rest of libav.
---
libavfilter/vf_hqdn3d.c | 308 +++
1 files changed, 152 insertions(+), 156 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 9e01606..1fa70c6 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -32,165 +32,161 @@
#include video.h
typedef struct {
-int Coefs[4][512*16];
-unsigned int *Line;
-unsigned short *Frame[3];
+int coefs[4][512*16];
+uint32_t *line;
+uint16_t *frame_prev[3];
int hsub, vsub;
} HQDN3DContext;
-static inline unsigned int LowPassMul(unsigned int PrevMul, unsigned int
CurrMul, int *Coef)
+static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
{
-//int dMul= (PrevMul0xFF)-(CurrMul0xFF);
-int dMul= PrevMul-CurrMul;
-unsigned int d=((dMul+0x10007FF)12);
-return CurrMul + Coef[d];
+int dmul = prev-cur;
+unsigned int d = (dmul+0x10007FF)12; // 0x1000 to convert to unsigned,
7FF for rounding
+return cur + coef[d];
}
-static void deNoiseTemporal(unsigned char *FrameSrc,
-unsigned char *FrameDest,
-unsigned short *FrameAnt,
-int W, int H, int sStride, int dStride,
-int *Temporal)
+static void denoise_temporal(uint8_t *src, uint8_t *dst,
+ uint16_t *frame_ant,
+ int w, int h, int sstride, int dstride,
+ int *temporal)
{
-long X, Y;
-unsigned int PixelDst;
-
-for (Y = 0; Y H; Y++) {
-for (X = 0; X W; X++) {
-PixelDst = LowPassMul(FrameAnt[X]8, FrameSrc[X]16, Temporal);
-FrameAnt[X] = ((PixelDst+0x107F)8);
-FrameDest[X]= ((PixelDst+0x10007FFF)16);
+long x, y;
+uint32_t pixel;
+
+for (y = 0; y h; y++) {
+for (x = 0; x w; x++) {
+pixel = lowpass(frame_ant[x]8, src[x]16, temporal);
+frame_ant[x] = ((pixel+0x107F)8);
+dst[x]= ((pixel+0x10007FFF)16);
}
-FrameSrc += sStride;
-FrameDest += dStride;
-FrameAnt += W;
+src += sstride;
+dst += dstride;
+frame_ant += w;
}
}
-static void deNoiseSpacial(unsigned char *Frame,
- unsigned char *FrameDest,
- unsigned int *LineAnt,
- int W, int H, int sStride, int dStride,
- int *Horizontal, int *Vertical)
+static void denoise_spatial(uint8_t *src, uint8_t *dst,
+uint32_t *line_ant,
+int w, int h, int sstride, int dstride,
+int *horizontal, int *vertical)
{
-long X, Y;
-long sLineOffs = 0, dLineOffs = 0;
-unsigned int PixelAnt;
-unsigned int PixelDst;
+long x, y;
+long sline_offs = 0, dline_offs = 0;
+uint32_t pixel_ant;
+uint32_t pixel;
/* First pixel has no left nor top neighbor. */
-PixelDst = LineAnt[0] = PixelAnt = Frame[0]16;
-FrameDest[0]= ((PixelDst+0x10007FFF)16);
+pixel = line_ant[0] = pixel_ant = src[0]16;
+dst[0]= ((pixel+0x10007FFF)16);
/* First line has no top neighbor, only left. */
-for (X = 1; X W; X++) {
-PixelDst = LineAnt[X] = LowPassMul(PixelAnt, Frame[X]16, Horizontal);
-FrameDest[X]= ((PixelDst+0x10007FFF)16);
+for (x = 1; x w; x++) {
+pixel = line_ant[x] = lowpass(pixel_ant, src[x]16, horizontal);
+dst[x]= ((pixel+0x10007FFF)16);
}
-for (Y = 1; Y H; Y++) {
-unsigned int PixelAnt;
-sLineOffs += sStride, dLineOffs += dStride;
+for (y = 1; y h; y++) {
+uint32_t pixel_ant;
+sline_offs += sstride, dline_offs += dstride;
/* First pixel on each line doesn't have previous pixel */
-PixelAnt = Frame[sLineOffs]16;
-PixelDst = LineAnt[0] = LowPassMul(LineAnt[0], PixelAnt, Vertical);
-FrameDest[dLineOffs]= ((PixelDst+0x10007FFF)16);
+pixel_ant = src[sline_offs]16;
+pixel = line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
+dst[dline_offs]= ((pixel+0x10007FFF)16);
-for (X = 1; X W; X++) {
-unsigned int PixelDst;
+for (x = 1; x w; x++) {
+uint32_t pixel;
/* The rest are normal */
-PixelAnt = LowPassMul(PixelAnt, Frame[sLineOffs+X]16,
Horizontal);
-PixelDst = LineAnt[X] = LowPassMul(LineAnt[X], PixelAnt, Vertical);
-FrameDest[dLineOffs+X]= ((PixelDst+0x10007FFF)16);
+pixel_ant = lowpass(pixel_ant, src[sline_offs+x]16, horizontal);
+pixel = line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
+dst[dline_offs+x]= ((pixel+0x10007FFF)16);
}
}
}
-static void deNoise(unsigned char *Frame,
-
[libav-devel] [PATCH 3/7] vf_hqdn3d: simplify and optimize
14% faster on penryn, 2% on sandybridge, 9% on bulldozer
---
libavfilter/vf_hqdn3d.c | 157 +++---
1 files changed, 51 insertions(+), 106 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 1fa70c6..bb81b79 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -1,6 +1,7 @@
/*
* Copyright (c) 2003 Daniel Moreno comac AT comac DOT darktech DOT org
* Copyright (c) 2010 Baptiste Coudurier
+ * Copyright (c) 2012 Loren Merritt
*
* This file is part of Libav, ported from MPlayer.
