diff options
| -rw-r--r-- | test/convolve_test.cc | 2 | ||||
| -rw-r--r-- | vp9/common/mips/msa/vp9_convolve8_avg_msa.c | 679 | ||||
| -rw-r--r-- | vp9/common/vp9_rtcd_defs.pl | 2 | ||||
| -rw-r--r-- | vp9/vp9_common.mk | 1 |
4 files changed, 682 insertions, 2 deletions
diff --git a/test/convolve_test.cc b/test/convolve_test.cc index 856507703..d5cb793f1 100644 --- a/test/convolve_test.cc +++ b/test/convolve_test.cc @@ -1820,7 +1820,7 @@ const ConvolveFunctions convolve8_msa( vp9_convolve_copy_msa, vp9_convolve_avg_msa, vp9_convolve8_horiz_msa, vp9_convolve8_avg_horiz_msa, vp9_convolve8_vert_msa, vp9_convolve8_avg_vert_msa, - vp9_convolve8_msa, vp9_convolve8_avg_c, 0); + vp9_convolve8_msa, vp9_convolve8_avg_msa, 0); INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest, ::testing::Values( make_tuple(4, 4, &convolve8_msa), diff --git a/vp9/common/mips/msa/vp9_convolve8_avg_msa.c b/vp9/common/mips/msa/vp9_convolve8_avg_msa.c new file mode 100644 index 000000000..e9f3a9dc3 --- /dev/null +++ b/vp9/common/mips/msa/vp9_convolve8_avg_msa.c @@ -0,0 +1,679 @@ +/* + * Copyright (c) 2015 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#include "./vp9_rtcd.h" +#include "vp9/common/mips/msa/vp9_convolve_msa.h" + +static void common_hv_8ht_8vt_and_aver_dst_4w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3, tmp0, tmp1; + v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3; + v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; + v8i16 hz_out7, hz_out8, hz_out9, res0, res1, vec0, vec1, vec2, vec3, vec4; + v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3; + + mask0 = LD_UB(&mc_filt_mask_arr[16]); + src -= (3 + 3 * src_stride); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); + XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); + src += (7 * src_stride); + + hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8); + + filt = LD_SH(filter_vert); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3); + + ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src7, src8, src9, src10); + XORI_B4_128_SB(src7, src8, src9, src10); + src += (4 * src_stride); + + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8); + vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6); + res0 = FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + + hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8); + vec4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8); + res1 = FILT_8TAP_DPADD_S_H(vec1, vec2, vec3, vec4, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2); + + SRARI_H2_SH(res0, res1, FILTER_BITS); + SAT_SH2_SH(res0, res1, 7); + PCKEV_B2_UB(res0, res0, res1, res1, tmp0, tmp1); + XORI_B2_128_UB(tmp0, tmp1); + AVER_UB2_UB(tmp0, dst0, tmp1, dst2, tmp0, tmp1); + ST4x4_UB(tmp0, tmp1, 0, 1, 0, 1, dst, dst_stride); + dst += (4 * dst_stride); + + hz_out5 = hz_out9; + vec0 = vec2; + vec1 = vec3; + vec2 = vec4; + } +} + +static void common_hv_8ht_8vt_and_aver_dst_8w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10; + v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3; + v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3; + v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3; + v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; + v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3; + v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9; + + mask0 = LD_UB(&mc_filt_mask_arr[0]); + src -= (3 + 3 * src_stride); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3); + + mask1 = mask0 + 2; + mask2 = mask0 + 4; + mask3 = mask0 + 6; + + LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6); + src += (7 * src_stride); + + XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6); + hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0, + filt_hz1, filt_hz2, filt_hz3); + + filt = LD_SH(filter_vert); + SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3); + + ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1); + ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4); + ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src7, src8, src9, src10); + XORI_B4_128_SB(src7, src8, src9, src10); + src += (4 * src_stride); + + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + + hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6); + tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + + hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7); + tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + + hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8); + tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + + hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3, + filt_hz0, filt_hz1, filt_hz2, filt_hz3); + out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9); + tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1, + filt_vt2, filt_vt3); + + SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7); + CONVERT_UB_AVG_ST8x4_UB(tmp0, tmp1, tmp2, tmp3, dst0, dst1, dst2, dst3, + dst, dst_stride); + dst += (4 * dst_stride); + + hz_out6 = hz_out10; + out0 = out2; + out1 = out3; + out2 = out8; + out4 = out6; + out5 = out7; + out6 = out9; + } +} + +static void common_hv_8ht_8vt_and_aver_dst_16w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 2; multiple8_cnt--;) { + common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + src += 8; + dst += 8; + } +} + +static void common_hv_8ht_8vt_and_aver_dst_32w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 4; multiple8_cnt--;) { + common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + src += 8; + dst += 8; + } +} + +static void common_hv_8ht_8vt_and_aver_dst_64w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 8; multiple8_cnt--;) { + common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + src += 8; + dst += 8; + } +} + +static void common_hv_2ht_2vt_and_aver_dst_4x4_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + v16i8 src0, src1, src2, src3, src4, mask; + v16u8 filt_hz, filt_vt, vec0, vec1; + v16u8 dst0, dst1, dst2, dst3, res0, res1; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1, filt; + + mask = LD_SB(&mc_filt_mask_arr[16]); + + /* rearranging filter */ + filt = LD_UH(filter_horiz); + filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0); + + filt = LD_UH(filter_vert); + filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0); + + LD_SB5(src, src_stride, src0, src1, src2, src3, src4); + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8); + hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1); + AVER_UB2_UB(res0, dst0, res1, dst2, res0, res1); + ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); +} + +static void common_hv_2ht_2vt_and_aver_dst_4x8_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask; + v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3, res0, res1, res2, res3; + v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6; + v8u16 hz_out7, hz_out8, tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + mask = LD_SB(&mc_filt_mask_arr[16]); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + filt_hz = (v16u8)__msa_splati_h(filt, 0); + + filt = LD_SH(filter_vert); + filt_vt = (v16u8)__msa_splati_h(filt, 0); + + LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7); + src += (8 * src_stride); + src8 = LD_SB(src); + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS); + hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS); + hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS); + hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS); + SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1, + hz_out3, hz_out5, 8); + hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6); + + LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7); + ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, + dst4, dst6); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3); + DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt, + tmp0, tmp1, tmp2, tmp3); + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + SAT_UH4_UH(tmp0, tmp1, tmp2, tmp3, 7); + PCKEV_B4_UB(tmp0, tmp0, tmp1, tmp1, tmp2, tmp2, tmp3, tmp3, res0, res1, + res2, res3); + AVER_UB4_UB(res0, dst0, res1, dst2, res2, dst4, res3, dst6, res0, res1, + res2, res3); + ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride); + dst += (4 * dst_stride); + ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride); +} + +static void common_hv_2ht_2vt_and_aver_dst_4w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + if (4 == height) { + common_hv_2ht_2vt_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert); + } else if (8 == height) { + common_hv_2ht_2vt_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert); + } +} + +static void common_hv_2ht_2vt_and_aver_dst_8x4_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert) { + v16i8 src0, src1, src2, src3, src4, mask; + v16u8 filt_hz, filt_vt, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3; + v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + filt_hz = (v16u8)__msa_splati_h(filt, 0); + + filt = LD_SH(filter_vert); + filt_vt = (v16u8)__msa_splati_h(filt, 0); + + LD_SB5(src, src_stride, src0, src1, src2, src3, src4); + src += (5 * src_stride); + + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp0 = __msa_dotp_u_h(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp1 = __msa_dotp_u_h(vec1, filt_vt); + + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp2 = __msa_dotp_u_h(vec2, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp3 = __msa_dotp_u_h(vec3, filt_vt); + + SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS); + SAT_UH4_UH(tmp0, tmp1, tmp2, tmp3, 7); + PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, + dst, dst_stride); +} + +static void common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, mask; + v16u8 filt_hz, filt_vt, vec0, dst0, dst1, dst2, dst3; + v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3; + v8i16 filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + filt_hz = (v16u8)__msa_splati_h(filt, 0); + + filt = LD_SH(filter_vert); + filt_vt = (v16u8)__msa_splati_h(filt, 0); + + src0 = LD_SB(src); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src1, src2, src3, src4); + src += (4 * src_stride); + + hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp0 = __msa_dotp_u_h(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp1 = __msa_dotp_u_h(vec0, filt_vt); + + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + + hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0); + tmp2 = __msa_dotp_u_h(vec0, filt_vt); + + hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1); + tmp3 = __msa_dotp_u_h(vec0, filt_vt); + + SRARI_H2_UH(tmp2, tmp3, FILTER_BITS); + SAT_UH2_UH(tmp2, tmp3, 7); + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, + dst, dst_stride); + dst += (4 * dst_stride); + } +} + +static void common_hv_2ht_2vt_and_aver_dst_8w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + if (4 == height) { + common_hv_2ht_2vt_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert); + } else { + common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, + height); + } +} + +static void common_hv_2ht_2vt_and_aver_dst_16w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + uint32_t loop_cnt; + v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask; + v16u8 filt_hz, filt_vt, vec0, vec1, dst0, dst1, dst2, dst3; + v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1; + v8i16 filt; + + mask = LD_SB(&mc_filt_mask_arr[0]); + + /* rearranging filter */ + filt = LD_SH(filter_horiz); + filt_hz = (v16u8)__msa_splati_h(filt, 0); + + filt = LD_SH(filter_vert); + filt_vt = (v16u8)__msa_splati_h(filt, 0); + + LD_SB2(src, 8, src0, src1); + src += src_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + + for (loop_cnt = (height >> 2); loop_cnt--;) { + LD_SB4(src, src_stride, src0, src2, src4, src6); + LD_SB4(src + 8, src_stride, src1, src3, src5, src7); + src += (4 * src_stride); + LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3); + + hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst); + dst += dst_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + PCKEV_AVG_ST_UB(tmp1, tmp0, dst1, dst); + dst += dst_stride; + + hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS); + hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst); + dst += dst_stride; + + hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS); + hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS); + ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1); + DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1); + SRARI_H2_UH(tmp0, tmp1, FILTER_BITS); + SAT_UH2_UH(tmp0, tmp1, 7); + PCKEV_AVG_ST_UB(tmp1, tmp0, dst3, dst); + dst += dst_stride; + } +} + +static void common_hv_2ht_2vt_and_aver_dst_32w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 2; multiple8_cnt--;) { + common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + src += 16; + dst += 16; + } +} + +static void common_hv_2ht_2vt_and_aver_dst_64w_msa(const uint8_t *src, + int32_t src_stride, + uint8_t *dst, + int32_t dst_stride, + int8_t *filter_horiz, + int8_t *filter_vert, + int32_t height) { + int32_t multiple8_cnt; + for (multiple8_cnt = 4; multiple8_cnt--;) { + common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride, + filter_horiz, filter_vert, height); + src += 16; + dst += 16; + } +} + +void vp9_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h) { + int8_t cnt, filt_hor[8], filt_ver[8]; + + if (16 != x_step_q4 || 16 != y_step_q4) { + vp9_convolve8_avg_c(src, src_stride, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); + return; + } + + if (((const int32_t *)filter_x)[1] == 0x800000 && + ((const int32_t *)filter_y)[1] == 0x800000) { + vp9_convolve_avg(src, src_stride, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); + return; + } + + for (cnt = 0; cnt < 8; ++cnt) { + filt_hor[cnt] = filter_x[cnt]; + filt_ver[cnt] = filter_y[cnt]; + } + + if (((const int32_t *)filter_x)[0] == 0 && + ((const int32_t *)filter_y)[0] == 0) { + switch (w) { + case 4: + common_hv_2ht_2vt_and_aver_dst_4w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + &filt_hor[3], &filt_ver[3], h); + break; + case 8: + common_hv_2ht_2vt_and_aver_dst_8w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + &filt_hor[3], &filt_ver[3], h); + break; + case 16: + common_hv_2ht_2vt_and_aver_dst_16w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + &filt_hor[3], &filt_ver[3], h); + break; + case 32: + common_hv_2ht_2vt_and_aver_dst_32w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + &filt_hor[3], &filt_ver[3], h); + break; + case 64: + common_hv_2ht_2vt_and_aver_dst_64w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + &filt_hor[3], &filt_ver[3], h); + break; + default: + vp9_convolve8_avg_c(src, src_stride, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); + break; + } + } else if (((const int32_t *)filter_x)[0] == 0 || + ((const int32_t *)filter_y)[0] == 0) { + vp9_convolve8_avg_c(src, src_stride, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); + } else { + switch (w) { + case 4: + common_hv_8ht_8vt_and_aver_dst_4w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + filt_hor, filt_ver, h); + break; + case 8: + common_hv_8ht_8vt_and_aver_dst_8w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + filt_hor, filt_ver, h); + break; + case 16: + common_hv_8ht_8vt_and_aver_dst_16w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + filt_hor, filt_ver, h); + break; + case 32: + common_hv_8ht_8vt_and_aver_dst_32w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + filt_hor, filt_ver, h); + break; + case 64: + common_hv_8ht_8vt_and_aver_dst_64w_msa(src, (int32_t)src_stride, + dst, (int32_t)dst_stride, + filt_hor, filt_ver, h); + break; + default: + vp9_convolve8_avg_c(src, src_stride, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); + break; + } + } +} diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl index da62f09b3..65de9d20b 100644 --- a/vp9/common/vp9_rtcd_defs.pl +++ b/vp9/common/vp9_rtcd_defs.pl @@ -301,7 +301,7 @@ add_proto qw/void vp9_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride specialize qw/vp9_convolve8_vert sse2 ssse3 neon dspr2 msa/, "$avx2_ssse3"; add_proto qw/void vp9_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"; -specialize qw/vp9_convolve8_avg sse2 ssse3 neon dspr2/; +specialize qw/vp9_convolve8_avg sse2 ssse3 neon dspr2 msa/; add_proto qw/void vp9_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h"; specialize qw/vp9_convolve8_avg_horiz sse2 ssse3 neon dspr2 msa/; diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk index 878dff369..166926b3e 100644 --- a/vp9/vp9_common.mk +++ b/vp9/vp9_common.mk @@ -133,6 +133,7 @@ VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_mblpf_vert_loopfilter_ds # common (msa) VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_macros_msa.h VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_convolve8_avg_horiz_msa.c +VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_convolve8_avg_msa.c VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_convolve8_avg_vert_msa.c VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_convolve8_horiz_msa.c VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_convolve8_msa.c |