Lines Matching refs:PARALLEL_COLS_53
75 #define PARALLEL_COLS_53 (2*VREG_INT_COUNT) macro
638 LOAD(&tmp[PARALLEL_COLS_53 * i + 0])); in opj_idwt53_v_final_memcpy()
640 LOAD(&tmp[PARALLEL_COLS_53 * i + VREG_INT_COUNT])); in opj_idwt53_v_final_memcpy()
664 assert(PARALLEL_COLS_53 == 16); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
667 assert(PARALLEL_COLS_53 == 8); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
701 STORE(tmp + PARALLEL_COLS_53 * (i + 0), s0c_0); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
702 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0c_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
705 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
707 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
711 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + 0, s0n_0); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
712 STORE(tmp + PARALLEL_COLS_53 * (i + 0) + VREG_INT_COUNT, s0n_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
719 STORE(tmp + PARALLEL_COLS_53 * (len - 1), tmp_len_minus_1); in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
721 STORE(tmp + PARALLEL_COLS_53 * (len - 2), in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
727 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
730 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
735 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
737 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas0_mcols_SSE2_OR_AVX2()
766 assert(PARALLEL_COLS_53 == 16); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
769 assert(PARALLEL_COLS_53 == 8); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
782 STORE(tmp + PARALLEL_COLS_53 * 0, ADD(LOADU(in_even + 0), dc_0)); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
788 STORE(tmp + PARALLEL_COLS_53 * 0 + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
802 STORE(tmp + PARALLEL_COLS_53 * i, dc_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
803 STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
806 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + 0, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
808 STORE(tmp + PARALLEL_COLS_53 * (i + 1) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
816 STORE(tmp + PARALLEL_COLS_53 * i, dc_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
817 STORE(tmp + PARALLEL_COLS_53 * i + VREG_INT_COUNT, dc_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
827 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + 0, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
829 STORE(tmp + PARALLEL_COLS_53 * (len - 2) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
832 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, dn_0); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
833 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, dn_1); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
835 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + 0, ADD(s1_0, dc_0)); in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
836 STORE(tmp + PARALLEL_COLS_53 * (len - 1) + VREG_INT_COUNT, in opj_idwt53_v_cas1_mcols_SSE2_OR_AVX2()
980 if (len > 1 && nb_cols == PARALLEL_COLS_53) { in opj_idwt53_v()
1023 if (len > 2 && nb_cols == PARALLEL_COLS_53) { in opj_idwt53_v()
1370 for (j = job->min_j; j + PARALLEL_COLS_53 <= job->max_j; in opj_dwt_decode_v_func()
1371 j += PARALLEL_COLS_53) { in opj_dwt_decode_v_func()
1373 PARALLEL_COLS_53); in opj_dwt_decode_v_func()
1412 if (h.mem_count > (SIZE_MAX / PARALLEL_COLS_53 / sizeof(OPJ_INT32))) { in opj_dwt_decode_tile()
1419 h_mem_size = h.mem_count * PARALLEL_COLS_53 * sizeof(OPJ_INT32); in opj_dwt_decode_tile()
1495 for (j = 0; j + PARALLEL_COLS_53 <= rw; in opj_dwt_decode_tile()
1496 j += PARALLEL_COLS_53) { in opj_dwt_decode_tile()
1497 opj_idwt53_v(&v, &tiledp[j], (OPJ_SIZE_T)w, PARALLEL_COLS_53); in opj_dwt_decode_tile()