[FFmpeg-devel] avcodec/mips: preload data in hevc sao edge 135 degree filter msa functions

Submitted by Manojkumar Bhosale on Oct. 9, 2017, 8:43 a.m.

Details

Message ID 70293ACCC3BA6A4E81FFCA024C7A86E1E05942CE@PUMAIL01.pu.imgtec.org
State New
Headers show

Commit Message

Manojkumar Bhosale Oct. 9, 2017, 8:43 a.m.
LGTM

-----Original Message-----
From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of kaustubh.raste@imgtec.com

Sent: Monday, October 9, 2017 11:31 AM
To: ffmpeg-devel@ffmpeg.org
Cc: Kaustubh Raste
Subject: [FFmpeg-devel] [PATCH] avcodec/mips: preload data in hevc sao edge 135 degree filter msa functions

From: Kaustubh Raste <kaustubh.raste@imgtec.com>


Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>

---
 libavcodec/mips/hevc_lpf_sao_msa.c |  194 ++++++++++++++++++++++++------------
 1 file changed, 132 insertions(+), 62 deletions(-)

         offset_mask1 = (v8i16) (__msa_hadd_u_h(diff_minus11, diff_minus11) + 2);
 
-        VSHF_H2_SH(edge_idx, edge_idx, sao_offset, sao_offset, offset_mask0,
-                   offset_mask0, offset_mask0, offset_mask0);
-        VSHF_H2_SH(edge_idx, edge_idx, sao_offset, sao_offset, offset_mask1,
-                   offset_mask1, offset_mask1, offset_mask1);
-        ILVEV_B2_SH(src_zero0, zero, src_zero1, zero, src00, src01);
-        ADD2(offset_mask0, src00, offset_mask1, src01, offset_mask0,
-             offset_mask1);
-        CLIP_SH2_0_255(offset_mask0, offset_mask1);
-        dst0 = __msa_pckev_b((v16i8) offset_mask1, (v16i8) offset_mask0);
+        offset = __msa_pckev_b((v16i8) offset_mask1, (v16i8) offset_mask0);
+        dst0 = __msa_pckev_b((v16i8) src_zero1, (v16i8) src_zero0);
+
+        VSHF_B2_SB(edge_idx, edge_idx, sao_offset, sao_offset, offset, offset,
+                   offset, offset);
+
+        dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+        dst0 = __msa_adds_s_b(dst0, offset);
+        dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
 
         src_minus10 = src10;
         src_minus11 = src11;
 
+        /* load in advance */
+        LD_UB2(src_orig + src_stride, src_stride, src10, src11);
+
         dst_val0 = __msa_copy_u_w((v4i32) dst0, 0);
         dst_val1 = __msa_copy_u_w((v4i32) dst0, 2);
 
@@ -2286,8 +2290,46 @@ static void hevc_sao_edge_filter_135degree_4width_msa(uint8_t *dst,
         SW(dst_val1, dst);
 
         dst += dst_stride;
-        src_orig += (src_stride << 1);
     }
+
+    SLDI_B2_0_SB(src_minus11, src10, src_zero0, src_zero1, 1);
+    SLDI_B2_0_UB(src_minus10, src_minus11, src_minus10, src_minus11, 
+ 2);
+
+    ILVR_B2_UB(src10, src_minus10, src11, src_minus11, src_minus10,
+               src_minus11);
+    ILVR_B2_SB(src_zero0, src_zero0, src_zero1, src_zero1, src_zero0,
+               src_zero1);
+
+    cmp_minus10 = ((v16u8) src_zero0 == src_minus10);
+    diff_minus10 = __msa_nor_v(cmp_minus10, cmp_minus10);
+    cmp_minus10 = (src_minus10 < (v16u8) src_zero0);
+    diff_minus10 = __msa_bmnz_v(diff_minus10, const1, cmp_minus10);
+
+    cmp_minus11 = ((v16u8) src_zero1 == src_minus11);
+    diff_minus11 = __msa_nor_v(cmp_minus11, cmp_minus11);
+    cmp_minus11 = (src_minus11 < (v16u8) src_zero1);
+    diff_minus11 = __msa_bmnz_v(diff_minus11, const1, cmp_minus11);
+
+    offset_mask0 = (v8i16) (__msa_hadd_u_h(diff_minus10, diff_minus10) + 2);
+    offset_mask1 = (v8i16) (__msa_hadd_u_h(diff_minus11, diff_minus11) 
+ + 2);
+
+    offset = __msa_pckev_b((v16i8) offset_mask1, (v16i8) offset_mask0);
+    dst0 = __msa_pckev_b((v16i8) src_zero1, (v16i8) src_zero0);
+
+    VSHF_B2_SB(edge_idx, edge_idx, sao_offset, sao_offset, offset, offset,
+               offset, offset);
+
+    dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+    dst0 = __msa_adds_s_b(dst0, offset);
+    dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+
+    dst_val0 = __msa_copy_u_w((v4i32) dst0, 0);
+    dst_val1 = __msa_copy_u_w((v4i32) dst0, 2);
+
+    SW(dst_val0, dst);
+    dst += dst_stride;
+    SW(dst_val1, dst);
+    dst += dst_stride;
 }
 
