[FFmpeg-devel] [PATCH] avcodec/mips: Unrolled loops and expanded functions in avc put mc 10 & 30 msa functions

kaustubh.raste at imgtec.com kaustubh.raste at imgtec.com
Mon Sep 18 11:30:13 EEST 2017


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

Signed-off-by: Kaustubh Raste <kaustubh.raste at imgtec.com>
---
 libavcodec/mips/h264qpel_msa.c |  284 +++++++++++++++++++++++++++++++++++++++-
 1 file changed, 278 insertions(+), 6 deletions(-)

diff --git a/libavcodec/mips/h264qpel_msa.c b/libavcodec/mips/h264qpel_msa.c
index 05dffea..b7f6c3d 100644
--- a/libavcodec/mips/h264qpel_msa.c
+++ b/libavcodec/mips/h264qpel_msa.c
@@ -3065,37 +3065,309 @@ void ff_avg_h264_qpel4_mc00_msa(uint8_t *dst, const uint8_t *src,
 void ff_put_h264_qpel16_mc10_msa(uint8_t *dst, const uint8_t *src,
                                  ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_16w_msa(src - 2, stride, dst, stride, 16, 0);
+    uint32_t loop_cnt;
+    v16i8 dst0, dst1, dst2, dst3, src0, src1, src2, src3, src4, src5, src6;
+    v16i8 mask0, mask1, mask2, mask3, mask4, mask5, src7, vec11;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10;
+    v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2);
+    mask3 = mask0 + 8;
+    mask4 = mask1 + 8;
+    mask5 = mask2 + 8;
+    src -= 2;
+
+    for (loop_cnt = 4; loop_cnt--;) {
+        LD_SB2(src, 16, src0, src1);
+        src += stride;
+        LD_SB2(src, 16, src2, src3);
+        src += stride;
+        LD_SB2(src, 16, src4, src5);
+        src += stride;
+        LD_SB2(src, 16, src6, src7);
+        src += stride;
+
+        XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+        VSHF_B2_SB(src0, src0, src0, src1, mask0, mask3, vec0, vec3);
+        VSHF_B2_SB(src2, src2, src2, src3, mask0, mask3, vec6, vec9);
+        VSHF_B2_SB(src0, src0, src0, src1, mask1, mask4, vec1, vec4);
+        VSHF_B2_SB(src2, src2, src2, src3, mask1, mask4, vec7, vec10);
+        VSHF_B2_SB(src0, src0, src0, src1, mask2, mask5, vec2, vec5);
+        VSHF_B2_SB(src2, src2, src2, src3, mask2, mask5, vec8, vec11);
+        HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3);
+        DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b,
+                     minus5b, res0, res1, res2, res3);
+        DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b,
+                     plus20b, res0, res1, res2, res3);
+        VSHF_B2_SB(src4, src4, src4, src5, mask0, mask3, vec0, vec3);
+        VSHF_B2_SB(src6, src6, src6, src7, mask0, mask3, vec6, vec9);
+        VSHF_B2_SB(src4, src4, src4, src5, mask1, mask4, vec1, vec4);
+        VSHF_B2_SB(src6, src6, src6, src7, mask1, mask4, vec7, vec10);
+        VSHF_B2_SB(src4, src4, src4, src5, mask2, mask5, vec2, vec5);
+        VSHF_B2_SB(src6, src6, src6, src7, mask2, mask5, vec8, vec11);
+        HADD_SB4_SH(vec0, vec3, vec6, vec9, res4, res5, res6, res7);
+        DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b,
+                     minus5b, res4, res5, res6, res7);
+        DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b,
+                     plus20b, res4, res5, res6, res7);
+        SLDI_B2_SB(src1, src3, src0, src2, src0, src2, 2);
+        SLDI_B2_SB(src5, src7, src4, src6, src4, src6, 2);
+        SRARI_H4_SH(res0, res1, res2, res3, 5);
+        SRARI_H4_SH(res4, res5, res6, res7, 5);
+        SAT_SH4_SH(res0, res1, res2, res3, 7);
+        SAT_SH4_SH(res4, res5, res6, res7, 7);
+        PCKEV_B2_SB(res1, res0, res3, res2, dst0, dst1);
+        PCKEV_B2_SB(res5, res4, res7, res6, dst2, dst3);
+        dst0 = __msa_aver_s_b(dst0, src0);
+        dst1 = __msa_aver_s_b(dst1, src2);
+        dst2 = __msa_aver_s_b(dst2, src4);
+        dst3 = __msa_aver_s_b(dst3, src6);
+        XORI_B4_128_SB(dst0, dst1, dst2, dst3);
+        ST_SB4(dst0, dst1, dst2, dst3, dst, stride);
+        dst += (4 * stride);
+    }
 }
 
