Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

avcodec/mips: Refine ff_h264_h_lpf_luma_inter_msa

Browse Source 
Using mask to avoid judgment, H264 4K decoding speed
improved about 0.1fps tested on 3A4000

Signed-off-by: Shiyou Yin <yinshiyou-hf@loongson.cn>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
pull/362/head
gxw 4 years ago committed by Michael Niedermayer
parent 6458c6bdb4
commit 464d28c070
1 changed files with 171 additions and 294 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 463
      libavcodec/mips/h264dsp_msa.c

463
libavcodec/mips/h264dsp_msa.c
Unescape View File

@ -1284,284 +1284,160 @@ static void avc_loopfilter_cb_or_cr_intra_edge_ver_msa(uint8_t *data_cb_or_cr,
} }
} }
static void avc_loopfilter_luma_inter_edge_ver_msa(uint8_t *data, static void avc_loopfilter_luma_inter_edge_ver_msa(uint8_t* pPix, uint32_t iStride,
uint8_t bs0, uint8_t bs1, uint8_t iAlpha, uint8_t iBeta,
uint8_t bs2, uint8_t bs3, uint8_t* pTc)
uint8_t tc0, uint8_t tc1,
uint8_t tc2, uint8_t tc3,
uint8_t alpha_in,
uint8_t beta_in,
ptrdiff_t img_width)
{ {
v16u8 tmp_vec, bs = { 0 }; v16u8 p0, p1, p2, q0, q1, q2;
v16i8 iTc, negiTc, negTc, flags, f;
tmp_vec = (v16u8) __msa_fill_b(bs0); v8i16 p0_l, p0_r, p1_l, p1_r, p2_l, p2_r, q0_l, q0_r, q1_l, q1_r, q2_l, q2_r;
bs = (v16u8) __msa_insve_w((v4i32) bs, 0, (v4i32) tmp_vec); v8i16 tc_l, tc_r, negTc_l, negTc_r;
tmp_vec = (v16u8) __msa_fill_b(bs1); v8i16 iTc_l, iTc_r, negiTc_l, negiTc_r;
bs = (v16u8) __msa_insve_w((v4i32) bs, 1, (v4i32) tmp_vec); // Use for temporary variable
tmp_vec = (v16u8) __msa_fill_b(bs2); v8i16 t0, t1, t2, t3;
bs = (v16u8) __msa_insve_w((v4i32) bs, 2, (v4i32) tmp_vec); v16u8 alpha, beta;
tmp_vec = (v16u8) __msa_fill_b(bs3); v16u8 bDetaP0Q0, bDetaP1P0, bDetaQ1Q0, bDetaP2P0, bDetaQ2Q0;
bs = (v16u8) __msa_insve_w((v4i32) bs, 3, (v4i32) tmp_vec); v16i8 const_1_b = __msa_ldi_b(1);
v8i16 const_1_h = __msa_ldi_h(1);
if (!__msa_test_bz_v(bs)) { v8i16 const_4_h = __msa_ldi_h(4);
uint8_t *src = data - 4; v8i16 const_not_255_h = __msa_ldi_h(~255);
v16u8 p3_org, p2_org, p1_org, p0_org, q0_org, q1_org, q2_org, q3_org;
v16u8 p0_asub_q0, p1_asub_p0, q1_asub_q0, alpha, beta;
v16u8 is_less_than, is_less_than_beta, is_less_than_alpha;
v16u8 is_bs_greater_than0;
v16u8 tc = { 0 };
v16i8 zero = { 0 }; v16i8 zero = { 0 };
v16i8 tc = { pTc[0 >> 2], pTc[1 >> 2], pTc[2 >> 2], pTc[3 >> 2],