*
@@ -51,13 +52,13 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
int *temporal)
{
long x, y;
-uint32_t pixel;
+uint32_t tmp;
for (y = 0; y h; y++) {
for (x = 0; x w; x++) {
-pixel = lowpass(frame_ant[x]8, src[x]16, temporal);
-frame_ant[x] = ((pixel+0x107F)8);
-dst[x]= ((pixel+0x10007FFF)16);
+tmp = lowpass(frame_ant[x]8, src[x]16, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
src += sstride;
dst += dstride;
@@ -66,111 +67,66 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
static void denoise_spatial(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant,
+uint32_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int *horizontal, int *vertical)
+int *spatial, int *temporal)
{
long x, y;
-long sline_offs = 0, dline_offs = 0;
uint32_t pixel_ant;
-uint32_t pixel;
+uint32_t tmp;
-/* First pixel has no left nor top neighbor. */
-pixel = line_ant[0] = pixel_ant = src[0]16;
-dst[0]= ((pixel+0x10007FFF)16);
-
-/* First line has no top neighbor, only left. */
-for (x = 1; x w; x++) {
-pixel = line_ant[x] = lowpass(pixel_ant, src[x]16, horizontal);
-dst[x]= ((pixel+0x10007FFF)16);
+/* First line has no top neighbor. Only left one for each tmp and
+ * last frame */
+pixel_ant = src[0]16;
+for (x = 0; x w; x++) {
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]16,
spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
for (y = 1; y h; y++) {
-uint32_t pixel_ant;
-sline_offs += sstride, dline_offs += dstride;
-/* First pixel on each line doesn't have previous pixel */
-pixel_ant = src[sline_offs]16;
-pixel = line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
-dst[dline_offs]= ((pixel+0x10007FFF)16);
-
-for (x = 1; x w; x++) {
-uint32_t pixel;
-/* The rest are normal */
-pixel_ant = lowpass(pixel_ant, src[sline_offs+x]16, horizontal);
-pixel = line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
-dst[dline_offs+x]= ((pixel+0x10007FFF)16);
+src += sstride;
+dst += dstride;
+frame_ant += w;
+pixel_ant = src[0]16;
+for (x = 0; x w-1; x++) {
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
+pixel_ant = lowpass(pixel_ant, src[x+1]16, spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
}
static void denoise(uint8_t *src, uint8_t *dst,
uint32_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
-int *horizontal, int *vertical, int *temporal)
+int *spatial, int *temporal)
{
long x, y;
-long sline_offs = 0, dline_offs = 0;
-uint32_t pixel_ant;
-uint32_t pixel;
uint16_t *frame_ant = *frame_ant_ptr;
-
if (!frame_ant) {
+uint8_t *frame_src = src;
*frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
-for (y = 0; y h; y++) {
-uint16_t *frame_dst = frame_ant+y*w;
-uint8_t *frame_src = src+y*sstride;
+for (y = 0; y h; y++, src += sstride, frame_ant += w)
for (x = 0; x w; x++)
-frame_dst[x] = frame_src[x]8;
-}
+frame_ant[x] = src[x]8;
+src = frame_src;
+frame_ant = *frame_ant_ptr;
}
-if (!horizontal[0] !vertical[0]) {
+if (spatial[0])
+denoise_spatial(src, dst, line_ant, frame_ant,
+w, h, sstride, dstride, spatial, temporal);
+else
denoise_temporal(src, dst, frame_ant
[libav-devel] [PATCH 5/7] vf_hqdn3d: support 10bit colordepth
---
libavfilter/vf_hqdn3d.c | 68 +-
1 files changed, 49 insertions(+), 19 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 46902fa..02f0cd1 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -27,6 +27,7 @@
*/
#include libavutil/pixdesc.h
+#include libavutil/intreadwrite.h
#include avfilter.h
#include formats.h
#include internal.h
@@ -37,18 +38,25 @@ typedef struct {
uint16_t *line;
uint16_t *frame_prev[3];
int hsub, vsub;
+int depth;
} HQDN3DContext;
+#define RIGHTSHIFT(a,b) (((a)+(((1(b))-1)1))(b))
+#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) (16-depth))
+#define STORE(x,val) (depth==8 ? dst[x] = RIGHTSHIFT(val, 16-depth)\
+: AV_WN16A(dst+(x)*2, RIGHTSHIFT(val, 16-depth)))
+
static inline uint32_t lowpass(int prev, int cur, int16_t *coef)
{
int d = (prev-cur)4;
return cur + coef[d];
}
+av_always_inline
static void denoise_temporal(uint8_t *src, uint8_t *dst,
uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int16_t *temporal)
+ int16_t *temporal, int depth)
{
long x, y;
uint32_t tmp;
@@ -57,8 +65,8 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
for (y = 0; y h; y++) {
for (x = 0; x w; x++) {
-frame_ant[x] = tmp = lowpass(frame_ant[x], src[x]8, temporal);
-dst[x] = (tmp+0x7F)8;
+frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal);
+STORE(x, tmp);
}
src += sstride;
dst += dstride;
@@ -66,10 +74,11 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
}
+av_always_inline
static void denoise_spatial(uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int16_t *spatial, int16_t *temporal)
+int16_t *spatial, int16_t *temporal, int depth)
{
long x, y;
uint32_t pixel_ant;
@@ -80,34 +89,35 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
/* First line has no top neighbor. Only left one for each tmp and
* last frame */
-pixel_ant = src[0]8;
+pixel_ant = LOAD(0);
for (x = 0; x w; x++) {
-line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]8, spatial);
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
for (y = 1; y h; y++) {
src += sstride;
dst += dstride;
frame_ant += w;
-pixel_ant = src[0]8;
+pixel_ant = LOAD(0);
