[libav-devel] [PATCH 1/4] h264/aarch64: sign extend int stride in loop filter asm
---
libavcodec/aarch64/h264dsp_neon.S | 3 +++
1 file changed, 3 insertions(+)
diff --git a/libavcodec/aarch64/h264dsp_neon.S
b/libavcodec/aarch64/h264dsp_neon.S
index 9b4610a4d4..60ffa24500 100644
--- a/libavcodec/aarch64/h264dsp_neon.S
+++ b/libavcodec/aarch64/h264dsp_neon.S
@@ -130,6 +130,7 @@ endfunc
function ff_h264_h_loop_filter_luma_neon, export=1
h264_loop_filter_start
+sxtwx1, w1
sub x0, x0, #4
ld1 {v6.8B}, [x0], x1
@@ -210,6 +211,7 @@ endfunc
function ff_h264_v_loop_filter_chroma_neon, export=1
h264_loop_filter_start
+sxtwx1, w1
sub x0, x0, x1, lsl #1
ld1 {v18.8B}, [x0], x1
@@ -228,6 +230,7 @@ endfunc
function ff_h264_h_loop_filter_chroma_neon, export=1
h264_loop_filter_start
+sxtwx1, w1
sub x0, x0, #2
ld1 {v18.S}[0], [x0], x1
--
2.20.1
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
[libav-devel] [PATCH 2/4] checkasm/h264: add loop filter tests
---
tests/checkasm/h264dsp.c | 124 +++
1 file changed, 124 insertions(+)
diff --git a/tests/checkasm/h264dsp.c b/tests/checkasm/h264dsp.c
index f355a72a74..706fc79397 100644
--- a/tests/checkasm/h264dsp.c
+++ b/tests/checkasm/h264dsp.c
@@ -28,6 +28,7 @@
#include "libavutil/intreadwrite.h"
static const uint32_t pixel_mask[3] = { 0x, 0x01ff01ff, 0x03ff03ff };
+static const uint32_t pixel_mask_lf[3] = { 0xff0fff0f, 0x01ff000f, 0x03ff000f
};
#define SIZEOF_PIXEL ((bit_depth + 7) / 8)
#define SIZEOF_COEF (2 * ((bit_depth + 7) / 8))
@@ -312,9 +313,132 @@ static void check_idct_multiple(void)
}
}
+
+static void check_loop_filter(void)
+{
+LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
+LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
+LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
+H264DSPContext h;
+int bit_depth;
+int alphas[36], betas[36];
+int8_t tc0[36][4];
+
+declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, int stride,
+ int alpha, int beta, int8_t *tc0);
+
+for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
+int i, j, a, c;
+uint32_t mask = pixel_mask_lf[bit_depth - 8];
+ff_h264dsp_init(, bit_depth, 1);
+for (i = 35, a = 255, c = 250; i >= 0; i--) {
+alphas[i] = a << (bit_depth - 8);
+betas[i] = (i + 1) / 2 << (bit_depth - 8);
+tc0[i][0] = tc0[i][3] = (c + 6) / 10;
+tc0[i][1] = (c + 7) / 15;
+tc0[i][2] = (c + 9) / 20;
+a = a*9/10;
+c = c*9/10;
+}
+
+#define CHECK_LOOP_FILTER(name, align, ...) \
+do {\
+if (check_func(h.name, #name "_%dbpp", bit_depth)) {\
+for (j = 0; j < 36; j++) { \
+intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \
+for (i = 0; i < 1024; i+=4) { \
+AV_WN32A(dst + i, rnd() & mask);\
+} \
+memcpy(dst0, dst, 32 * 16 * 2); \
+memcpy(dst1, dst, 32 * 16 * 2); \
+\
+call_ref(dst0 + off, 32, alphas[j], betas[j], tc0[j]); \
+call_new(dst1 + off, 32, alphas[j], betas[j], tc0[j]); \
