Lines Matching refs:coeffs
308 static void hevc_idct_4x4_msa(int16_t *coeffs)
315 LD_SH2(coeffs, 8, in0, in1);
328 ST_SH2(in0, in1, coeffs, 8);
331 static void hevc_idct_8x8_msa(int16_t *coeffs)
336 LD_SH8(coeffs, 8, in0, in1, in2, in3, in4, in5, in6, in7);
343 ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, coeffs, 8);
346 static void hevc_idct_16x16_msa(int16_t *coeffs)
351 int16_t *src = coeffs;
381 buf_ptr = coeffs;
404 buf_ptr = coeffs + 8;
408 LD_SH8(coeffs, 16, in0, in1, in2, in3, in4, in5, in6, in7);
411 ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, coeffs, 16);
413 LD_SH8((coeffs + 8), 16, in0, in1, in2, in3, in4, in5, in6, in7);
416 LD_SH8((coeffs + 128), 16, in8, in9, in10, in11, in12, in13, in14, in15);
417 ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, (coeffs + 128), 16);
420 ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, (coeffs + 8), 16);
422 LD_SH8((coeffs + 136), 16, in0, in1, in2, in3, in4, in5, in6, in7);
425 ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, (coeffs + 136), 16);
428 static void hevc_idct_8x32_column_msa(int16_t *coeffs, uint8_t buf_pitch,
436 int16_t *src0 = (coeffs + buf_pitch);
437 int16_t *src1 = (coeffs + 2 * buf_pitch);
438 int16_t *src2 = (coeffs + 4 * buf_pitch);
439 int16_t *src3 = (coeffs);
589 ST_SH(in0, (coeffs + i * buf_pitch));
595 ST_SH(in0, (coeffs + (31 - i) * buf_pitch));
601 static void hevc_idct_transpose_32x8_to_8x32(int16_t *coeffs, int16_t *tmp_buf)
607 LD_SH8(coeffs + i * 8, 32, in0, in1, in2, in3, in4, in5, in6, in7);
614 static void hevc_idct_transpose_8x32_to_32x8(int16_t *tmp_buf, int16_t *coeffs)
623 ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, coeffs + i * 8, 32);
627 static void hevc_idct_32x32_msa(int16_t *coeffs)
630 int16_t *src = coeffs;
644 hevc_idct_8x32_column_msa((coeffs + col_cnt * 8), buf_pitch, round);
652 src = (coeffs + 32 * 8 * row_cnt);
660 static void hevc_idct_dc_4x4_msa(int16_t *coeffs)
665 val = (coeffs[0] + 1) >> 1;
669 ST_SH2(dst, dst, coeffs, 8);
672 static void hevc_idct_dc_8x8_msa(int16_t *coeffs)
677 val = (coeffs[0] + 1) >> 1;
681 ST_SH8(dst, dst, dst, dst, dst, dst, dst, dst, coeffs, 8);
684 static void hevc_idct_dc_16x16_msa(int16_t *coeffs)
690 val = (coeffs[0] + 1) >> 1;
695 ST_SH8(dst, dst, dst, dst, dst, dst, dst, dst, coeffs, 8);
696 coeffs += 8 * 8;
700 static void hevc_idct_dc_32x32_msa(int16_t *coeffs)
706 val = (coeffs[0] + 1) >> 1;
711 ST_SH8(dst, dst, dst, dst, dst, dst, dst, dst, coeffs, 8);
712 coeffs += 8 * 8;
716 static void hevc_addblk_4x4_msa(int16_t *coeffs, uint8_t *dst, int32_t stride)
723 LD_SH2(coeffs, 8, in0, in1);
733 static void hevc_addblk_8x8_msa(int16_t *coeffs, uint8_t *dst, int32_t stride)
743 LD_SH8(coeffs, 8, in0, in1, in2, in3, in4, in5, in6, in7);
769 static void hevc_addblk_16x16_msa(int16_t *coeffs, uint8_t *dst, int32_t stride)
780 LD_SH4(coeffs, 16, in0, in2, in4, in6);
781 LD_SH4((coeffs + 8), 16, in1, in3, in5, in7);
782 coeffs += 64;
802 LD_SH4(coeffs, 16, in0, in2, in4, in6);
803 LD_SH4((coeffs + 8), 16, in1, in3, in5, in7);
804 coeffs += 64;
836 static void hevc_addblk_32x32_msa(int16_t *coeffs, uint8_t *dst, int32_t stride)
849 LD_SH4(coeffs, 16, in0, in2, in4, in6);
850 LD_SH4((coeffs + 8), 16, in1, in3, in5, in7);
851 coeffs += 64;
873 LD_SH4(coeffs, 16, in0, in2, in4, in6);
874 LD_SH4((coeffs + 8), 16, in1, in3, in5, in7);
875 coeffs += 64;
906 LD_SH4(coeffs, 16, in0, in2, in4, in6);
907 LD_SH4((coeffs + 8), 16, in1, in3, in5, in7);
941 static void hevc_idct_luma_4x4_msa(int16_t *coeffs)
946 LD_SH2(coeffs, 8, in0, in1);
960 ST_SH2(dst0, dst1, coeffs, 8);
963 void ff_hevc_idct_4x4_msa(int16_t *coeffs, int col_limit)
965 hevc_idct_4x4_msa(coeffs);
968 void ff_hevc_idct_8x8_msa(int16_t *coeffs, int col_limit)
970 hevc_idct_8x8_msa(coeffs);
973 void ff_hevc_idct_16x16_msa(int16_t *coeffs, int col_limit)
975 hevc_idct_16x16_msa(coeffs);
978 void ff_hevc_idct_32x32_msa(int16_t *coeffs, int col_limit)
980 hevc_idct_32x32_msa(coeffs);
983 void ff_hevc_addblk_4x4_msa(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
985 hevc_addblk_4x4_msa(coeffs, dst, stride);
988 void ff_hevc_addblk_8x8_msa(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
990 hevc_addblk_8x8_msa(coeffs, dst, stride);
993 void ff_hevc_addblk_16x16_msa(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
995 hevc_addblk_16x16_msa(coeffs, dst, stride);
998 void ff_hevc_addblk_32x32_msa(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride)
1000 hevc_addblk_32x32_msa(coeffs, dst, stride);
1003 void ff_hevc_idct_dc_4x4_msa(int16_t *coeffs)
1005 hevc_idct_dc_4x4_msa(coeffs);
1008 void ff_hevc_idct_dc_8x8_msa(int16_t *coeffs)
1010 hevc_idct_dc_8x8_msa(coeffs);
1013 void ff_hevc_idct_dc_16x16_msa(int16_t *coeffs)
1015 hevc_idct_dc_16x16_msa(coeffs);
1018 void ff_hevc_idct_dc_32x32_msa(int16_t *coeffs)
1020 hevc_idct_dc_32x32_msa(coeffs);
1023 void ff_hevc_idct_luma_4x4_msa(int16_t *coeffs)
1025 hevc_idct_luma_4x4_msa(coeffs);