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-pixel_ant = lowpass(pixel_ant, src[x+1]8, spatial);
+pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
}
-static void denoise(uint8_t *src, uint8_t *dst,
-uint16_t *line_ant, uint16_t **frame_ant_ptr,
-int w, int h, int sstride, int dstride,
-int16_t *spatial, int16_t *temporal)
+av_always_inline
+static void denoise_depth(uint8_t *src, uint8_t *dst,
+ uint16_t *line_ant, uint16_t **frame_ant_ptr,
+ int w, int h, int sstride, int dstride,
+ int16_t *spatial, int16_t *temporal, int depth)
{
long x, y;
uint16_t *frame_ant = *frame_ant_ptr;
@@ -116,19 +126,25 @@ static void denoise(uint8_t *src, uint8_t *dst,
*frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
for (y = 0; y h; y++, src += sstride, frame_ant += w)
for (x = 0; x w; x++)
-frame_ant[x] = src[x]8;
+frame_ant[x] = LOAD(x);
src = frame_src;
frame_ant = *frame_ant_ptr;
}
if (spatial[0])
denoise_spatial(src, dst, line_ant, frame_ant,
-w, h, sstride, dstride, spatial, temporal);
+w, h, sstride, dstride, spatial, temporal, depth);
else
denoise_temporal(src, dst, frame_ant,
- w, h, sstride, dstride, temporal);
+ w, h, sstride, dstride, temporal, depth);
}
+#define denoise(...) \
+switch (hqdn3d-depth) {\
+case 8:
[libav-devel] [PATCH 6/7] vf_hqdn3d: support 16bit colordepth
---
libavfilter/vf_hqdn3d.c | 70 +++
1 files changed, 46 insertions(+), 24 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 02f0cd1..521fef8 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -34,21 +34,24 @@
#include video.h
typedef struct {
-int16_t coefs[4][512*16];
+int16_t *coefs[4];
uint16_t *line;
uint16_t *frame_prev[3];
+double strength[4];
int hsub, vsub;
int depth;
} HQDN3DContext;
+#define LUT_BITS (depth==16 ? 8 : 4)
#define RIGHTSHIFT(a,b) (((a)+(((1(b))-1)1))(b))
#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) (16-depth))
#define STORE(x,val) (depth==8 ? dst[x] = RIGHTSHIFT(val, 16-depth)\
: AV_WN16A(dst+(x)*2, RIGHTSHIFT(val, 16-depth)))
-static inline uint32_t lowpass(int prev, int cur, int16_t *coef)
+av_always_inline
+static inline uint32_t lowpass(int prev, int cur, int16_t *coef, int depth)
{
-int d = (prev-cur)4;
+int d = (prev - cur) (8 - LUT_BITS);
return cur + coef[d];
}
@@ -61,11 +64,11 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
long x, y;
uint32_t tmp;
-temporal += 0x1000;
+temporal += 256 LUT_BITS;
for (y = 0; y h; y++) {
for (x = 0; x w; x++) {
-frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal);
+frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal,
depth);
STORE(x, tmp);
}
src += sstride;
@@ -84,15 +87,15 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
uint32_t pixel_ant;
uint32_t tmp;
-spatial += 0x1000;
-temporal += 0x1000;
+spatial += 256 LUT_BITS;
+temporal += 256 LUT_BITS;
/* First line has no top neighbor. Only left one for each tmp and
* last frame */
pixel_ant = LOAD(0);
for (x = 0; x w; x++) {
-line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial);
-frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial,
depth);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
STORE(x, tmp);
}
@@ -102,13 +105,13 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
frame_ant += w;
pixel_ant = LOAD(0);
for (x = 0; x w-1; x++) {
-line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial);
-frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial,
depth);
+pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial, depth);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
STORE(x, tmp);
}
-line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial, depth);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal, depth);
STORE(x, tmp);
}
}
@@ -119,6 +122,8 @@ static void denoise_depth(uint8_t *src, uint8_t *dst,
int w, int h, int sstride, int dstride,
int16_t *spatial, int16_t *temporal, int depth)
{
+// FIXME: For 16bit depth, frame_ant could be a pointer to the previous
+// filtered frame rather than a separate buffer.
long x, y;
uint16_t *frame_ant = *frame_ant_ptr;
if (!frame_ant) {
@@ -143,24 +148,26 @@ static void denoise_depth(uint8_t *src, uint8_t *dst,
switch (hqdn3d-depth) {\
case 8: denoise_depth(__VA_ARGS__, 8); break;\
case 10: denoise_depth(__VA_ARGS__, 10); break;\
+case 16: denoise_depth(__VA_ARGS__, 16); break;\
}
-static void precalc_coefs(int16_t *ct, double dist25)
+static int16_t *precalc_coefs(double dist25, int depth)
{
int i;
double gamma, simil, C;
+int16_t *ct = av_malloc((512LUT_BITS)*sizeof(int16_t));
gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.1);
-for (i = -255*16; i = 255*16; i++) {
-// lowpass() truncates (not rounds) the diff, so +15/32 for the
midpoint of the bin.
-double f = (i + 15.0/32.0) / 16.0;
+for (i = -255LUT_BITS; i = 255LUT_BITS; i++) {
+double f = ((i(9-LUT_BITS)) + (1(8-LUT_BITS)) - 1) / 512.0; //
midpoint of the bin
simil = 1.0 - FFABS(f) / 255.0;
C = pow(simil, gamma) * 256.0 * f;
-ct[16*256+i] = lrint(C);
+ct[(256LUT_BITS)+i] = lrint(C);
}
ct[0] = !!dist25;
+return ct;
}
#define PARAM1_DEFAULT 4.0
@@ -208,6 +215,11 @@ static int init(AVFilterContext *ctx, const char *args)
}
}
+hqdn3d-strength[0] = lum_spac;
+
[libav-devel] [PATCH 7/7] vf_hqdn3d: x86 asm
13% faster on penryn, 16% on sandybridge, 15% on bulldozer
Not simd; a compiler should have generated this, but gcc didn't.