+if (memcmp(dst0, dst1, 32 * 16 * SIZEOF_PIXEL)) { \
+fprintf(stderr, #name ": j:%d, alpha:%d beta:%d " \
+"tc0:{%d,%d,%d,%d}\n", j, alphas[j], betas[j],
\
+tc0[j][0], tc0[j][1], tc0[j][2], tc0[j][3]); \
+fail(); \
+} \
+bench_new(dst1, 32, alphas[j], betas[j], tc0[j]); \
+} \
+} \
+} while (0)
+
+CHECK_LOOP_FILTER(h264_v_loop_filter_luma, 1);
+CHECK_LOOP_FILTER(h264_h_loop_filter_luma, 0);
+CHECK_LOOP_FILTER(h264_h_loop_filter_luma_mbaff, 0);
+CHECK_LOOP_FILTER(h264_v_loop_filter_chroma, 1);
+CHECK_LOOP_FILTER(h264_h_loop_filter_chroma, 0);
+CHECK_LOOP_FILTER(h264_h_loop_filter_chroma_mbaff, 0);
+#undef CHECK_LOOP_FILTER
+}
+}
+
+static void check_loop_filter_intra(void)
+{
+LOCAL_ALIGNED_16(uint8_t, dst, [32 * 16 * 2]);
+LOCAL_ALIGNED_16(uint8_t, dst0, [32 * 16 * 2]);
+LOCAL_ALIGNED_16(uint8_t, dst1, [32 * 16 * 2]);
+H264DSPContext h;
+int bit_depth;
+int alphas[36], betas[36];
+
+declare_func_emms(AV_CPU_FLAG_MMX, void, uint8_t *pix, int stride,
+ int alpha, int beta);
+
+for (bit_depth = 8; bit_depth <= 10; bit_depth++) {
+int i, j, a;
+uint32_t mask = pixel_mask_lf[bit_depth - 8];
+ff_h264dsp_init(, bit_depth, 1);
+for (i = 35, a = 255; i >= 0; i--) {
+alphas[i] = a << (bit_depth - 8);
+betas[i] = (i + 1) / 2 << (bit_depth - 8);
+a = a*9/10;
+}
+
+#define CHECK_LOOP_FILTER(name, align) \
+do {\
+if (check_func(h.name, #name "_%dbpp", bit_depth)) {\
+for (j = 0; j < 36; j++) { \
+intptr_t off = 8 * 32 + (j & 15) * 4 * !align; \
+for (i = 0; i < 1024; i+=4) { \
+AV_WN32A(dst + i, rnd() & mask);
[libav-devel] [PATCH 4/4] h264/aarch64: add intra loop filter neon asm
Add my neon asm from x264 relicensed under the LGPL 2.1 or later. Ported
(x264 uses nv12 chroma) and optimized.
Cycle count for checkasm --bench on a Snapdragon 820e:
h264_h_loop_filter_luma_intra_8bpp_c: 60.0
h264_h_loop_filter_luma_intra_8bpp_neon: 54.2
h264_v_loop_filter_luma_intra_8bpp_c: 148.3
h264_v_loop_filter_luma_intra_8bpp_neon: 73.8
h264_h_loop_filter_chroma_intra_8bpp_c: 27.8
h264_h_loop_filter_chroma_intra_8bpp_neon: 21.4
h264_h_loop_filter_chroma_mbaff_intra_8bpp_c: 15.8
h264_h_loop_filter_chroma_mbaff_intra_8bpp_neon: 15.7
h264_v_loop_filter_chroma_intra_8bpp_c: 45.8
h264_v_loop_filter_chroma_intra_8bpp_neon: 17.3
---
libavcodec/aarch64/h264dsp_init_aarch64.c | 16 ++
libavcodec/aarch64/h264dsp_neon.S | 297 ++
2 files changed, 313 insertions(+)
diff --git a/libavcodec/aarch64/h264dsp_init_aarch64.c
b/libavcodec/aarch64/h264dsp_init_aarch64.c
index b106f11134..07bda2ff07 100644
--- a/libavcodec/aarch64/h264dsp_init_aarch64.c
+++ b/libavcodec/aarch64/h264dsp_init_aarch64.c
@@ -29,10 +29,20 @@ void ff_h264_v_loop_filter_luma_neon(uint8_t *pix, int
stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_h_loop_filter_luma_neon(uint8_t *pix, int stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_luma_intra_neon(uint8_t *pix, int stride, int alpha,
