Lines Matching refs:PARALLEL_COLS_53
75 #define PARALLEL_COLS_53 (2*VREG_INT_COUNT) macro
640 LOAD(&tmp[PARALLEL_COLS_53 * i + 0])); in opj_idwt53_v_final_memcpy()
642 LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT])); in opj_idwt53_v_final_memcpy()
666 assert(PARALLEL_COLS_53 == 16); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
669 assert(PARALLEL_COLS_53 == 8); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
703 STORE(tmp + PARALLEL_COLS_53 * (i + 0), s0c_0); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
704 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0c_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
707 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
709 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
713 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + 0, s0n_0); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
714 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0n_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
721 STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
723 STORE(tmp + PARALLEL_COLS_53 * (len - 2), in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
729 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
732 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
737 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
739 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
768 assert(PARALLEL_COLS_53 == 16); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
771 assert(PARALLEL_COLS_53 == 8); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
784 STORE(tmp + PARALLEL_COLS_53 * 0, ADD(LOADU(in_even + 0), dc_0)); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
790 STORE(tmp + PARALLEL_COLS_53 * 0 + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
804 STORE(tmp + PARALLEL_COLS_53 * i, dc_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
805 STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
808 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
810 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
818 STORE(tmp + PARALLEL_COLS_53 * i, dc_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
819 STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
829 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + 0, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
831 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
834 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, dn_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
835 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, dn_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
837 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, ADD(s1_0, dc_0)); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
838 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
982 if (len > 1 && nb_cols == PARALLEL_COLS_53) { in opj_idwt53_v()
1025 if (len > 2 && nb_cols == PARALLEL_COLS_53) { in opj_idwt53_v()
1372 for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j; in opj_dwt_decode_v_func()
1373 j += PARALLEL_COLS_53) { in opj_dwt_decode_v_func()
1375 PARALLEL_COLS_53); in opj_dwt_decode_v_func()
1414 if (h.mem_count > (SIZE_MAX / PARALLEL_COLS_53 / sizeof(OPJ_INT32))) { in opj_dwt_decode_tile()
1421 h_mem_size = h.mem_count * PARALLEL_COLS_53 * sizeof(OPJ_INT32); in opj_dwt_decode_tile()
1497 for (j = 0; j + PARALLEL_COLS_53 <= rw; in opj_dwt_decode_tile()
1498 j += PARALLEL_COLS_53) { in opj_dwt_decode_tile()
1499 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53); in opj_dwt_decode_tile()