---
libavfilter/vf_hqdn3d.c| 25 +--
libavfilter/x86/Makefile |1 +
libavfilter/x86/hqdn3d.asm | 105
libavutil/x86/x86inc.asm |1 +
4 files changed, 128 insertions(+), 4 deletions(-)
create mode 100644 libavfilter/x86/hqdn3d.asm
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 521fef8..98529ce 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -40,8 +40,13 @@ typedef struct {
double strength[4];
int hsub, vsub;
int depth;
+void (*denoise_row[17])(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
} HQDN3DContext;
+void ff_hqdn3d_row_8_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+void ff_hqdn3d_row_10_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+void ff_hqdn3d_row_16_x86(uint8_t *src, uint8_t *dst, uint16_t *line_ant,
uint16_t *frame_ant, ptrdiff_t w, int16_t *spatial, int16_t *temporal);
+
#define LUT_BITS (depth==16 ? 8 : 4)
#define RIGHTSHIFT(a,b) (((a)+(((1(b))-1)1))(b))
#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) (16-depth))
@@ -78,7 +83,8 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
av_always_inline
-static void denoise_spatial(uint8_t *src, uint8_t *dst,
+static void denoise_spatial(HQDN3DContext *hqdn3d,
+uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
int16_t *spatial, int16_t *temporal, int depth)
@@ -103,6 +109,10 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
src += sstride;
dst += dstride;
frame_ant += w;
+if (hqdn3d-denoise_row[depth]) {
+hqdn3d-denoise_row[depth](src, dst, line_ant, frame_ant, w,
spatial, temporal);
+continue;
+}
pixel_ant = LOAD(0);
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial,
depth);
@@ -117,7 +127,8 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
}
av_always_inline
-static void denoise_depth(uint8_t *src, uint8_t *dst,
+static void denoise_depth(HQDN3DContext *hqdn3d,
+ uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
int16_t *spatial, int16_t *temporal, int depth)
@@ -137,7 +148,7 @@ static void denoise_depth(uint8_t *src, uint8_t *dst,
}
if (spatial[0])
-denoise_spatial(src, dst, line_ant, frame_ant,
+denoise_spatial(hqdn3d, src, dst, line_ant, frame_ant,
w, h, sstride, dstride, spatial, temporal, depth);
else
denoise_temporal(src, dst, frame_ant,
@@ -293,6 +304,12 @@ static int config_input(AVFilterLink *inlink)
}
}
+#if HAVE_YASM
+hqdn3d-denoise_row[ 8] = ff_hqdn3d_row_8_x86;
+hqdn3d-denoise_row[10] = ff_hqdn3d_row_10_x86;
+hqdn3d-denoise_row[16] = ff_hqdn3d_row_16_x86;
+#endif
+
return 0;
}
@@ -310,7 +327,7 @@ static int end_frame(AVFilterLink *inlink)
int ret, c;
for (c = 0; c 3; c++) {
-denoise(inpic-data[c], outpic-data[c],
+denoise(hqdn3d, inpic-data[c], outpic-data[c],
hqdn3d-line, hqdn3d-frame_prev[c],
inpic-video-w (!!c * hqdn3d-hsub),
inpic-video-h (!!c * hqdn3d-vsub),
diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
index e98693d..46fc84f 100644
--- a/libavfilter/x86/Makefile
+++ b/libavfilter/x86/Makefile
@@ -1,2 +1,3 @@
MMX-OBJS-$(CONFIG_YADIF_FILTER) += x86/yadif.o
MMX-OBJS-$(CONFIG_GRADFUN_FILTER)+= x86/gradfun.o
+YASM-OBJS-$(CONFIG_HQDN3D_FILTER)+= x86/hqdn3d.o
diff --git a/libavfilter/x86/hqdn3d.asm b/libavfilter/x86/hqdn3d.asm
new file mode 100644
index 000..c15c119
--- /dev/null
+++ b/libavfilter/x86/hqdn3d.asm
@@ -0,0 +1,105 @@
+;**
+;* Copyright (c) 2012 Loren Merritt
+;*
+;* This file is part of Libav.
+;*
+;* Libav is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 of the License, or (at your option) any later version.
+;*
+;* Libav is distributed in the hope that it will be useful,
+;* but WITHOUT ANY WARRANTY; without even the implied warranty
[libav-devel] [PATCH 4/7] vf_hqdn3d: reduce intermediate precision
11% faster on penryn, 7% on sandybridge, 5% on bulldozer
Negligible change to output.
---
libavfilter/vf_hqdn3d.c | 62 --
1 files changed, 32 insertions(+), 30 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index bb81b79..46902fa 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -33,32 +33,32 @@
#include video.h
typedef struct {
-int coefs[4][512*16];
-uint32_t *line;
+int16_t coefs[4][512*16];
+uint16_t *line;
uint16_t *frame_prev[3];
int hsub, vsub;
} HQDN3DContext;
-static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
+static inline uint32_t lowpass(int prev, int cur, int16_t *coef)
{
-int dmul = prev-cur;
-unsigned int d = (dmul+0x10007FF)12; // 0x1000 to convert to unsigned,
7FF for rounding
+int d = (prev-cur)4;
return cur + coef[d];
}
static void denoise_temporal(uint8_t *src, uint8_t *dst,
uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int *temporal)
+ int16_t *temporal)
{
long x, y;
uint32_t tmp;
+temporal += 0x1000;
+
for (y = 0; y h; y++) {
for (x = 0; x w; x++) {
-tmp = lowpass(frame_ant[x]8, src[x]16, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+frame_ant[x] = tmp = lowpass(frame_ant[x], src[x]8, temporal);
+dst[x] = (tmp+0x7F)8;
}
src += sstride;
dst += dstride;
@@ -67,47 +67,47 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
static void denoise_spatial(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant, uint16_t *frame_ant,
+uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int *spatial, int *temporal)
+int16_t *spatial, int16_t *temporal)
{
long x, y;
uint32_t pixel_ant;
uint32_t tmp;
+spatial += 0x1000;
+temporal += 0x1000;
+
/* First line has no top neighbor. Only left one for each tmp and
* last frame */
-pixel_ant = src[0]16;
+pixel_ant = src[0]8;
for (x = 0; x w; x++) {
-line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]16,
spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]8, spatial);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
for (y = 1; y h; y++) {
src += sstride;
dst += dstride;
frame_ant += w;
-pixel_ant = src[0]16;
+pixel_ant = src[0]8;
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-pixel_ant = lowpass(pixel_ant, src[x+1]16, spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+pixel_ant = lowpass(pixel_ant, src[x+1]8, spatial);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
}
static void denoise(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant, uint16_t **frame_ant_ptr,
+uint16_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
-int *spatial, int *temporal)
+int16_t *spatial, int16_t *temporal)
{
long x, y;
uint16_t *frame_ant = *frame_ant_ptr;
@@ -129,16 +129,18 @@ static void denoise(uint8_t *src, uint8_t *dst,
w, h, sstride, dstride, temporal);
}
-static void precalc_coefs(int *ct, double dist25)
+static void precalc_coefs(int16_t *ct, double dist25)
{
int i;
double gamma, simil, C;
-gamma = log(0.25) / log(1.0 - dist25/255.0 - 0.1);
+gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.1);
for (i = -255*16; i = 255*16; i++) {
-simil = 1.0 - FFABS(i) / (16*255.0);
-C = pow(simil, gamma) * 65536.0 * i / 16.0;
+// lowpass() truncates (not rounds) the diff, so +15/32 for the
midpoint of the bin.
