From 3c2f7eb3398ea5e6a6d072e75922a317b73c7e8d Mon Sep 17 00:00:00 2001 From: Kaustubh Raste Date: Fri, 7 Oct 2016 18:43:31 +0530 Subject: [PATCH] Optimize vp9 loopfilter msa functions Updated code to process in 8bit as saturation/clipping takes care of overflow Removed unused macro Change-Id: I113df60286fb28b216df800d95b2d3695ef71440 --- vpx_dsp/mips/loopfilter_16_msa.c | 4 +- vpx_dsp/mips/loopfilter_4_msa.c | 4 +- vpx_dsp/mips/loopfilter_8_msa.c | 4 +- vpx_dsp/mips/loopfilter_msa.h | 141 ++++++++----------------------- 4 files changed, 40 insertions(+), 113 deletions(-) diff --git a/vpx_dsp/mips/loopfilter_16_msa.c b/vpx_dsp/mips/loopfilter_16_msa.c index 6ee2456ca..b73d56bd5 100644 --- a/vpx_dsp/mips/loopfilter_16_msa.c +++ b/vpx_dsp/mips/loopfilter_16_msa.c @@ -449,7 +449,7 @@ static void mb_lpf_horizontal_edge(uint8_t *src, int32_t pitch, LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, mask, flat); VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); @@ -779,7 +779,7 @@ int32_t vpx_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48, /* flat4 */ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); /* filter4 */ - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); diff --git a/vpx_dsp/mips/loopfilter_4_msa.c b/vpx_dsp/mips/loopfilter_4_msa.c index e0079665f..9500cd2fd 100644 --- a/vpx_dsp/mips/loopfilter_4_msa.c +++ b/vpx_dsp/mips/loopfilter_4_msa.c @@ -27,7 +27,7 @@ void vpx_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch, LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, mask, flat); - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); p1_d = __msa_copy_u_d((v2i64)p1_out, 0); p0_d = __msa_copy_u_d((v2i64)p0_out, 0); @@ -86,7 +86,7 @@ void vpx_lpf_vertical_4_msa(uint8_t *src, int32_t pitch, q3); LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, mask, flat); - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1); ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1); ILVRL_H2_SH(vec1, vec0, vec2, vec3); diff --git a/vpx_dsp/mips/loopfilter_8_msa.c b/vpx_dsp/mips/loopfilter_8_msa.c index 403e5dc51..a22c62bb3 100644 --- a/vpx_dsp/mips/loopfilter_8_msa.c +++ b/vpx_dsp/mips/loopfilter_8_msa.c @@ -32,7 +32,7 @@ void vpx_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch, LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev, mask, flat); VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); @@ -177,7 +177,7 @@ void vpx_lpf_vertical_8_msa(uint8_t *src, int32_t pitch, /* flat4 */ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat); /* filter4 */ - VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); + VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out); flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat); diff --git a/vpx_dsp/mips/loopfilter_msa.h b/vpx_dsp/mips/loopfilter_msa.h index 4063e5e6b..fe6574eb6 100644 --- a/vpx_dsp/mips/loopfilter_msa.h +++ b/vpx_dsp/mips/loopfilter_msa.h @@ -13,113 +13,40 @@ #include "vpx_dsp/mips/macros_msa.h" -#define VP9_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \ - p1_out, p0_out, q0_out, q1_out) \ - { \ - v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \ - v16i8 filt, filt1, filt2, cnst4b, cnst3b; \ - v8i16 q0_sub_p0_r, filt_r, cnst3h; \ - \ - p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \ - p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \ - q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \ - q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \ - \ - filt = __msa_subs_s_b(p1_m, q1_m); \ - filt = filt & (v16i8)hev_in; \ - q0_sub_p0 = q0_m - p0_m; \ - filt_sign = __msa_clti_s_b(filt, 0); \ - \ - cnst3h = __msa_ldi_h(3); \ - q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \ - q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \ - filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \ - filt_r += q0_sub_p0_r; \ - filt_r = __msa_sat_s_h(filt_r, 7); \ - \ - /* combine left and right part */ \ - filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r); \ - \ - filt = filt & (v16i8)mask_in; \ - cnst4b = __msa_ldi_b(4); \ - filt1 = __msa_adds_s_b(filt, cnst4b); \ - filt1 >>= 3; \ - \ - cnst3b = __msa_ldi_b(3); \ - filt2 = __msa_adds_s_b(filt, cnst3b); \ - filt2 >>= 3; \ - \ - q0_m = __msa_subs_s_b(q0_m, filt1); \ - q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \ - p0_m = __msa_adds_s_b(p0_m, filt2); \ - p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \ - \ - filt = __msa_srari_b(filt1, 1); \ - hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \ - filt = filt & (v16i8)hev_in; \ - \ - q1_m = __msa_subs_s_b(q1_m, filt); \ - q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \ - p1_m = __msa_adds_s_b(p1_m, filt); \ - p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \ - } - -#define VP9_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \ - p1_out, p0_out, q0_out, q1_out) \ - { \ - v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \ - v16i8 filt, filt1, filt2, cnst4b, cnst3b; \ - v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h; \ - \ - p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \ - p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \ - q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \ - q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \ - \ - filt = __msa_subs_s_b(p1_m, q1_m); \ - \ - filt = filt & (v16i8)hev_in; \ - \ - q0_sub_p0 = q0_m - p0_m; \ - filt_sign = __msa_clti_s_b(filt, 0); \ - \ - cnst3h = __msa_ldi_h(3); \ - q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \ - q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \ - filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \ - filt_r += q0_sub_p0_r; \ - filt_r = __msa_sat_s_h(filt_r, 7); \ - \ - q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0); \ - q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h); \ - filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt); \ - filt_l += q0_sub_p0_l; \ - filt_l = __msa_sat_s_h(filt_l, 7); \ - \ - filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r); \ - filt = filt & (v16i8)mask_in; \ - \ - cnst4b = __msa_ldi_b(4); \ - filt1 = __msa_adds_s_b(filt, cnst4b); \ - filt1 >>= 3; \ - \ - cnst3b = __msa_ldi_b(3); \ - filt2 = __msa_adds_s_b(filt, cnst3b); \ - filt2 >>= 3; \ - \ - q0_m = __msa_subs_s_b(q0_m, filt1); \ - q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \ - p0_m = __msa_adds_s_b(p0_m, filt2); \ - p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \ - \ - filt = __msa_srari_b(filt1, 1); \ - hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \ - filt = filt & (v16i8)hev_in; \ - \ - q1_m = __msa_subs_s_b(q1_m, filt); \ - q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \ - p1_m = __msa_adds_s_b(p1_m, filt); \ - p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \ +#define VP9_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask, hev, p1_out, \ + p0_out, q0_out, q1_out) \ + { \ + v16i8 p1_m, p0_m, q0_m, q1_m, filt, q0_sub_p0, t1, t2; \ + const v16i8 cnst4b = __msa_ldi_b(4); \ + const v16i8 cnst3b = __msa_ldi_b(3); \ + \ + p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \ + p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \ + q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \ + q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \ + \ + filt = __msa_subs_s_b(p1_m, q1_m); \ + filt &= hev; \ + q0_sub_p0 = __msa_subs_s_b(q0_m, p0_m); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt = __msa_adds_s_b(filt, q0_sub_p0); \ + filt &= mask; \ + t1 = __msa_adds_s_b(filt, cnst4b); \ + t1 >>= cnst3b; \ + t2 = __msa_adds_s_b(filt, cnst3b); \ + t2 >>= cnst3b; \ + q0_m = __msa_subs_s_b(q0_m, t1); \ + q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \ + p0_m = __msa_adds_s_b(p0_m, t2); \ + p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \ + filt = __msa_srari_b(t1, 1); \ + hev = __msa_xori_b(hev, 0xff); \ + filt &= hev; \ + q1_m = __msa_subs_s_b(q1_m, filt); \ + q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \ + p1_m = __msa_adds_s_b(p1_m, filt); \ + p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \ } #define VP9_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out) \ -- 2.40.0