tmp_vec = (v16u8) __msa_fill_b(tc0); pTc[4 >> 2], pTc[5 >> 2], pTc[6 >> 2], pTc[7 >> 2],
tc = (v16u8) __msa_insve_w((v4i32) tc, 0, (v4i32) tmp_vec); pTc[8 >> 2], pTc[9 >> 2], pTc[10 >> 2], pTc[11 >> 2],
tmp_vec = (v16u8) __msa_fill_b(tc1); pTc[12 >> 2], pTc[13 >> 2], pTc[14 >> 2], pTc[15 >> 2] };
tc = (v16u8) __msa_insve_w((v4i32) tc, 1, (v4i32) tmp_vec); negTc = zero - tc;
tmp_vec = (v16u8) __msa_fill_b(tc2); iTc = tc;
tc = (v16u8) __msa_insve_w((v4i32) tc, 2, (v4i32) tmp_vec);
tmp_vec = (v16u8) __msa_fill_b(tc3); // Load data from pPix
tc = (v16u8) __msa_insve_w((v4i32) tc, 3, (v4i32) tmp_vec); LD_SH8(pPix - 3, iStride, t0, t1, t2, t3, q1_l, q1_r, q2_l, q2_r);
LD_SH8(pPix + 8 * iStride - 3, iStride, p0_l, p0_r, p1_l, p1_r,
is_bs_greater_than0 = (zero < bs); p2_l, p2_r, q0_l, q0_r);
TRANSPOSE16x8_UB_UB(t0, t1, t2, t3, q1_l, q1_r, q2_l, q2_r,
{ p0_l, p0_r, p1_l, p1_r, p2_l, p2_r, q0_l, q0_r,
v16u8 row0, row1, row2, row3, row4, row5, row6, row7; p2, p1, p0, q0, q1, q2, alpha, beta);
v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
alpha = (v16u8)__msa_fill_b(iAlpha);
LD_UB8(src, img_width, beta = (v16u8)__msa_fill_b(iBeta);
row0, row1, row2, row3, row4, row5, row6, row7);
src += (8 * img_width); bDetaP0Q0 = __msa_asub_u_b(p0, q0);
LD_UB8(src, img_width, bDetaP1P0 = __msa_asub_u_b(p1, p0);
row8, row9, row10, row11, row12, row13, row14, row15); bDetaQ1Q0 = __msa_asub_u_b(q1, q0);
bDetaP2P0 = __msa_asub_u_b(p2, p0);
TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, bDetaQ2Q0 = __msa_asub_u_b(q2, q0);
row8, row9, row10, row11, bDetaP0Q0 = (v16u8)__msa_clt_u_b(bDetaP0Q0, alpha);
row12, row13, row14, row15, bDetaP1P0 = (v16u8)__msa_clt_u_b(bDetaP1P0, beta);
p3_org, p2_org, p1_org, p0_org, bDetaQ1Q0 = (v16u8)__msa_clt_u_b(bDetaQ1Q0, beta);
q0_org, q1_org, q2_org, q3_org); bDetaP2P0 = (v16u8)__msa_clt_u_b(bDetaP2P0, beta);
} bDetaQ2Q0 = (v16u8)__msa_clt_u_b(bDetaQ2Q0, beta);
p0_asub_q0 = __msa_asub_u_b(p0_org, q0_org); // Unsigned extend p0, p1, p2, q0, q1, q2 from 8 bits to 16 bits
p1_asub_p0 = __msa_asub_u_b(p1_org, p0_org); ILVRL_B2_SH(zero, p0, p0_r, p0_l);
q1_asub_q0 = __msa_asub_u_b(q1_org, q0_org); ILVRL_B2_SH(zero, p1, p1_r, p1_l);
ILVRL_B2_SH(zero, p2, p2_r, p2_l);
alpha = (v16u8) __msa_fill_b(alpha_in); ILVRL_B2_SH(zero, q0, q0_r, q0_l);
beta = (v16u8) __msa_fill_b(beta_in); ILVRL_B2_SH(zero, q1, q1_r, q1_l);