+double f = (i + 15.0/32.0) / 16.0;
+simil = 1.0 - FFABS(f) / 255.0;
+C = pow(simil, gamma) * 256.0 * f;
Re: [libav-devel] [PATCH 3/7] vf_hqdn3d: simplify and optimize
On Thu, 26 Jul 2012, Ronald S. Bultje wrote: Hi, On Thu, Jul 26, 2012 at 3:51 PM, Loren Merritt lor...@u.washington.edu wrote: 14% faster on penryn, 2% on sandybridge, 9% on bulldozer --- libavfilter/vf_hqdn3d.c | 157 +++--- 1 files changed, 51 insertions(+), 106 deletions(-) Looks good. I am going to ask a very stupid question: why is this faster? I see a lot of simplification, which is good, but I'm not quite sure which part actually has a clear speed impact. Old code's sline_offs and dline_offs confused gcc into incrementing the src and dst pointers rather than using x as an index reg. Old code did horizontal(x), vertical(x), temporal(x). There's a dependency chain between those 3 filters, so you need to interleave multiple loop iterations to get maximum throughput. OOE might theoretically handle that, but doesn't do so perfectly on the CPUs I tested. New code does vertical(x), horizontal(x+1), temporal(x); which requires less OOE. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 5/7] vf_hqdn3d: support 9 and 10bit colordepth
---
libavfilter/vf_hqdn3d.c | 72 ++
1 files changed, 53 insertions(+), 19 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 46902fa..138b527 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -27,6 +27,7 @@
*/
#include libavutil/pixdesc.h
+#include libavutil/intreadwrite.h
#include avfilter.h
#include formats.h
#include internal.h
@@ -37,18 +38,25 @@ typedef struct {
uint16_t *line;
uint16_t *frame_prev[3];
int hsub, vsub;
+int depth;
} HQDN3DContext;
+#define RIGHTSHIFT(a,b) (((a)+(((1(b))-1)1))(b))
+#define LOAD(x) ((depth==8 ? src[x] : AV_RN16A(src+(x)*2)) (16-depth))
+#define STORE(x,val) (depth==8 ? dst[x] = RIGHTSHIFT(val, 16-depth)\
+: AV_WN16A(dst+(x)*2, RIGHTSHIFT(val, 16-depth)))
+
static inline uint32_t lowpass(int prev, int cur, int16_t *coef)
{
int d = (prev-cur)4;
return cur + coef[d];
}
+av_always_inline
static void denoise_temporal(uint8_t *src, uint8_t *dst,
uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int16_t *temporal)
+ int16_t *temporal, int depth)
{
long x, y;
uint32_t tmp;
@@ -57,8 +65,8 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
for (y = 0; y h; y++) {
for (x = 0; x w; x++) {
-frame_ant[x] = tmp = lowpass(frame_ant[x], src[x]8, temporal);
-dst[x] = (tmp+0x7F)8;
+frame_ant[x] = tmp = lowpass(frame_ant[x], LOAD(x), temporal);
+STORE(x, tmp);
}
src += sstride;
dst += dstride;
@@ -66,10 +74,11 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
}
+av_always_inline
static void denoise_spatial(uint8_t *src, uint8_t *dst,
uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int16_t *spatial, int16_t *temporal)
+int16_t *spatial, int16_t *temporal, int depth)
{
long x, y;
uint32_t pixel_ant;
@@ -80,34 +89,35 @@ static void denoise_spatial(uint8_t *src, uint8_t *dst,
/* First line has no top neighbor. Only left one for each tmp and
* last frame */
-pixel_ant = src[0]8;
+pixel_ant = LOAD(0);
for (x = 0; x w; x++) {
-line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]8, spatial);
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, LOAD(x), spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
for (y = 1; y h; y++) {
src += sstride;
dst += dstride;
frame_ant += w;
-pixel_ant = src[0]8;
+pixel_ant = LOAD(0);
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-pixel_ant = lowpass(pixel_ant, src[x+1]8, spatial);
+pixel_ant = lowpass(pixel_ant, LOAD(x+1), spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
-dst[x] = (tmp+0x7F)8;
+STORE(x, tmp);
}
}
-static void denoise(uint8_t *src, uint8_t *dst,
-uint16_t *line_ant, uint16_t **frame_ant_ptr,
-int w, int h, int sstride, int dstride,
-int16_t *spatial, int16_t *temporal)
+av_always_inline
+static void denoise_depth(uint8_t *src, uint8_t *dst,
+ uint16_t *line_ant, uint16_t **frame_ant_ptr,
+ int w, int h, int sstride, int dstride,
+ int16_t *spatial, int16_t *temporal, int depth)
{
long x, y;
uint16_t *frame_ant = *frame_ant_ptr;
@@ -116,19 +126,26 @@ static void denoise(uint8_t *src, uint8_t *dst,
*frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
for (y = 0; y h; y++, src += sstride, frame_ant += w)
for (x = 0; x w; x++)
-frame_ant[x] = src[x]8;
+frame_ant[x] = LOAD(x);
src = frame_src;
frame_ant = *frame_ant_ptr;
}
if (spatial[0])
denoise_spatial(src, dst, line_ant, frame_ant,
-w, h, sstride, dstride, spatial, temporal);
+w, h, sstride, dstride, spatial, temporal, depth);
else
denoise_temporal(src, dst, frame_ant,
- w, h, sstride, dstride, temporal);
+ w, h, sstride, dstride, temporal, depth);
}
+#define denoise(...) \
+switch (hqdn3d-depth) {\
+case 8:
Re: [libav-devel] [PATCH 2/2] lavr: add x86-optimized mixing functions
On Tue, 24 Jul 2012, Justin Ruggles wrote: Adds optimized functions for mixing 3 through 8 input channels to 1 and 2 output channels in fltp or s16p format with flt coeffs. --- libavresample/utils.c |7 +- libavresample/x86/audio_mix.asm| 293 libavresample/x86/audio_mix_init.c | 130 3 files changed, 427 insertions(+), 3 deletions(-) LGTM. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [libav-commits] x86: fft: convert sse inline asm to yasm
On Wed, 25 Jul 2012, Luca Barbato wrote:
On 07/25/2012 07:40 AM, Jason Garrett-Glaser wrote:
Do the x264 functions sign-extend all their integer arguments? Or put
differently, does the problem occur for 32-bit builds also, or only
for 64-bit builds?
