[FFmpeg-devel] [PATCH] avcodec/mips: Unrolled loops and expanded functions in avc put mc 10 & 30 msa functions
Manojkumar Bhosale
Manojkumar.Bhosale at imgtec.com
Tue Sep 19 14:14:56 EEST 2017
LGTM
-----Original Message-----
From: ffmpeg-devel [mailto:ffmpeg-devel-bounces at ffmpeg.org] On Behalf Of kaustubh.raste at imgtec.com
Sent: Monday, September 18, 2017 2:00 PM
To: ffmpeg-devel at ffmpeg.org
Cc: Kaustubh Raste
Subject: [FFmpeg-devel] [PATCH] avcodec/mips: Unrolled loops and expanded functions in avc put mc 10 & 30 msa functions
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
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel at ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
More information about the ffmpeg-devel
mailing list