ILVRL_B2_SH(zero, q2, q2_r, q2_l);
is_less_than_alpha = (p0_asub_q0 < alpha); // Signed extend tc, negTc from 8 bits to 16 bits
is_less_than_beta = (p1_asub_p0 < beta); flags = __msa_clt_s_b(tc, zero);
is_less_than = is_less_than_beta & is_less_than_alpha; ILVRL_B2(v8i16, flags, tc, tc_r, tc_l);
is_less_than_beta = (q1_asub_q0 < beta); flags = __msa_clt_s_b(negTc, zero);
is_less_than = is_less_than_beta & is_less_than; ILVRL_B2(v8i16, flags, negTc, negTc_r, negTc_l);
is_less_than = is_less_than & is_bs_greater_than0;
f = (v16i8)bDetaP0Q0 & (v16i8)bDetaP1P0 & (v16i8)bDetaQ1Q0;
if (!__msa_test_bz_v(is_less_than)) { flags = f & (v16i8)bDetaP2P0;
v16i8 negate_tc, sign_negate_tc; flags = __msa_ceq_b(flags, zero);
v16u8 p0, q0, p2_asub_p0, q2_asub_q0; iTc += ((~flags) & const_1_b);
v8i16 tc_r, tc_l, negate_tc_r, i16_negatetc_l; flags = f & (v16i8)bDetaQ2Q0;
v8i16 p1_org_r, p0_org_r, q0_org_r, q1_org_r; flags = __msa_ceq_b(flags, zero);
v8i16 p1_org_l, p0_org_l, q0_org_l, q1_org_l; iTc += ((~flags) & const_1_b);
v8i16 p0_r, q0_r, p0_l, q0_l; negiTc = zero - iTc;
// Signed extend iTc, negiTc from 8 bits to 16 bits
negate_tc = zero - (v16i8) tc; flags = __msa_clt_s_b(iTc, zero);
sign_negate_tc = __msa_clti_s_b(negate_tc, 0); ILVRL_B2(v8i16, flags, iTc, iTc_r, iTc_l);
flags = __msa_clt_s_b(negiTc, zero);
ILVRL_B2_SH(sign_negate_tc, negate_tc, negate_tc_r, i16_negatetc_l); ILVRL_B2(v8i16, flags, negiTc, negiTc_r, negiTc_l);
UNPCK_UB_SH(tc, tc_r, tc_l); // Calculate the left part
UNPCK_UB_SH(p1_org, p1_org_r, p1_org_l); // p1
UNPCK_UB_SH(p0_org, p0_org_r, p0_org_l); t0 = (p2_l + ((p0_l + q0_l + const_1_h) >> 1) - (p1_l << 1)) >> 1;
UNPCK_UB_SH(q0_org, q0_org_r, q0_org_l); t0 = __msa_max_s_h(negTc_l, t0);
t0 = __msa_min_s_h(tc_l, t0);
p2_asub_p0 = __msa_asub_u_b(p2_org, p0_org); t1 = p1_l + t0;
is_less_than_beta = (p2_asub_p0 < beta); // q1
is_less_than_beta = is_less_than_beta & is_less_than; t0 = (q2_l + ((p0_l + q0_l + const_1_h) >> 1) - (q1_l << 1)) >> 1;
t0 = __msa_max_s_h(negTc_l, t0);
if (!__msa_test_bz_v(is_less_than_beta)) { t0 = __msa_min_s_h(tc_l, t0);
v16u8 p1; t2 = q1_l + t0;
v8i16 p1_r = { 0 }; // iDeta
v8i16 p1_l = { 0 }; t0 = (((q0_l - p0_l) << 2) + (p1_l - q1_l) + const_4_h) >> 3;
v8i16 p2_org_r = (v8i16) __msa_ilvr_b(zero, (v16i8) p2_org); t0 = __msa_max_s_h(negiTc_l, t0);