+ int beta);
+void ff_h264_h_loop_filter_luma_intra_neon(uint8_t *pix, int stride, int alpha,
+ int beta);
void ff_h264_v_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
int beta, int8_t *tc0);
void ff_h264_h_loop_filter_chroma_neon(uint8_t *pix, int stride, int alpha,
int beta, int8_t *tc0);
+void ff_h264_v_loop_filter_chroma_intra_neon(uint8_t *pix, int stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_intra_neon(uint8_t *pix, int stride,
+ int alpha, int beta);
+void ff_h264_h_loop_filter_chroma_mbaff_intra_neon(uint8_t *pix, int stride,
+ int alpha, int beta);
void ff_weight_h264_pixels_16_neon(uint8_t *dst, int stride, int height,
int log2_den, int weight, int offset);
@@ -77,8 +87,14 @@ av_cold void ff_h264dsp_init_aarch64(H264DSPContext *c,
const int bit_depth,
if (have_neon(cpu_flags) && bit_depth == 8) {
c->h264_v_loop_filter_luma = ff_h264_v_loop_filter_luma_neon;
c->h264_h_loop_filter_luma = ff_h264_h_loop_filter_luma_neon;
+c->h264_v_loop_filter_luma_intra=
ff_h264_v_loop_filter_luma_intra_neon;
+c->h264_h_loop_filter_luma_intra=
ff_h264_h_loop_filter_luma_intra_neon;
+
c->h264_v_loop_filter_chroma = ff_h264_v_loop_filter_chroma_neon;
c->h264_h_loop_filter_chroma = ff_h264_h_loop_filter_chroma_neon;
+c->h264_v_loop_filter_chroma_intra =
ff_h264_v_loop_filter_chroma_intra_neon;
+c->h264_h_loop_filter_chroma_intra =
ff_h264_h_loop_filter_chroma_intra_neon;
+c->h264_h_loop_filter_chroma_mbaff_intra =
ff_h264_h_loop_filter_chroma_mbaff_intra_neon;
c->weight_h264_pixels_tab[0] = ff_weight_h264_pixels_16_neon;
c->weight_h264_pixels_tab[1] = ff_weight_h264_pixels_8_neon;
diff --git a/libavcodec/aarch64/h264dsp_neon.S
b/libavcodec/aarch64/h264dsp_neon.S
index b649f1d018..448e575b8c 100644
--- a/libavcodec/aarch64/h264dsp_neon.S
+++ b/libavcodec/aarch64/h264dsp_neon.S
@@ -1,6 +1,7 @@
/*
* Copyright (c) 2008 Mans Rullgard
* Copyright (c) 2013 Janne Grunau
+ * Copyright (c) 2014 Janne Grunau
*
* This file is part of Libav.
*
@@ -181,6 +182,203 @@ function ff_h264_h_loop_filter_luma_neon, export=1
ret
endfunc
+
+.macro h264_loop_filter_start_intra
+orr w4, w2, w3
+cbnzw4, 1f
+ret
+1:
+sxtwx1, w1
+dup v30.16b, w2// alpha
+dup v31.16b, w3// beta
+.endm
+
+.macro h264_loop_filter_luma_intra
+uabdv16.16b, v7.16b, v0.16b// abs(p0 - q0)
+uabdv17.16b, v6.16b, v7.16b// abs(p1 - p0)
+uabdv18.16b, v1.16b, v0.16b// abs(q1 - q0)
+cmhiv19.16b, v30.16b, v16.16b // < alpha
+cmhiv17.16b, v31.16b, v17.16b // < beta
+cmhiv18.16b, v31.16b, v18.16b // < beta
+
+moviv29.16b, #2
+ushrv30.16b, v30.16b, #2// alpha >> 2
+add v30.16b, v30.16b, v29.16b // (alpha >> 2) + 2
+cmhiv16.16b, v30.16b, v16.16b // < (alpha >> 2) + 2
+
+and v19.16b, v19.16b, v17.16b
+and v19.16b,
[libav-devel] [PATCH 3/4] h264/aarch64: optimize neon loop filter
Exit as soon as possible if no filtering will be done.