Yes they do, and such a problem wouldn't target solely _avx functions;
that wouldn't make any sense.
How about start by checking for a missing vzeroupper?
Is there a simple way to do that? We could fix that part sooner than later.
Done.
... And it does find a hit, ff_mix_1_to_2_fltp_flt_avx.
--Loren Merritt#!/usr/bin/perl -w
$exe = $ARGV[0];
@ARGV==1 and -f $exe and -x $exe or die
usage: missing_vzeroupper.pl avconv\n.
Finds functions that use ymm and fail to reset the ymm state with vzeroupper,\n.
which thus incur a large speed penalty when mixed with non-avx xmm.\n;
open FH, -|, objdump, -d, -M, intel, $exe or die failed to run objdump: $!\n;
@funcs = split /^[0-9a-f]+ ([^. ]+):/m, join , FH;
close FH or die;
# The fft functions are ok because the vzeroupper happens in fft_dispatch.
$exceptions = qr/^_?(fft\d+|pass)_/;
$err = 0;
shift @funcs;
while(@funcs) {
my $funcname = shift @funcs;
my $asm = shift @funcs;
if($asm =~ /\bymm\d/ and $asm !~ /\bvzeroupper\b/ and $funcname !~ $exceptions) {
print $funcname\n;
$err = 1;
}
}
exit $err;
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH 14/15] lavr: x86: optimized 2-channel flt to fltp conversion
On Wed, 25 Jul 2012, Justin Ruggles wrote: On 07/25/2012 01:03 AM, Ronald S. Bultje wrote: (Can this be implemented in YMM with the current instructions available?) Not really sure if it would help. Maybe something like this? vpermilps m0, [srcq+2*lenq], q0213 vextractf128 m1, m0, 1 SBUTTERFLY2 qdq, m0, m1, m2 movdqa [dst0q+lenq], m0 movdqa [dst1q+lenq], m1 Is that even worth testing? No. This function should be memory-bound. And 32byte memory ops aren't any higher thoughput than 16byte memory ops unless they come in the ratio of 2 loads per 1 store. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
Re: [libav-devel] [PATCH] x86inc: automatically insert vzeroupper for YMM functions.
On Wed, 25 Jul 2012, Ronald S. Bultje wrote: diff --git a/libavutil/x86/x86inc.asm b/libavutil/x86/x86inc.asm index 42ba97a..4b523e9 100644 --- a/libavutil/x86/x86inc.asm +++ b/libavutil/x86/x86inc.asm @@ -369,11 +369,14 @@ DECLARE_REG 14, R15, R15D, R15W, R15B, 120 %macro RET 0 WIN64_RESTORE_XMM_INTERNAL rsp POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7 +%if mmsize == 32 +vzeroupper +%endif ret %endmacro %macro REP_RET 0 -%if regs_used 7 || xmm_regs_used 6 +%if regs_used 7 || xmm_regs_used 6 || mmsize == 32 RET %else rep ret There's 3 implementations of RET, for the 3 ABIs. --Loren Merritt ___ libav-devel mailing list libav-devel@libav.org https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 1/7] vf_hqdn3d: cosmetics
Change code style to match the rest of libav.
---
libavfilter/vf_hqdn3d.c | 309 +++
1 files changed, 153 insertions(+), 156 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index af69d41..8e2a6a2 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -32,165 +32,162 @@
#include video.h
typedef struct {
-int Coefs[4][512*16];
-unsigned int *Line;
-unsigned short *Frame[3];
+int coefs[4][512*16];
+uint32_t *line;
+uint16_t *frame_prev[3];
int hsub, vsub;
} HQDN3DContext;
-static inline unsigned int LowPassMul(unsigned int PrevMul, unsigned int
CurrMul, int *Coef)
+static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
{
-//int dMul= (PrevMul0xFF)-(CurrMul0xFF);
-int dMul= PrevMul-CurrMul;
-unsigned int d=((dMul+0x10007FF)12);
-return CurrMul + Coef[d];
+//int dmul= (prev0xFF)-(cur0xFF);
+int dmul = prev-cur;
+uint32_t d = ((dmul+0x10007FF)12);
+return cur + coef[d];
}
-static void deNoiseTemporal(unsigned char *FrameSrc,
-unsigned char *FrameDest,
-unsigned short *FrameAnt,
-int W, int H, int sStride, int dStride,
-int *Temporal)
+static void denoise_temporal(uint8_t *src, uint8_t *dst,
+ uint16_t *frame_ant,
+ int w, int h, int sstride, int dstride,
+ int *temporal)
{
-long X, Y;
-unsigned int PixelDst;
-
-for (Y = 0; Y H; Y++) {
-for (X = 0; X W; X++) {
-PixelDst = LowPassMul(FrameAnt[X]8, FrameSrc[X]16, Temporal);
-FrameAnt[X] = ((PixelDst+0x107F)8);
-FrameDest[X]= ((PixelDst+0x10007FFF)16);
+long x, y;
+uint32_t pixel;
+
+for (y = 0; y h; y++) {
+for (x = 0; x w; x++) {
+pixel = lowpass(frame_ant[x]8, src[x]16, temporal);
+frame_ant[x] = ((pixel+0x107F)8);
+dst[x]= ((pixel+0x10007FFF)16);
}
-FrameSrc += sStride;
-FrameDest += dStride;
-FrameAnt += W;
+src += sstride;
+dst += dstride;
+frame_ant += w;
}
}
-static void deNoiseSpacial(unsigned char *Frame,
- unsigned char *FrameDest,
- unsigned int *LineAnt,
- int W, int H, int sStride, int dStride,
- int *Horizontal, int *Vertical)
+static void denoise_spatial(uint8_t *src, uint8_t *dst,
+uint32_t *line_ant,
+int w, int h, int sstride, int dstride,
+int *horizontal, int *vertical)
{
-long X, Y;
-long sLineOffs = 0, dLineOffs = 0;
-unsigned int PixelAnt;
-unsigned int PixelDst;
+long x, y;
+long sline_offs = 0, dline_offs = 0;
+uint32_t pixel_ant;
+uint32_t pixel;
/* First pixel has no left nor top neighbor. */
-PixelDst = LineAnt[0] = PixelAnt = Frame[0]16;
-FrameDest[0]= ((PixelDst+0x10007FFF)16);
+pixel = line_ant[0] = pixel_ant = src[0]16;
+dst[0]= ((pixel+0x10007FFF)16);
/* First line has no top neighbor, only left. */
-for (X = 1; X W; X++) {
-PixelDst = LineAnt[X] = LowPassMul(PixelAnt, Frame[X]16, Horizontal);
