[FFmpeg-devel] [PATCH 2/6] avcodec/h264: add avx 8-bit 4:2:0 chroma h deblock/loop filter

[James Darnley]

~1.14x faster (93 vs. 81 cycles) compared with mmxext function
---
 libavcodec/x86/h264_deblock.asm | 70 +++++++++++++++++++++++++++++++++++++++++
 libavcodec/x86/h264dsp_init.c   |  3 ++
 2 files changed, 73 insertions(+)

diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm
index 2e84ca3..0465c9f 100644
--- a/libavcodec/x86/h264_deblock.asm
+++ b/libavcodec/x86/h264_deblock.asm
@@ -1059,6 +1059,54 @@ ff_chroma_intra_body_mmxext:
     paddb  m2, m6
     ret
 
+%macro LOAD_8_ROWS 8
+    movd m0, %1
+    movd m1, %2
+    movd m2, %3
+    movd m3, %4
+    movd m4, %5
+    movd m5, %6
+    movd m6, %7
+    movd m7, %8
+%endmacro
+
+%macro STORE_8_ROWS 8
+    movd %1, m0
+    movd %2, m1
+    movd %3, m2
+    movd %4, m3
+    movd %5, m4
+    movd %6, m5
+    movd %7, m6
+    movd %8, m7
+%endmacro
+
+%macro TRANSPOSE_8x4B_XMM 0
+    punpcklbw m0, m1
+    punpcklbw m2, m3
+    punpcklbw m4, m5
+    punpcklbw m6, m7
+    punpcklwd m0, m2
+    punpcklwd m4, m6
+    punpckhdq m2, m0, m4
+    punpckldq m0, m4
+    MOVHL m1, m0
+    MOVHL m3, m2
+%endmacro
+
+%macro TRANSPOSE_4x8B_XMM 0
+    punpcklbw m0, m1
+    punpcklbw m2, m3
+    punpckhwd m4, m0, m2
+    punpcklwd m0, m2
+    MOVHL m6, m4
+    MOVHL m2, m0
+    pshufd m1, m0, 1
+    pshufd m3, m2, 1
+    pshufd m5, m4, 1
+    pshufd m7, m6, 1
+%endmacro
+
 %macro CHROMA_INTER_BODY_XMM 1
     LOAD_MASK alpha_d, beta_d
     movd m6, [tc0_q]
@@ -1078,6 +1126,15 @@ ff_chroma_intra_body_mmxext:
     sub %1, stride_q
 %endmacro
 
+%macro CHROMA_H_START_XMM 2
+    movsxdifnidn stride_q, stride_d
+    dec alpha_d
+    dec beta_d
+    lea %2, [3*stride_q]
+    mov %1,  pix_q
+    add %1,  %2
+%endmacro
+
 %macro DEBLOCK_CHROMA_XMM 1
 
 INIT_XMM %1
@@ -1093,6 +1150,19 @@ cglobal deblock_v_chroma_8, 5, 6, 8, pix_, stride_, alpha_, beta_, tc0_
     movq [pix_q], m2
 RET
 
+cglobal deblock_h_chroma_8, 5, 7, 8, 0-16, pix_, stride_, alpha_, beta_, tc0_
+    CHROMA_H_START_XMM r5, r6
+    LOAD_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6)
+    TRANSPOSE_8x4B_XMM
+    movq [rsp], m0
+    movq [rsp + 8], m3
+    CHROMA_INTER_BODY_XMM 1
+    movq m0, [rsp]
+    movq m3, [rsp + 8]
+    TRANSPOSE_4x8B_XMM
+    STORE_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6)
+RET
+
 %endmacro ; DEBLOCK_CHROMA_XMM
 
 DEBLOCK_CHROMA_XMM avx
diff --git a/libavcodec/x86/h264dsp_init.c b/libavcodec/x86/h264dsp_init.c
index 6794aa5..0b15471 100644
--- a/libavcodec/x86/h264dsp_init.c
+++ b/libavcodec/x86/h264dsp_init.c
@@ -319,6 +319,9 @@ av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth,
 #endif
 
             c->h264_v_loop_filter_chroma       = ff_deblock_v_chroma_8_avx;
+            if (chroma_format_idc <= 1) {
+                c->h264_h_loop_filter_chroma       = ff_deblock_h_chroma_8_avx;
+            }
         }
     } else if (bit_depth == 10) {
         if (EXTERNAL_MMXEXT(cpu_flags)) {
-- 
2.8.3