 void ff_put_h264_qpel16_mc30_msa(uint8_t *dst, const uint8_t *src,
                                  ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_16w_msa(src - 2, stride, dst, stride, 16, 1);
+    uint32_t loop_cnt;
+    v16i8 dst0, dst1, dst2, dst3, src0, src1, src2, src3, src4, src5, src6;
+    v16i8 mask0, mask1, mask2, mask3, mask4, mask5, src7, vec11;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10;
+    v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2);
+    mask3 = mask0 + 8;
+    mask4 = mask1 + 8;
+    mask5 = mask2 + 8;
+    src -= 2;
+
+    for (loop_cnt = 4; loop_cnt--;) {
+        LD_SB2(src, 16, src0, src1);
+        src += stride;
+        LD_SB2(src, 16, src2, src3);
+        src += stride;
+        LD_SB2(src, 16, src4, src5);
+        src += stride;
+        LD_SB2(src, 16, src6, src7);
+        src += stride;
+
+        XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+        VSHF_B2_SB(src0, src0, src0, src1, mask0, mask3, vec0, vec3);
+        VSHF_B2_SB(src2, src2, src2, src3, mask0, mask3, vec6, vec9);
+        VSHF_B2_SB(src0, src0, src0, src1, mask1, mask4, vec1, vec4);
+        VSHF_B2_SB(src2, src2, src2, src3, mask1, mask4, vec7, vec10);
+        VSHF_B2_SB(src0, src0, src0, src1, mask2, mask5, vec2, vec5);
+        VSHF_B2_SB(src2, src2, src2, src3, mask2, mask5, vec8, vec11);
+        HADD_SB4_SH(vec0, vec3, vec6, vec9, res0, res1, res2, res3);
+        DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b,
+                     minus5b, res0, res1, res2, res3);
+        DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b,
+                     plus20b, res0, res1, res2, res3);
+        VSHF_B2_SB(src4, src4, src4, src5, mask0, mask3, vec0, vec3);
+        VSHF_B2_SB(src6, src6, src6, src7, mask0, mask3, vec6, vec9);
+        VSHF_B2_SB(src4, src4, src4, src5, mask1, mask4, vec1, vec4);
+        VSHF_B2_SB(src6, src6, src6, src7, mask1, mask4, vec7, vec10);
+        VSHF_B2_SB(src4, src4, src4, src5, mask2, mask5, vec2, vec5);
+        VSHF_B2_SB(src6, src6, src6, src7, mask2, mask5, vec8, vec11);
+        HADD_SB4_SH(vec0, vec3, vec6, vec9, res4, res5, res6, res7);
+        DPADD_SB4_SH(vec1, vec4, vec7, vec10, minus5b, minus5b, minus5b,
+                     minus5b, res4, res5, res6, res7);
+        DPADD_SB4_SH(vec2, vec5, vec8, vec11, plus20b, plus20b, plus20b,
+                     plus20b, res4, res5, res6, res7);
+        SLDI_B2_SB(src1, src3, src0, src2, src0, src2, 3);
+        SLDI_B2_SB(src5, src7, src4, src6, src4, src6, 3);
+        SRARI_H4_SH(res0, res1, res2, res3, 5);
+        SRARI_H4_SH(res4, res5, res6, res7, 5);
+        SAT_SH4_SH(res0, res1, res2, res3, 7);
+        SAT_SH4_SH(res4, res5, res6, res7, 7);
+        PCKEV_B2_SB(res1, res0, res3, res2, dst0, dst1);
+        PCKEV_B2_SB(res5, res4, res7, res6, dst2, dst3);
+        dst0 = __msa_aver_s_b(dst0, src0);
+        dst1 = __msa_aver_s_b(dst1, src2);
+        dst2 = __msa_aver_s_b(dst2, src4);
+        dst3 = __msa_aver_s_b(dst3, src6);
+        XORI_B4_128_SB(dst0, dst1, dst2, dst3);
+        ST_SB4(dst0, dst1, dst2, dst3, dst, stride);
+        dst += (4 * stride);
+    }
 }
 