-FrameDest[X]= ((PixelDst+0x10007FFF)16);
+for (x = 1; x w; x++) {
+pixel = line_ant[x] = lowpass(pixel_ant, src[x]16, horizontal);
+dst[x]= ((pixel+0x10007FFF)16);
}
-for (Y = 1; Y H; Y++) {
-unsigned int PixelAnt;
-sLineOffs += sStride, dLineOffs += dStride;
+for (y = 1; y h; y++) {
+uint32_t pixel_ant;
+sline_offs += sstride, dline_offs += dstride;
/* First pixel on each line doesn't have previous pixel */
-PixelAnt = Frame[sLineOffs]16;
-PixelDst = LineAnt[0] = LowPassMul(LineAnt[0], PixelAnt, Vertical);
-FrameDest[dLineOffs]= ((PixelDst+0x10007FFF)16);
+pixel_ant = src[sline_offs]16;
+pixel = line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
+dst[dline_offs]= ((pixel+0x10007FFF)16);
-for (X = 1; X W; X++) {
-unsigned int PixelDst;
+for (x = 1; x w; x++) {
+uint32_t pixel;
/* The rest are normal */
-PixelAnt = LowPassMul(PixelAnt, Frame[sLineOffs+X]16,
Horizontal);
-PixelDst = LineAnt[X] = LowPassMul(LineAnt[X], PixelAnt, Vertical);
-FrameDest[dLineOffs+X]= ((PixelDst+0x10007FFF)16);
+pixel_ant = lowpass(pixel_ant, src[sline_offs+x]16, horizontal);
+pixel = line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
+dst[dline_offs+x]= ((pixel+0x10007FFF)16);
}
}
}
-static void deNoise(unsigned char *Frame,
-
[libav-devel] [PATCH 2/7] vf_hqdn3d: simplify and optimize
14% faster on penryn, 2% on sandybridge, 9% on bulldozer
---
libavfilter/vf_hqdn3d.c | 166 +++---
1 files changed, 54 insertions(+), 112 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index 8e2a6a2..fdb2ecb 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -1,6 +1,7 @@
/*
* Copyright (c) 2003 Daniel Moreno comac AT comac DOT darktech DOT org
* Copyright (c) 2010 Baptiste Coudurier
+ * Copyright (c) 2012 Loren Merritt
*
* This file is part of Libav, ported from MPlayer.
*
@@ -40,9 +41,8 @@ typedef struct {
static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
{
-//int dmul= (prev0xFF)-(cur0xFF);
int dmul = prev-cur;
-uint32_t d = ((dmul+0x10007FF)12);
+unsigned int d = (dmul+0x10007FF)12; // 0x1000 to convert to unsigned,
7FF for rounding
return cur + coef[d];
}
@@ -51,14 +51,13 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
int w, int h, int sstride, int dstride,
int *temporal)
{
-long x, y;
-uint32_t pixel;
+uint32_t tmp;
-for (y = 0; y h; y++) {
-for (x = 0; x w; x++) {
-pixel = lowpass(frame_ant[x]8, src[x]16, temporal);
-frame_ant[x] = ((pixel+0x107F)8);
-dst[x]= ((pixel+0x10007FFF)16);
+for (long y = 0; y h; y++) {
+for (long x = 0; x w; x++) {
+tmp = lowpass(frame_ant[x]8, src[x]16, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
src += sstride;
dst += dstride;
@@ -66,112 +65,67 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
}
}
+av_always_inline
static void denoise_spatial(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant,
+uint32_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int *horizontal, int *vertical)
+int *spatial, int *temporal)
{
long x, y;
-long sline_offs = 0, dline_offs = 0;
uint32_t pixel_ant;
-uint32_t pixel;
-
-/* First pixel has no left nor top neighbor. */
-pixel = line_ant[0] = pixel_ant = src[0]16;
-dst[0]= ((pixel+0x10007FFF)16);
+uint32_t tmp;
-/* First line has no top neighbor, only left. */
-for (x = 1; x w; x++) {
-pixel = line_ant[x] = lowpass(pixel_ant, src[x]16, horizontal);
-dst[x]= ((pixel+0x10007FFF)16);
+/* First line has no top neighbor. Only left one for each tmp and
+ * last frame */
+pixel_ant = src[0]16;
+for (x = 0; x w; x++) {
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]16,
spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
for (y = 1; y h; y++) {
-uint32_t pixel_ant;
-sline_offs += sstride, dline_offs += dstride;
-/* First pixel on each line doesn't have previous pixel */
-pixel_ant = src[sline_offs]16;
-pixel = line_ant[0] = lowpass(line_ant[0], pixel_ant, vertical);
-dst[dline_offs]= ((pixel+0x10007FFF)16);
-
-for (x = 1; x w; x++) {
-uint32_t pixel;
-/* The rest are normal */
-pixel_ant = lowpass(pixel_ant, src[sline_offs+x]16, horizontal);
-pixel = line_ant[x] = lowpass(line_ant[x], pixel_ant, vertical);
-dst[dline_offs+x]= ((pixel+0x10007FFF)16);
+src += sstride;
+dst += dstride;
+frame_ant += w;
+pixel_ant = src[0]16;
+for (x = 0; x w-1; x++) {
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
+pixel_ant = lowpass(pixel_ant, src[x+1]16, spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
+line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
+tmp = lowpass(frame_ant[x]8, tmp, temporal);
+frame_ant[x] = (tmp+0x7F)8;
+dst[x] = (tmp+0x7FFF)16;
}
}
static void denoise(uint8_t *src, uint8_t *dst,
uint32_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
-int *horizontal, int *vertical, int *temporal)
+int *spatial, int *temporal)
{
-long x, y;
-long sline_offs = 0, dline_offs = 0;
-uint32_t pixel_ant;
-uint32_t pixel;
uint16_t *frame_ant = *frame_ant_ptr;
-
if (!frame_ant) {
+uint8_t *frame_src = src;
*frame_ant_ptr = frame_ant = av_malloc(w*h*sizeof(uint16_t));
-for (y = 0; y h; y++) {
-uint16_t *frame_dst = frame_ant+y*w;
-uint8_t *frame_src
[libav-devel] [PATCH 3/7] vf_hqdn3d: reduce intermediate precision
11% faster on penryn, 7% on sandybridge, 5% on bulldozer
Negligible change to output.