 static void hevc_sao_edge_filter_135degree_8width_msa(uint8_t *dst, @@ -2298,23 +2340,24 @@ static void hevc_sao_edge_filter_135degree_8width_msa(uint8_t *dst,
                                                       int32_t height)  {
     uint8_t *src_orig;
-    int32_t h_cnt;
     uint64_t dst_val0, dst_val1;
-    v8i16 edge_idx = { 1, 2, 0, 3, 4, 0, 0, 0 };
+    v16i8 edge_idx = { 1, 2, 0, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 
+ };
     v16u8 const1 = (v16u8) __msa_ldi_b(1);
-    v16i8 zero = { 0 };
-    v16i8 src_zero0, src_zero1, dst0, dst1;
+    v16i8 offset, sao_offset = LD_SB(sao_offset_val);
     v16u8 cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
     v16u8 src_minus10, src10, src_minus11, src11;
-    v8i16 sao_offset, src00, offset_mask0, src01, offset_mask1;
+    v16i8 src_zero0, src_zero1, dst0;
+    v8i16 offset_mask0, offset_mask1;
 
-    sao_offset = LD_SH(sao_offset_val);
+    sao_offset = __msa_pckev_b(sao_offset, sao_offset);
     src_orig = src - 1;
 
+    /* load in advance */
     LD_UB2(src_orig - src_stride, src_stride, src_minus10, src_minus11);
+    LD_UB2(src_orig + src_stride, src_stride, src10, src11);
 
-    for (h_cnt = (height >> 1); h_cnt--;) {
-        LD_UB2(src_orig + src_stride, src_stride, src10, src11);
+    for (height -= 2; height; height -= 2) {
+        src_orig += (src_stride << 1);
 
         SLDI_B2_0_SB(src_minus11, src10, src_zero0, src_zero1, 1);
         SLDI_B2_0_UB(src_minus10, src_minus11, src_minus10, src_minus11, 2); @@ -2336,30 +2379,68 @@ static void hevc_sao_edge_filter_135degree_8width_msa(uint8_t *dst,
         offset_mask0 = (v8i16) (__msa_hadd_u_h(diff_minus10, diff_minus10) + 2);
         offset_mask1 = (v8i16) (__msa_hadd_u_h(diff_minus11, diff_minus11) + 2);
 
-        VSHF_H2_SH(edge_idx, edge_idx, sao_offset, sao_offset, offset_mask0,
-                   offset_mask0, offset_mask0, offset_mask0);
-        VSHF_H2_SH(edge_idx, edge_idx, sao_offset, sao_offset, offset_mask1,
-                   offset_mask1, offset_mask1, offset_mask1);
-        ILVEV_B2_SH(src_zero0, zero, src_zero1, zero, src00, src01);
-        ADD2(offset_mask0, src00, offset_mask1, src01, offset_mask0,
-             offset_mask1);
-        CLIP_SH2_0_255(offset_mask0, offset_mask1);
-        PCKEV_B2_SB(offset_mask0, offset_mask0, offset_mask1, offset_mask1,
-                    dst0, dst1);
+        offset = __msa_pckev_b((v16i8) offset_mask1, (v16i8) offset_mask0);
+        dst0 = __msa_pckev_b((v16i8) src_zero1, (v16i8) src_zero0);
+
+        VSHF_B2_SB(edge_idx, edge_idx, sao_offset, sao_offset, offset, offset,
+                   offset, offset);
+
+        dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+        dst0 = __msa_adds_s_b(dst0, offset);
+        dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
 
         src_minus10 = src10;
         src_minus11 = src11;
 
+        /* load in advance */
+        LD_UB2(src_orig + src_stride, src_stride, src10, src11);
+
         dst_val0 = __msa_copy_u_d((v2i64) dst0, 0);
-        dst_val1 = __msa_copy_u_d((v2i64) dst1, 0);
+        dst_val1 = __msa_copy_u_d((v2i64) dst0, 1);
 