 void ff_put_h264_qpel8_mc10_msa(uint8_t *dst, const uint8_t *src,
                                 ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_8w_msa(src - 2, stride, dst, stride, 8, 0);
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2;
+    v16i8 tmp0, tmp1, tmp2, tmp3, vec11;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10;
+    v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2);
+    LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+    VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1);
+    VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2, vec3);
+    HADD_SB4_SH(vec0, vec1, vec2, vec3, res0, res1, res2, res3);
+    VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4, vec5);
+    VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6, vec7);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b,
+                 res0, res1, res2, res3);
+    VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec8, vec9);
+    VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec10, vec11);
+    DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b,
+                 res0, res1, res2, res3);
+    VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec1);
+    VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec2, vec3);
+    HADD_SB4_SH(vec0, vec1, vec2, vec3, res4, res5, res6, res7);
+    VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec4, vec5);
+    VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec6, vec7);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b,
+                 res4, res5, res6, res7);
+    VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec8, vec9);
+    VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec10, vec11);
+    DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b,
+                 res4, res5, res6, res7);
+    SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 2);
+    SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 2);
+    SLDI_B2_SB(src4, src5, src4, src5, src4, src5, 2);
+    SLDI_B2_SB(src6, src7, src6, src7, src6, src7, 2);
+    PCKEV_D2_SB(src1, src0, src3, src2, src0, src1);
+    PCKEV_D2_SB(src5, src4, src7, src6, src4, src5);
+    SRARI_H4_SH(res0, res1, res2, res3, 5);
+    SRARI_H4_SH(res4, res5, res6, res7, 5);
+    SAT_SH4_SH(res0, res1, res2, res3, 7);
+    SAT_SH4_SH(res4, res5, res6, res7, 7);
+    PCKEV_B2_SB(res1, res0, res3, res2, tmp0, tmp1);
+    PCKEV_B2_SB(res5, res4, res7, res6, tmp2, tmp3);
+    tmp0 = __msa_aver_s_b(tmp0, src0);
+    tmp1 = __msa_aver_s_b(tmp1, src1);
+    tmp2 = __msa_aver_s_b(tmp2, src4);
+    tmp3 = __msa_aver_s_b(tmp3, src5);
+    XORI_B4_128_SB(tmp0, tmp1, tmp2, tmp3);
+    ST8x8_UB(tmp0, tmp1, tmp2, tmp3, dst, stride);
 }
 
 void ff_put_h264_qpel8_mc30_msa(uint8_t *dst, const uint8_t *src,
                                 ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_8w_msa(src - 2, stride, dst, stride, 8, 1);
+    v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask0, mask1, mask2;
+    v16i8 tmp0, tmp1, tmp2, tmp3, vec11;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9, vec10;
+    v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[0], 16, mask0, mask1, mask2);
+    LD_SB8(src - 2, stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    XORI_B8_128_SB(src0, src1, src2, src3, src4, src5, src6, src7);
+    VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0, vec1);
+    VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2, vec3);
+    HADD_SB4_SH(vec0, vec1, vec2, vec3, res0, res1, res2, res3);
+    VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4, vec5);
+    VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6, vec7);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b,
+                 res0, res1, res2, res3);
+    VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec8, vec9);
+    VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec10, vec11);
+    DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b,
+                 res0, res1, res2, res3);
+    VSHF_B2_SB(src4, src4, src5, src5, mask0, mask0, vec0, vec1);
+    VSHF_B2_SB(src6, src6, src7, src7, mask0, mask0, vec2, vec3);
+    HADD_SB4_SH(vec0, vec1, vec2, vec3, res4, res5, res6, res7);
+    VSHF_B2_SB(src4, src4, src5, src5, mask1, mask1, vec4, vec5);
+    VSHF_B2_SB(src6, src6, src7, src7, mask1, mask1, vec6, vec7);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, minus5b, minus5b, minus5b, minus5b,
+                 res4, res5, res6, res7);
+    VSHF_B2_SB(src4, src4, src5, src5, mask2, mask2, vec8, vec9);
+    VSHF_B2_SB(src6, src6, src7, src7, mask2, mask2, vec10, vec11);
+    DPADD_SB4_SH(vec8, vec9, vec10, vec11, plus20b, plus20b, plus20b, plus20b,
+                 res4, res5, res6, res7);
+    SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 3);
+    SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 3);
+    SLDI_B2_SB(src4, src5, src4, src5, src4, src5, 3);
+    SLDI_B2_SB(src6, src7, src6, src7, src6, src7, 3);
+    PCKEV_D2_SB(src1, src0, src3, src2, src0, src1);
+    PCKEV_D2_SB(src5, src4, src7, src6, src4, src5);
+    SRARI_H4_SH(res0, res1, res2, res3, 5);
+    SRARI_H4_SH(res4, res5, res6, res7, 5);
+    SAT_SH4_SH(res0, res1, res2, res3, 7);
+    SAT_SH4_SH(res4, res5, res6, res7, 7);
+    PCKEV_B2_SB(res1, res0, res3, res2, tmp0, tmp1);
+    PCKEV_B2_SB(res5, res4, res7, res6, tmp2, tmp3);
+    tmp0 = __msa_aver_s_b(tmp0, src0);
+    tmp1 = __msa_aver_s_b(tmp1, src1);
+    tmp2 = __msa_aver_s_b(tmp2, src4);
+    tmp3 = __msa_aver_s_b(tmp3, src5);
+    XORI_B4_128_SB(tmp0, tmp1, tmp2, tmp3);
+    ST8x8_UB(tmp0, tmp1, tmp2, tmp3, dst, stride);
 }
 