v8i16 p2_org_l = (v8i16) __msa_ilvl_b(zero, (v16i8) p2_org); t0 = __msa_min_s_h(iTc_l, t0);
p1_l = t1;
AVC_LPF_P1_OR_Q1(p0_org_r, q0_org_r, p1_org_r, p2_org_r, q1_l = t2;
negate_tc_r, tc_r, p1_r); // p0
AVC_LPF_P1_OR_Q1(p0_org_l, q0_org_l, p1_org_l, p2_org_l, t1 = p0_l + t0;
i16_negatetc_l, tc_l, p1_l); t2 = t1 & const_not_255_h;
t3 = __msa_cle_s_h((v8i16)zero, t1);
p1 = (v16u8) __msa_pckev_b((v16i8) p1_l, (v16i8) p1_r); flags = (v16i8)__msa_ceq_h(t2, (v8i16)zero);
p1_org = __msa_bmnz_v(p1_org, p1, is_less_than_beta); p0_l = (t1 & (v8i16)flags) + (t3 & (v8i16)(~flags));
// q0
is_less_than_beta = __msa_andi_b(is_less_than_beta, 1); t1 = q0_l - t0;
tc = tc + is_less_than_beta; t2 = t1 & const_not_255_h;
} t3 = __msa_cle_s_h((v8i16)zero, t1);
flags = (v16i8)__msa_ceq_h(t2, (v8i16)zero);
q2_asub_q0 = __msa_asub_u_b(q2_org, q0_org); q0_l = (t1 & (v8i16)flags) + (t3 & (v8i16)(~flags));
is_less_than_beta = (q2_asub_q0 < beta);
is_less_than_beta = is_less_than_beta & is_less_than; // Calculate the right part
// p1
q1_org_r = (v8i16) __msa_ilvr_b(zero, (v16i8) q1_org); t0 = (p2_r + ((p0_r + q0_r + const_1_h) >> 1) - (p1_r << 1)) >> 1;
q1_org_l = (v8i16) __msa_ilvl_b(zero, (v16i8) q1_org); t0 = __msa_max_s_h(negTc_r, t0);
t0 = __msa_min_s_h(tc_r, t0);
if (!__msa_test_bz_v(is_less_than_beta)) { t1 = p1_r + t0;
v16u8 q1; // q1
v8i16 q1_r = { 0 }; t0 = (q2_r + ((p0_r + q0_r + const_1_h) >> 1) - (q1_r << 1)) >> 1;
v8i16 q1_l = { 0 }; t0 = __msa_max_s_h(negTc_r, t0);
v8i16 q2_org_r = (v8i16) __msa_ilvr_b(zero, (v16i8) q2_org); t0 = __msa_min_s_h(tc_r, t0);
v8i16 q2_org_l = (v8i16) __msa_ilvl_b(zero, (v16i8) q2_org); t2 = q1_r + t0;
// iDeta
AVC_LPF_P1_OR_Q1(p0_org_r, q0_org_r, q1_org_r, q2_org_r, t0 = (((q0_r - p0_r) << 2) + (p1_r - q1_r) + const_4_h) >> 3;
negate_tc_r, tc_r, q1_r); t0 = __msa_max_s_h(negiTc_r, t0);
AVC_LPF_P1_OR_Q1(p0_org_l, q0_org_l, q1_org_l, q2_org_l, t0 = __msa_min_s_h(iTc_r, t0);
i16_negatetc_l, tc_l, q1_l); p1_r = t1;
q1_r = t2;
q1 = (v16u8) __msa_pckev_b((v16i8) q1_l, (v16i8) q1_r); // p0
q1_org = __msa_bmnz_v(q1_org, q1, is_less_than_beta); t1 = p0_r + t0;
t2 = t1 & const_not_255_h;
is_less_than_beta = __msa_andi_b(is_less_than_beta, 1); t3 = __msa_cle_s_h((v8i16)zero, t1);
tc = tc + is_less_than_beta; flags = (v16i8)__msa_ceq_h(t2, (v8i16)zero);
} p0_r = (t1 & (v8i16)flags) + (t3 & (v8i16)(~flags));