         SD(dst_val0, dst);
         dst += dst_stride;
         SD(dst_val1, dst);
         dst += dst_stride;
-
-        src_orig += (src_stride << 1);
     }
+
+    SLDI_B2_0_SB(src_minus11, src10, src_zero0, src_zero1, 1);
+    SLDI_B2_0_UB(src_minus10, src_minus11, src_minus10, src_minus11, 2);
+    ILVR_B2_UB(src10, src_minus10, src11, src_minus11, src_minus10,
+               src_minus11);
+    ILVR_B2_SB(src_zero0, src_zero0, src_zero1, src_zero1, src_zero0,
+               src_zero1);
+
+    cmp_minus10 = ((v16u8) src_zero0 == src_minus10);
+    diff_minus10 = __msa_nor_v(cmp_minus10, cmp_minus10);
+    cmp_minus10 = (src_minus10 < (v16u8) src_zero0);
+    diff_minus10 = __msa_bmnz_v(diff_minus10, const1, cmp_minus10);
+
+    cmp_minus11 = ((v16u8) src_zero1 == src_minus11);
+    diff_minus11 = __msa_nor_v(cmp_minus11, cmp_minus11);
+    cmp_minus11 = (src_minus11 < (v16u8) src_zero1);
+    diff_minus11 = __msa_bmnz_v(diff_minus11, const1, cmp_minus11);
+
+    offset_mask0 = (v8i16) (__msa_hadd_u_h(diff_minus10, diff_minus10) + 2);
+    offset_mask1 = (v8i16) (__msa_hadd_u_h(diff_minus11, diff_minus11) 
+ + 2);
+
+    offset = __msa_pckev_b((v16i8) offset_mask1, (v16i8) offset_mask0);
+    dst0 = __msa_pckev_b((v16i8) src_zero1, (v16i8) src_zero0);
+
+    VSHF_B2_SB(edge_idx, edge_idx, sao_offset, sao_offset, offset, offset,
+               offset, offset);
+
+    dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+    dst0 = __msa_adds_s_b(dst0, offset);
+    dst0 = (v16i8) __msa_xori_b((v16u8) dst0, 128);
+
+    dst_val0 = __msa_copy_u_d((v2i64) dst0, 0);
+    dst_val1 = __msa_copy_u_d((v2i64) dst0, 1);
+
+    SD(dst_val0, dst);
+    dst += dst_stride;
+    SD(dst_val1, dst);
+    dst += dst_stride;
 }
 
 static void hevc_sao_edge_filter_135degree_16multiple_msa(uint8_t *dst, @@ -2372,7 +2453,7 @@ static void hevc_sao_edge_filter_135degree_16multiple_msa(uint8_t *dst,
                                                           int32_t height)  {
     uint8_t *src_orig, *dst_orig;
-    int32_t h_cnt, v_cnt;
+    int32_t v_cnt;
     v16i8 edge_idx = { 1, 2, 0, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
     v16u8 const1 = (v16u8) __msa_ldi_b(1);
     v16u8 dst0, dst1, dst2, dst3;
@@ -2383,20 +2464,18 @@ static void hevc_sao_edge_filter_135degree_16multiple_msa(uint8_t *dst,
     v16u8 src_plus10, src_plus11, src_plus12, src_plus13;
     v16i8 src_minus12, src_minus13, src_zero0, src_zero1, src_zero2, src_zero3;
     v16i8 offset_mask0, offset_mask1, offset_mask2, offset_mask3, sao_offset;
-    v8i16 src0, src1, src2, src3, src4, src5, src6, src7;
-    v8i16 temp0, temp1, temp2, temp3, temp4, temp5, temp6, temp7;
 
     sao_offset = LD_SB(sao_offset_val);
     sao_offset = __msa_pckev_b(sao_offset, sao_offset);
 
-    for (h_cnt = (height >> 2); h_cnt--;) {
+    for (; height; height -= 4) {
         src_orig = src - 1;
         dst_orig = dst;
 
-        LD_UB4(src_orig, src_stride,
-               src_minus11, src_plus10, src_plus11, src_plus12);
+        LD_UB4(src_orig, src_stride, src_minus11, src_plus10, src_plus11,
+               src_plus12);
 
-        for (v_cnt = 0; v_cnt < (width >> 4); v_cnt++) {
+        for (v_cnt = 0; v_cnt < width; v_cnt += 16) {
             src_minus10 = LD_UB(src_orig + 2 - src_stride);
             LD_UB4(src_orig + 16, src_stride, src10, src11, src12, src13);
             src_plus13 = LD_UB(src_orig + (src_stride << 2)); @@ -2463,23 +2542,14 @@ static void hevc_sao_edge_filter_135degree_16multiple_msa(uint8_t *dst,
             VSHF_B2_SB(edge_idx, edge_idx, sao_offset, sao_offset,
                        offset_mask3, offset_mask3, offset_mask3, offset_mask3);
 