Improves the checkasm --bench cycle count on a Snapdragon 820e:
h264_h_loop_filter_luma_8bpp_c: 72.4 -> 72.5
h264_h_loop_filter_luma_8bpp_neon: 97.1 -> 56.3
h264_v_loop_filter_luma_8bpp_c: 174.0 -> 173.5
h264_v_loop_filter_luma_8bpp_neon: 62.9 -> 60.9
h264_h_loop_filter_chroma_8bpp_c:30.2 -> 30.3
h264_h_loop_filter_chroma_8bpp_neon: 51.6 -> 25.7
h264_v_loop_filter_chroma_8bpp_c:57.3 -> 57.3
h264_v_loop_filter_chroma_8bpp_neon: 28.0 -> 24.0
---
libavcodec/aarch64/h264dsp_neon.S | 33 ++-
1 file changed, 19 insertions(+), 14 deletions(-)
diff --git a/libavcodec/aarch64/h264dsp_neon.S
b/libavcodec/aarch64/h264dsp_neon.S
index 60ffa24500..b649f1d018 100644
--- a/libavcodec/aarch64/h264dsp_neon.S
+++ b/libavcodec/aarch64/h264dsp_neon.S
@@ -54,9 +54,12 @@
uabdv17.16B, v20.16B, v16.16B // abs(p2 - p0)
and v21.16B, v21.16B, v28.16B
uabdv19.16B, v4.16B, v0.16B // abs(q2 - q0)
+and v21.16B, v21.16B, v30.16B // < beta
+shrnv30.8b, v21.8h, #4
+mov x7, v30.d[0]
cmhiv17.16B, v22.16B, v17.16B // < beta
-and v21.16B, v21.16B, v30.16B
cmhiv19.16B, v22.16B, v19.16B // < beta
+cbz x7, 9f
and v17.16B, v17.16B, v21.16B
and v19.16B, v19.16B, v21.16B
and v24.16B, v24.16B, v21.16B
@@ -124,7 +127,7 @@ function ff_h264_v_loop_filter_luma_neon, export=1
st1 {v16.16B}, [x0], x1
st1 {v0.16B}, [x0], x1
st1 {v19.16B}, [x0]
-
+9:
ret
endfunc
@@ -174,32 +177,34 @@ function ff_h264_h_loop_filter_luma_neon, export=1
st1 {v16.S}[3], [x0], x1
st1 {v0.S}[3], [x0], x1
st1 {v19.S}[3], [x0], x1
-
+9:
ret
endfunc
.macro h264_loop_filter_chroma
dup v22.8B, w2 // alpha
+dup v23.8B, w3 // beta
uxtlv24.8H, v24.8B
uabdv26.8B, v16.8B, v0.8B // abs(p0 - q0)
-uxtlv4.8H, v0.8B
uabdv28.8B, v18.8B, v16.8B // abs(p1 - p0)
+uabdv30.8B, v2.8B, v0.8B // abs(q1 - q0)
+cmhiv26.8B, v22.8B, v26.8B // < alpha
+cmhiv28.8B, v23.8B, v28.8B // < beta
+cmhiv30.8B, v23.8B, v30.8B // < beta
+uxtlv4.8H, v0.8B
+and v26.8B, v26.8B, v28.8B
usubw v4.8H, v4.8H, v16.8B
-sli v24.8H, v24.8H, #8
+and v26.8B, v26.8B, v30.8B
shl v4.8H, v4.8H, #2
-uabdv30.8B, v2.8B, v0.8B // abs(q1 - q0)
+mov x2, v26.d[0]
+sli v24.8H, v24.8H, #8
uaddw v4.8H, v4.8H, v18.8B
-cmhiv26.8B, v22.8B, v26.8B // < alpha
+cbz x2, 9f
usubw v4.8H, v4.8H, v2.8B
-dup v22.8B, w3 // beta
rshrn v4.8B, v4.8H, #3
-cmhiv28.8B, v22.8B, v28.8B // < beta
-cmhiv30.8B, v22.8B, v30.8B // < beta
sminv4.8B, v4.8B, v24.8B
neg v25.8B, v24.8B
-and v26.8B, v26.8B, v28.8B
smaxv4.8B, v4.8B, v25.8B
-and v26.8B, v26.8B, v30.8B
uxtlv22.8H, v0.8B
and v4.8B, v4.8B, v26.8B
uxtlv28.8H, v16.8B
@@ -224,7 +229,7 @@ function ff_h264_v_loop_filter_chroma_neon, export=1
sub x0, x0, x1, lsl #1
st1 {v16.8B}, [x0], x1
st1 {v0.8B}, [x0], x1
-
+9:
ret
endfunc
@@ -257,7 +262,7 @@ function ff_h264_h_loop_filter_chroma_neon, export=1
st1 {v16.S}[1], [x0], x1
st1 {v0.S}[1], [x0], x1
st1 {v2.S}[1], [x0], x1
-
+9:
ret
endfunc
--
2.20.1
___
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