// q0
{ t1 = q0_r - t0;
v8i16 threshold_r, negate_thresh_r; t2 = t1 & const_not_255_h;
v8i16 threshold_l, negate_thresh_l; t3 = __msa_cle_s_h((v8i16)zero, t1);
v16i8 negate_thresh, sign_negate_thresh; flags = (v16i8)__msa_ceq_h(t2, (v8i16)zero);
q0_r = (t1 & (v8i16)flags) + (t3 & (v8i16)(~flags));
negate_thresh = zero - (v16i8) tc;
sign_negate_thresh = __msa_clti_s_b(negate_thresh, 0); // Combined left and right
PCKEV_B4(v8i16, p1_l, p1_r, p0_l, p0_r, q0_l, q0_r, q1_l, q1_r,
ILVR_B2_SH(zero, tc, sign_negate_thresh, negate_thresh, t0, t1, t2, t3);
threshold_r, negate_thresh_r); flags = (v16i8)__msa_cle_s_b(zero, tc);
flags &= f;
AVC_LPF_P0Q0(q0_org_r, p0_org_r, p1_org_r, q1_org_r, p0 = (v16u8)(((v16i8)t1 & flags) + (p0 & (~flags)));
negate_thresh_r, threshold_r, p0_r, q0_r); q0 = (v16u8)(((v16i8)t2 & flags) + (q0 & (~flags)));
// Using t1, t2 as temporary flags
threshold_l = (v8i16) __msa_ilvl_b(zero, (v16i8) tc); t1 = (v8i16)(flags & (~(__msa_ceq_b((v16i8)bDetaP2P0, zero))));
negate_thresh_l = (v8i16) __msa_ilvl_b(sign_negate_thresh, p1 = (v16u8)(t0 & t1) + (p1 & (v16u8)(~t1));
negate_thresh); t2 = (v8i16)(flags & (~(__msa_ceq_b((v16i8)bDetaQ2Q0, zero))));
q1 = (v16u8)(t3 & t2) + (q1 & (v16u8)(~t2));
AVC_LPF_P0Q0(q0_org_l, p0_org_l, p1_org_l, q1_org_l,
negate_thresh_l, threshold_l, p0_l, q0_l); ILVRL_B2_SH(p0, p1, t0, t1);
} ILVRL_B2_SH(q1, q0, t2, t3);
ILVRL_H2_UB(t2, t0, p1, p0);
PCKEV_B2_UB(p0_l, p0_r, q0_l, q0_r, p0, q0); ILVRL_H2_UB(t3, t1, q0, q1);
// Store data to pPix
p0_org = __msa_bmnz_v(p0_org, p0, is_less_than); ST_W8(p1, p0, 0, 1, 2, 3, 0, 1, 2, 3, pPix - 2, iStride);
q0_org = __msa_bmnz_v(q0_org, q0, is_less_than); ST_W8(q0, q1, 0, 1, 2, 3, 0, 1, 2, 3, pPix + 8 * iStride - 2, iStride);
{
v16i8 tp0, tp1, tp2, tp3;
v8i16 tmp2, tmp5;
v4i32 tmp3, tmp4, tmp6, tmp7;
uint32_t out0, out2;
uint16_t out1, out3;
src = data - 3;
ILVRL_B2_SB(p1_org, p2_org, tp0, tp2);
ILVRL_B2_SB(q0_org, p0_org, tp1, tp3);
ILVRL_B2_SH(q2_org, q1_org, tmp2, tmp5);
ILVRL_H2_SW(tp1, tp0, tmp3, tmp4);
ILVRL_H2_SW(tp3, tp2, tmp6, tmp7);
out0 = __msa_copy_u_w(tmp3, 0);
out1 = __msa_copy_u_h(tmp2, 0);
out2 = __msa_copy_u_w(tmp3, 1);
out3 = __msa_copy_u_h(tmp2, 1);
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp3, 2);
out1 = __msa_copy_u_h(tmp2, 2);
out2 = __msa_copy_u_w(tmp3, 3);
out3 = __msa_copy_u_h(tmp2, 3);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp4, 0);