-            UNPCK_UB_SH(src_zero0, src0, src1);
-            UNPCK_SB_SH(offset_mask0, temp0, temp1);
-            UNPCK_UB_SH(src_zero1, src2, src3);
-            UNPCK_SB_SH(offset_mask1, temp2, temp3);
-            UNPCK_UB_SH(src_zero2, src4, src5);
-            UNPCK_SB_SH(offset_mask2, temp4, temp5);
-            UNPCK_UB_SH(src_zero3, src6, src7);
-            UNPCK_SB_SH(offset_mask3, temp6, temp7);
-
-            ADD4(temp0, src0, temp1, src1, temp2, src2, temp3, src3, temp0,
-                 temp1, temp2, temp3);
-            ADD4(temp4, src4, temp5, src5, temp6, src6, temp7, src7, temp4,
-                 temp5, temp6, temp7);
-            CLIP_SH4_0_255(temp0, temp1, temp2, temp3);
-            CLIP_SH4_0_255(temp4, temp5, temp6, temp7);
-            PCKEV_B4_UB(temp1, temp0, temp3, temp2, temp5, temp4, temp7, temp6,
-                        dst0, dst1, dst2, dst3);
+            XORI_B4_128_SB(src_zero0, src_zero1, src_zero2, src_zero3);
+
+            dst0 = (v16u8) __msa_adds_s_b((v16i8) src_zero0, offset_mask0);
+            dst1 = (v16u8) __msa_adds_s_b((v16i8) src_zero1, offset_mask1);
+            dst2 = (v16u8) __msa_adds_s_b((v16i8) src_zero2, offset_mask2);
+            dst3 = (v16u8) __msa_adds_s_b((v16i8) src_zero3, 
+ offset_mask3);
+
+            XORI_B4_128_UB(dst0, dst1, dst2, dst3);
 
             src_minus11 = src10;
             src_plus10 = src11;
--
1.7.9.5

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Comments

Michael Niedermayer Oct. 10, 2017, 9:34 p.m.
On Mon, Oct 09, 2017 at 08:43:51AM +0000, Manojkumar Bhosale wrote:
> LGTM

applied

thx

[...]

Patch hide | download patch | download mbox

diff --git a/libavcodec/mips/hevc_lpf_sao_msa.c b/libavcodec/mips/hevc_lpf_sao_msa.c
index c192265..5b5537a 100644
--- a/libavcodec/mips/hevc_lpf_sao_msa.c
+++ b/libavcodec/mips/hevc_lpf_sao_msa.c
@@ -2226,23 +2226,24 @@  static void hevc_sao_edge_filter_135degree_4width_msa(uint8_t *dst,
                                                       int32_t height)  {
     uint8_t *src_orig;
-    int32_t h_cnt;
     uint32_t dst_val0, dst_val1;
-    v8i16 edge_idx = { 1, 2, 0, 3, 4, 0, 0, 0 };
+    v16i8 edge_idx = { 1, 2, 0, 3, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 
+ };
     v16u8 const1 = (v16u8) __msa_ldi_b(1);
-    v16i8 zero = { 0 };
+    v16i8 offset, sao_offset = LD_SB(sao_offset_val);
     v16i8 src_zero0, src_zero1, dst0;
     v16u8 cmp_minus10, diff_minus10, cmp_minus11, diff_minus11;
     v16u8 src_minus10, src10, src_minus11, src11;
-    v8i16 offset_mask0, offset_mask1, sao_offset, src00, src01;
+    v8i16 offset_mask0, offset_mask1;
 
-    sao_offset = LD_SH(sao_offset_val);
+    sao_offset = __msa_pckev_b(sao_offset, sao_offset);
     src_orig = src - 1;
 
+    /* load in advance */
     LD_UB2(src_orig - src_stride, src_stride, src_minus10, src_minus11);
+    LD_UB2(src_orig + src_stride, src_stride, src10, src11);
 
-    for (h_cnt = (height >> 1); h_cnt--;) {
-        LD_UB2(src_orig + src_stride, src_stride, src10, src11);
+    for (height -= 2; height; height -= 2) {
+        src_orig += (src_stride << 1);
 
         SLDI_B2_0_SB(src_minus11, src10, src_zero0, src_zero1, 1);
         SLDI_B2_0_UB(src_minus10, src_minus11, src_minus10, src_minus11, 2); @@ -2265,19 +2266,22 @@ static void hevc_sao_edge_filter_135degree_4width_msa(uint8_t *dst,
         offset_mask0 = (v8i16) (__msa_hadd_u_h(diff_minus10, diff_minus10) + 2);