; ; SIMD for interpolation ; Copyright 2019 Anton Khirnov ; ; This program is free software: you can redistribute it and/or modify ; it under the terms of the GNU General Public License as published by ; the Free Software Foundation, either version 3 of the License, or ; (at your option) any later version. ; ; This program is distributed in the hope that it will be useful, ; but WITHOUT ANY WARRANTY; without even the implied warranty of ; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the ; GNU General Public License for more details. ; ; You should have received a copy of the GNU General Public License ; along with this program. If not, see . ;/ %include "config.asm" %include "util.asm" %include "x86inc.asm" SECTION .text ; double %define ELEM_SIZE 8 ; transpose a 4x4 matrix of doubles, supplied in %1-%4 ; %5, %6: temporary registers (clobbered) %macro TRANSPOSE4x4 6 unpcklpd %5, %1, %2 ; %5 = 00 10 02 12 unpckhpd %6, %1, %2 ; %6 = 01 11 03 13 unpcklpd %1, %3, %4 ; %1 = 20 30 22 32 unpckhpd %4, %3, %4 ; %4 = 21 31 23 33 vperm2f128 %3, %5, %1, 00110001b ; %3 = 02 12 22 32 vinsertf128 %1, %5, xmm%1, 1 ; %1 = 00 10 20 30 vinsertf128 %2, %6, xmm%4, 1 ; %2 = 01 11 21 31 vperm2f128 %4, %6, %4, 00110001b ; %4 = 03 13 23 33 %endmacro INIT_YMM avx2 cglobal transfer_interp_line_cont_4, 9, 11, 16, dst, dst_len, src, src_stride, idx_x, fact_x, fact_x_stride, fact_y, fact_y_stride,\ idx_x_val, fact_x3 shl fact_y_strideq, 3 shl fact_x_strideq, 3 shl src_strideq, 3 shl dst_lenq, 3 add dstq, dst_lenq add idx_xq, dst_lenq add fact_xq, dst_lenq neg dst_lenq ; from now on, the register that held the line size is used as the offset into data arrays %define offsetq dst_lenq ; load the y interpolation factors SPLATPD m0, [fact_yq + fact_y_strideq * 0], 0 ; fact y + 0 -> m0 SPLATPD m1, [fact_yq + fact_y_strideq * 1], 0 ; fact y + 1 -> m1 SPLATPD m2, [fact_yq + fact_y_strideq * 2], 0 ; fact y + 2 -> m2 add fact_yq, fact_y_strideq SPLATPD m3, [fact_yq + fact_y_strideq * 2], 0 ; fact y + 3 -> m3 ; reuse the now unneded fact_y[_stride] registers %define fact_x1q fact_yq %define fact_x2q fact_y_strideq lea fact_x1q, [fact_xq + fact_x_strideq] lea fact_x2q, [fact_xq + fact_x_strideq * 2] lea fact_x3q, [fact_x1q + fact_x_strideq * 2] .loop: ; load the x interpolation factors movu m4, [fact_xq + offsetq] ; fact x + 0 -> m4 movu m5, [fact_x1q + offsetq] ; fact x + 1 -> m5 movu m6, [fact_x2q + offsetq] ; fact x + 2 -> m6 movu m7, [fact_x3q + offsetq] ; fact x + 3 -> m7 ; load the x indices of the source values movu m15, [idx_xq + offsetq] psllq m15, m15, 3 movq idx_x_valq, xm15 movu m8, [srcq + idx_x_valq] vpermq m13, m15, 1 movq idx_x_valq, xm13 movu m9, [srcq + idx_x_valq] vpermq m13, m15, 2 movq idx_x_valq, xm13 movu m10, [srcq + idx_x_valq] vpermq m13, m15, 3 movq idx_x_valq, xm13 movu m11, [srcq + idx_x_valq] TRANSPOSE4x4 m8, m9, m10, m11, m12, m13 mulpd m14, m8, m4 vfmadd231pd m14, m9, m5 vfmadd231pd m14, m10, m6 vfmadd231pd m14, m11, m7 mulpd m14, m0 movq xm13, src_strideq vbroadcastsd m13, xm13 paddq m15, m13 movq idx_x_valq, xm15 movu m8, [srcq + idx_x_valq] vpermq m13, m15, 1 movq idx_x_valq, xm13 movu m9, [srcq + idx_x_valq] vpermq m13, m15, 2 movq idx_x_valq, xm13 movu m10, [srcq + idx_x_valq] vpermq m13, m15, 3 movq idx_x_valq, xm13 movu m11, [srcq + idx_x_valq] TRANSPOSE4x4 m8, m9, m10, m11, m12, m13 mulpd m13, m8, m4 vfmadd231pd m13, m9, m5 vfmadd231pd m13, m10, m6 vfmadd231pd m13, m11, m7 vfmadd231pd m14, m13, m1 movq xm13, src_strideq vbroadcastsd m13, xm13 paddq m15, m13 movq idx_x_valq, xm15 movu m8, [srcq + idx_x_valq] vpermq m13, m15, 1 movq idx_x_valq, xm13 movu m9, [srcq + idx_x_valq] vpermq m13, m15, 2 movq idx_x_valq, xm13 movu m10, [srcq + idx_x_valq] vpermq m13, m15, 3 movq idx_x_valq, xm13 movu m11, [srcq + idx_x_valq] TRANSPOSE4x4 m8, m9, m10, m11, m12, m13 mulpd m13, m8, m4 vfmadd231pd m13, m9, m5 vfmadd231pd m13, m10, m6 vfmadd231pd m13, m11, m7 vfmadd231pd m14, m13, m2 movq xm13, src_strideq vbroadcastsd m13, xm13 paddq m15, m13 movq idx_x_valq, xm15 movu m8, [srcq + idx_x_valq] vpermq m13, m15, 1 movq idx_x_valq, xm13 movu m9, [srcq + idx_x_valq] vpermq m13, m15, 2 movq idx_x_valq, xm13 movu m10, [srcq + idx_x_valq] vpermq m13, m15, 3 movq idx_x_valq, xm13 movu m11, [srcq + idx_x_valq] TRANSPOSE4x4 m8, m9, m10, m11, m12, m13 mulpd m13, m8, m4 vfmadd231pd m13, m9, m5 vfmadd231pd m13, m10, m6 vfmadd231pd m13, m11, m7 vfmadd231pd m14, m13, m3 add offsetq, mmsize jg .store_partial movu [dstq + offsetq - mmsize], m14 js .loop jmp .finish .store_partial: sub offsetq, ELEM_SIZE jz .store3 sub offsetq, ELEM_SIZE jz .store2 .store1: ; offsetq is now mmsize-2 after the write position movq [dstq + offsetq - mmsize + 2 * ELEM_SIZE], xm14 jmp .finish .store2: ; offsetq is now mmsize-2 after the write position movu [dstq + offsetq - mmsize + 2 * ELEM_SIZE], xm14 jmp .finish .store3: ; offsetq is now mmsize-1 after the write position movu [dstq + offsetq - mmsize + 1 * ELEM_SIZE], xm14 vextractf128 xm14, m14, 1 movq [dstq + offsetq - mmsize + 3 * ELEM_SIZE], xm14 .finish: RET