out1 = __msa_copy_u_h(tmp2, 4);
out2 = __msa_copy_u_w(tmp4, 1);
out3 = __msa_copy_u_h(tmp2, 5);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp4, 2);
out1 = __msa_copy_u_h(tmp2, 6);
out2 = __msa_copy_u_w(tmp4, 3);
out3 = __msa_copy_u_h(tmp2, 7);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp6, 0);
out1 = __msa_copy_u_h(tmp5, 0);
out2 = __msa_copy_u_w(tmp6, 1);
out3 = __msa_copy_u_h(tmp5, 1);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp6, 2);
out1 = __msa_copy_u_h(tmp5, 2);
out2 = __msa_copy_u_w(tmp6, 3);
out3 = __msa_copy_u_h(tmp5, 3);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp7, 0);
out1 = __msa_copy_u_h(tmp5, 4);
out2 = __msa_copy_u_w(tmp7, 1);
out3 = __msa_copy_u_h(tmp5, 5);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
out0 = __msa_copy_u_w(tmp7, 2);
out1 = __msa_copy_u_h(tmp5, 6);
out2 = __msa_copy_u_w(tmp7, 3);
out3 = __msa_copy_u_h(tmp5, 7);
src += img_width;
SW(out0, src);
SH(out1, (src + 4));
src += img_width;
SW(out2, src);
SH(out3, (src + 4));
}
}
}
} }
static void avc_loopfilter_luma_inter_edge_hor_msa(uint8_t *data, static void avc_loopfilter_luma_inter_edge_hor_msa(uint8_t *data,
@ -2180,23 +2056,24 @@ static void avc_h_loop_filter_chroma422_mbaff_msa(uint8_t *src,
void ff_h264_h_lpf_luma_inter_msa(uint8_t *data, ptrdiff_t img_width, void ff_h264_h_lpf_luma_inter_msa(uint8_t *data, ptrdiff_t img_width,
int alpha, int beta, int8_t *tc) int alpha, int beta, int8_t *tc)
{ {
uint8_t bs0 = 1; // uint8_t bs0 = 1;
uint8_t bs1 = 1; // uint8_t bs1 = 1;
uint8_t bs2 = 1; // uint8_t bs2 = 1;
uint8_t bs3 = 1; // uint8_t bs3 = 1;
//
if (tc[0] < 0) // if (tc[0] < 0)
bs0 = 0; // bs0 = 0;
if (tc[1] < 0) // if (tc[1] < 0)
bs1 = 0; // bs1 = 0;
if (tc[2] < 0) // if (tc[2] < 0)
bs2 = 0; // bs2 = 0;
if (tc[3] < 0) // if (tc[3] < 0)
bs3 = 0; // bs3 = 0;
//
avc_loopfilter_luma_inter_edge_ver_msa(data, bs0, bs1, bs2, bs3, // avc_loopfilter_luma_inter_edge_ver_msa(data, bs0, bs1, bs2, bs3,
tc[0], tc[1], tc[2], tc[3], // tc[0], tc[1], tc[2], tc[3],
alpha, beta, img_width); // alpha, beta, img_width);
avc_loopfilter_luma_inter_edge_ver_msa(data, img_width, alpha, beta, tc);
} }
void ff_h264_v_lpf_luma_inter_msa(uint8_t *data, ptrdiff_t img_width, void ff_h264_v_lpf_luma_inter_msa(uint8_t *data, ptrdiff_t img_width,

Write Preview
Loading…
Cancel
Save