32x8 helps more than 8x32, which is expected and in line with previous functions, as width-8 IDCTs are limited to YMM registers in the 2nd pass.
32x8
8x32
inv_txfm_add_8x32_dct_dct_0_10bpc_avx2: 36.5 inv_txfm_add_8x32_dct_dct_0_10bpc_avx512icl: 33.0 inv_txfm_add_8x32_dct_dct_1_10bpc_avx2: 163.2 inv_txfm_add_8x32_dct_dct_1_10bpc_avx512icl: 155.8 inv_txfm_add_8x32_dct_dct_2_10bpc_avx2: 199.5 inv_txfm_add_8x32_dct_dct_2_10bpc_avx512icl: 185.2 inv_txfm_add_8x32_dct_dct_4_10bpc_avx2: 285.5 inv_txfm_add_8x32_dct_dct_4_10bpc_avx512icl: 227.2 inv_txfm_add_32x8_dct_dct_0_10bpc_avx2: 21.9 inv_txfm_add_32x8_dct_dct_0_10bpc_avx512icl: 18.7 inv_txfm_add_32x8_dct_dct_1_10bpc_avx2: 348.5 inv_txfm_add_32x8_dct_dct_1_10bpc_avx512icl: 219.3 inv_txfm_add_32x8_dct_dct_2_10bpc_avx2: 348.5 inv_txfm_add_32x8_dct_dct_2_10bpc_avx512icl: 259.6 inv_txfm_add_32x8_dct_dct_4_10bpc_avx2: 348.5 inv_txfm_add_32x8_dct_dct_4_10bpc_avx512icl: 281.9 inv_txfm_add_8x32_identity_identity_2_10bpc_avx2: 28.4 inv_txfm_add_8x32_identity_identity_2_10bpc_avx512icl: 23.3 inv_txfm_add_8x32_identity_identity_4_10bpc_avx2: 53.2 inv_txfm_add_8x32_identity_identity_4_10bpc_avx512icl: 47.9 inv_txfm_add_32x8_identity_identity_2_10bpc_avx2: 30.4 inv_txfm_add_32x8_identity_identity_2_10bpc_avx512icl: 25.2 inv_txfm_add_32x8_identity_identity_4_10bpc_avx2: 51.9 inv_txfm_add_32x8_identity_identity_4_10bpc_avx512icl: 40.6
Henrik Gramner authored50babcfb
Henrik Gramner authored3136ae6a