 void ff_put_h264_qpel4_mc10_msa(uint8_t *dst, const uint8_t *src,
                                 ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_4w_msa(src - 2, stride, dst, stride, 4, 0);
+    v16i8 src0, src1, src2, src3, res, mask0, mask1, mask2;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5;
+    v8i16 res0, res1;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[48], 16, mask0, mask1, mask2);
+    LD_SB4(src - 2, stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1);
+    HADD_SB2_SH(vec0, vec1, res0, res1);
+    VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3);
+    DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, res0, res1);
+    VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5);
+    DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, res0, res1);
+    SRARI_H2_SH(res0, res1, 5);
+    SAT_SH2_SH(res0, res1, 7);
+    res = __msa_pckev_b((v16i8) res1, (v16i8) res0);
+    SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 2);
+    SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 2);
+    src0 = (v16i8) __msa_insve_w((v4i32) src0, 1, (v4i32) src1);
+    src1 = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3);
+    src0 = (v16i8) __msa_insve_d((v2i64) src0, 1, (v2i64) src1);
+    res = __msa_aver_s_b(res, src0);
+    res = (v16i8) __msa_xori_b((v16u8) res, 128);
+    ST4x4_UB(res, res, 0, 1, 2, 3, dst, stride);
 }
 
 void ff_put_h264_qpel4_mc30_msa(uint8_t *dst, const uint8_t *src,
                                 ptrdiff_t stride)
 {
-    avc_luma_hz_qrt_4w_msa(src - 2, stride, dst, stride, 4, 1);
+    v16i8 src0, src1, src2, src3, res, mask0, mask1, mask2;
+    v16i8 vec0, vec1, vec2, vec3, vec4, vec5;
+    v8i16 res0, res1;
+    v16i8 minus5b = __msa_ldi_b(-5);
+    v16i8 plus20b = __msa_ldi_b(20);
+
+    LD_SB3(&luma_mask_arr[48], 16, mask0, mask1, mask2);
+    LD_SB4(src - 2, stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0, vec1);
+    HADD_SB2_SH(vec0, vec1, res0, res1);
+    VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2, vec3);
+    DPADD_SB2_SH(vec2, vec3, minus5b, minus5b, res0, res1);
+    VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4, vec5);
+    DPADD_SB2_SH(vec4, vec5, plus20b, plus20b, res0, res1);
+    SRARI_H2_SH(res0, res1, 5);
+    SAT_SH2_SH(res0, res1, 7);
+    res = __msa_pckev_b((v16i8) res1, (v16i8) res0);
+    SLDI_B2_SB(src0, src1, src0, src1, src0, src1, 3);
+    SLDI_B2_SB(src2, src3, src2, src3, src2, src3, 3);
+    src0 = (v16i8) __msa_insve_w((v4i32) src0, 1, (v4i32) src1);
+    src1 = (v16i8) __msa_insve_w((v4i32) src2, 1, (v4i32) src3);
+    src0 = (v16i8) __msa_insve_d((v2i64) src0, 1, (v2i64) src1);
+    res = __msa_aver_s_b(res, src0);
+    res = (v16i8) __msa_xori_b((v16u8) res, 128);
+    ST4x4_UB(res, res, 0, 1, 2, 3, dst, stride);
 }
 
 void ff_put_h264_qpel16_mc20_msa(uint8_t *dst, const uint8_t *src,
-- 
1.7.9.5



More information about the ffmpeg-devel mailing list