---
libavfilter/vf_hqdn3d.c | 62 --
1 files changed, 32 insertions(+), 30 deletions(-)
diff --git a/libavfilter/vf_hqdn3d.c b/libavfilter/vf_hqdn3d.c
index fdb2ecb..505a8ab 100644
--- a/libavfilter/vf_hqdn3d.c
+++ b/libavfilter/vf_hqdn3d.c
@@ -33,31 +33,31 @@
#include video.h
typedef struct {
-int coefs[4][512*16];
-uint32_t *line;
+int16_t coefs[4][512*16];
+uint16_t *line;
uint16_t *frame_prev[3];
int hsub, vsub;
} HQDN3DContext;
-static inline uint32_t lowpass(unsigned int prev, unsigned int cur, int *coef)
+static inline uint32_t lowpass(int prev, int cur, int16_t *coef)
{
-int dmul = prev-cur;
-unsigned int d = (dmul+0x10007FF)12; // 0x1000 to convert to unsigned,
7FF for rounding
+int d = (prev-cur)4;
return cur + coef[d];
}
static void denoise_temporal(uint8_t *src, uint8_t *dst,
uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
- int *temporal)
+ int16_t *temporal)
{
uint32_t tmp;
+temporal += 0x1000;
+
for (long y = 0; y h; y++) {
for (long x = 0; x w; x++) {
-tmp = lowpass(frame_ant[x]8, src[x]16, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+frame_ant[x] = tmp = lowpass(frame_ant[x], src[x]8, temporal);
+dst[x] = (tmp+0x7F)8;
}
src += sstride;
dst += dstride;
@@ -67,47 +67,47 @@ static void denoise_temporal(uint8_t *src, uint8_t *dst,
av_always_inline
static void denoise_spatial(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant, uint16_t *frame_ant,
+uint16_t *line_ant, uint16_t *frame_ant,
int w, int h, int sstride, int dstride,
-int *spatial, int *temporal)
+int16_t *spatial, int16_t *temporal)
{
long x, y;
uint32_t pixel_ant;
uint32_t tmp;
+spatial += 0x1000;
+temporal += 0x1000;
+
/* First line has no top neighbor. Only left one for each tmp and
* last frame */
-pixel_ant = src[0]16;
+pixel_ant = src[0]8;
for (x = 0; x w; x++) {
-line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]16,
spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+line_ant[x] = tmp = pixel_ant = lowpass(pixel_ant, src[x]8, spatial);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
for (y = 1; y h; y++) {
src += sstride;
dst += dstride;
frame_ant += w;
-pixel_ant = src[0]16;
+pixel_ant = src[0]8;
for (x = 0; x w-1; x++) {
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-pixel_ant = lowpass(pixel_ant, src[x+1]16, spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+pixel_ant = lowpass(pixel_ant, src[x+1]8, spatial);
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
line_ant[x] = tmp = lowpass(line_ant[x], pixel_ant, spatial);
-tmp = lowpass(frame_ant[x]8, tmp, temporal);
-frame_ant[x] = (tmp+0x7F)8;
-dst[x] = (tmp+0x7FFF)16;
+frame_ant[x] = tmp = lowpass(frame_ant[x], tmp, temporal);
+dst[x] = (tmp+0x7F)8;
}
}
static void denoise(uint8_t *src, uint8_t *dst,
-uint32_t *line_ant, uint16_t **frame_ant_ptr,
+uint16_t *line_ant, uint16_t **frame_ant_ptr,
int w, int h, int sstride, int dstride,
-int *spatial, int *temporal)
+int16_t *spatial, int16_t *temporal)
{
uint16_t *frame_ant = *frame_ant_ptr;
if (!frame_ant) {
@@ -128,16 +128,18 @@ static void denoise(uint8_t *src, uint8_t *dst,
w, h, sstride, dstride, temporal);
}
-static void precalc_coefs(int *ct, double dist25)
+static void precalc_coefs(int16_t *ct, double dist25)
{
int i;
double gamma, simil, C;
-gamma = log(0.25) / log(1.0 - dist25/255.0 - 0.1);
+gamma = log(0.25) / log(1.0 - FFMIN(dist25,252.0)/255.0 - 0.1);
for (i = -255*16; i = 255*16; i++) {
-simil = 1.0 - FFABS(i) / (16*255.0);
-C = pow(simil, gamma) * 65536.0 * i / 16.0;
+// lowpass() truncates (not rounds) the diff, so +15/32 for the
midpoint of the bin.
+double f = (i + 15.0/32.0) / 16.0;
+simil = 1.0 - FFABS(f) / 255.0;
+C = pow(simil, gamma) * 256.0 * f;
