mips msa vp9 filter by weight optimization

filter by weight - average improvement ~2x-3x

Change-Id: I4832033335d339cdafdce697f07ce3e643920057

diff --git a/vp9/common/mips/msa/vp9_macros_msa.h b/vp9/common/mips/msa/vp9_macros_msa.h
index f1217d5..2043e13 100644 (file)
--- a/vp9/common/mips/msa/vp9_macros_msa.h
+++ b/vp9/common/mips/msa/vp9_macros_msa.h
@@ -244,6 +244,22 @@
   out3 = LW((psrc) + 3 * stride);                    \
 }
 
+/* Description : Load double words with stride
+   Arguments   : Inputs  - psrc    (source pointer to load from)
+                         - stride
+                 Outputs - out0, out1
+   Details     : Loads double word in 'out0' from (psrc)
+                 Loads double word in 'out1' from (psrc + stride)
+*/
+#define LD2(psrc, stride, out0, out1) {  \
+  out0 = LD((psrc));                     \
+  out1 = LD((psrc) + stride);            \
+}
+#define LD4(psrc, stride, out0, out1, out2, out3) {  \
+  LD2((psrc), stride, out0, out1);                   \
+  LD2((psrc) + 2 * stride, stride, out2, out3);      \
+}
+
 /* Description : Store 4 words with stride
    Arguments   : Inputs  - in0, in1, in2, in3, pdst, stride
    Details     : Stores word from 'in0' to (pdst)
@@ -482,6 +498,24 @@
   SD(out0_m, pdst);                       \
 }
 
+/* Description : Store as 8x2 byte block to destination memory from input vector
+   Arguments   : Inputs  - in, pdst, stride
+   Details     : Index 0 double word element from input vector 'in' is copied
+                 and stored to destination memory at (pdst)
+                 Index 1 double word element from input vector 'in' is copied
+                 and stored to destination memory at (pdst + stride)
+*/
+#define ST8x2_UB(in, pdst, stride) {        \
+  uint64_t out0_m, out1_m;                  \
+  uint8_t *pblk_8x2_m = (uint8_t *)(pdst);  \
+                                            \
+  out0_m = __msa_copy_u_d((v2i64)in, 0);    \
+  out1_m = __msa_copy_u_d((v2i64)in, 1);    \
+                                            \
+  SD(out0_m, pblk_8x2_m);                   \
+  SD(out1_m, pblk_8x2_m + stride);          \
+}
+
 /* Description : Store as 8x4 byte block to destination memory from input
                  vectors
    Arguments   : Inputs  - in0, in1, pdst, stride

diff --git a/vp9/common/mips/msa/vp9_mfqe_msa.c b/vp9/common/mips/msa/vp9_mfqe_msa.c
new file mode 100644 (file)
index 0000000..64cb9a8
--- /dev/null
+++ b/vp9/common/mips/msa/vp9_mfqe_msa.c
@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/mips/msa/vp9_macros_msa.h"
+
+static void filter_by_weight8x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    uint8_t *dst_ptr, int32_t dst_stride,
+                                    int32_t src_weight) {
+  int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  int32_t row;
+  uint64_t src0_d, src1_d, dst0_d, dst1_d;
+  v16i8 src0 = { 0 };
+  v16i8 src1 = { 0 };
+  v16i8 dst0 = { 0 };
+  v16i8 dst1 = { 0 };
+  v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
+
+  src_wt = __msa_fill_h(src_weight);
+  dst_wt = __msa_fill_h(dst_weight);
+
+  for (row = 2; row--;) {
+    LD2(src_ptr, src_stride, src0_d, src1_d);
+    src_ptr += (2 * src_stride);
+    LD2(dst_ptr, dst_stride, dst0_d, dst1_d);
+    INSERT_D2_SB(src0_d, src1_d, src0);
+    INSERT_D2_SB(dst0_d, dst1_d, dst0);
+
+    LD2(src_ptr, src_stride, src0_d, src1_d);
+    src_ptr += (2 * src_stride);
+    LD2((dst_ptr + 2 * dst_stride), dst_stride, dst0_d, dst1_d);
+    INSERT_D2_SB(src0_d, src1_d, src1);
+    INSERT_D2_SB(dst0_d, dst1_d, dst1);
+
+    UNPCK_UB_SH(src0, src_r, src_l);
+    UNPCK_UB_SH(dst0, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    dst0 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+    ST8x2_UB(dst0, dst_ptr, dst_stride);
+    dst_ptr += (2 * dst_stride);
+
+    UNPCK_UB_SH(src1, src_r, src_l);
+    UNPCK_UB_SH(dst1, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    dst1 = (v16i8)__msa_pckev_b((v16i8)res_h_l, (v16i8)res_h_r);
+    ST8x2_UB(dst1, dst_ptr, dst_stride);
+    dst_ptr += (2 * dst_stride);
+  }
+}
+
+static void filter_by_weight16x16_msa(const uint8_t *src_ptr,
+                                      int32_t src_stride,
+                                      uint8_t *dst_ptr,
+                                      int32_t dst_stride,
+                                      int32_t src_weight) {
+  int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
+  int32_t row;
+  v16i8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
+
+  src_wt = __msa_fill_h(src_weight);
+  dst_wt = __msa_fill_h(dst_weight);
+
+  for (row = 4; row--;) {
+    LD_SB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_SB4(dst_ptr, dst_stride, dst0, dst1, dst2, dst3);
+
+    UNPCK_UB_SH(src0, src_r, src_l);
+    UNPCK_UB_SH(dst0, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src1, src_r, src_l);
+    UNPCK_UB_SH(dst1, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src2, src_r, src_l);
+    UNPCK_UB_SH(dst2, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+
+    UNPCK_UB_SH(src3, src_r, src_l);
+    UNPCK_UB_SH(dst3, dst_r, dst_l);
+    res_h_r = (src_r * src_wt);
+    res_h_r += (dst_r * dst_wt);
+    res_h_l = (src_l * src_wt);
+    res_h_l += (dst_l * dst_wt);
+    SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
+    PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
+    dst_ptr += dst_stride;
+  }
+}
+
+void vp9_filter_by_weight8x8_msa(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 int src_weight) {
+  filter_by_weight8x8_msa(src, src_stride, dst, dst_stride, src_weight);
+}
+
+void vp9_filter_by_weight16x16_msa(const uint8_t *src, int src_stride,
+                                   uint8_t *dst, int dst_stride,
+                                   int src_weight) {
+  filter_by_weight16x16_msa(src, src_stride, dst, dst_stride, src_weight);
+}

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 5035126..2f262a6 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -276,10 +276,10 @@ specialize qw/vp9_plane_add_noise sse2/;
 $vp9_plane_add_noise_sse2=vp9_plane_add_noise_wmt;
 
 add_proto qw/void vp9_filter_by_weight16x16/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
-specialize qw/vp9_filter_by_weight16x16 sse2/;
+specialize qw/vp9_filter_by_weight16x16 sse2 msa/;
 
 add_proto qw/void vp9_filter_by_weight8x8/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int src_weight";
-specialize qw/vp9_filter_by_weight8x8 sse2/;
+specialize qw/vp9_filter_by_weight8x8 sse2 msa/;
 }
 
 #

diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index b01fdd1..6f091ee 100644 (file)
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -152,6 +152,10 @@ VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_8_msa.c
 VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_16_msa.c
 VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_msa.h
 
+ifeq ($(CONFIG_VP9_POSTPROC),yes)
+VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_mfqe_msa.c
+endif
+
 VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
 VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.h
 ifeq ($(ARCH_X86_64), yes)