32 |
; * |
; * |
33 |
; * History: |
; * History: |
34 |
; * |
; * |
35 |
|
; * 09.08.2002 sse2 dequant funcs revamped |
36 |
; * 14.06.2002 mmx+xmm dequant_* funcs revamped -Skal- |
; * 14.06.2002 mmx+xmm dequant_* funcs revamped -Skal- |
37 |
; * 24.02.2002 sse2 quant_intra / dequant_intra (have to use movdqu ???) |
; * 24.02.2002 sse2 quant_intra / dequant_intra (have to use movdqu ???) |
38 |
; * 17.04.2002 sse2 quant_inter / dequant_inter |
; * 17.04.2002 sse2 quant_inter / dequant_inter |
262 |
; |
; |
263 |
;=========================================================================== |
;=========================================================================== |
264 |
|
|
265 |
align 8 |
align 16 |
266 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
sse2_2047 times 8 dw 2047 |
|
mmx_32767_minus_2047 times 4 dw (32767-2047) |
|
267 |
|
|
268 |
align 16 |
align 16 |
269 |
mmx_2047 times 4 dw 2047 |
mmx_2047 times 4 dw 2047 |
270 |
|
|
271 |
align 16 |
align 8 |
272 |
sse2_pos_2047 times 8 dw 2047 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
273 |
sse2_neg_2048 times 8 dw -2048 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
274 |
|
|
275 |
|
|
276 |
section .text |
section .text |
855 |
|
|
856 |
ret |
ret |
857 |
|
|
858 |
|
|
859 |
;=========================================================================== |
;=========================================================================== |
860 |
; |
; |
861 |
; void dequant_intra_sse2(int16_t *data, |
; void dequant_intra_sse2(int16_t *data, |
864 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
865 |
; |
; |
866 |
;=========================================================================== |
;=========================================================================== |
867 |
|
align ALIGN |
|
align 16 |
|
868 |
cglobal dequant_intra_sse2 |
cglobal dequant_intra_sse2 |
869 |
dequant_intra_sse2: |
dequant_intra_sse2: |
870 |
|
mov edx, [esp+ 4] ; data |
871 |
push esi |
mov ecx, [esp+ 8] ; coeff |
872 |
push edi |
mov eax, [esp+12] ; quant |
|
|
|
|
mov edi, [esp + 8 + 4] ; data |
|
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
|
|
|
873 |
movq mm6, [mmx_add + eax*8 - 8] |
movq mm6, [mmx_add + eax*8 - 8] |
874 |
movq mm7, [mmx_mul + eax*8 - 8] |
movq mm7, [mmx_mul + eax*8 - 8] |
875 |
movq2dq xmm6, mm6 |
movq2dq xmm6, mm6 |
876 |
movq2dq xmm7, mm7 |
movq2dq xmm7, mm7 |
877 |
movlhps xmm6, xmm6 |
movlhps xmm6, xmm6 |
878 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
879 |
|
mov eax, -16 |
880 |
|
|
881 |
xor eax, eax |
align ALIGN |
882 |
|
.loop |
883 |
align 16 |
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
884 |
.das2_loop |
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
|
movdqa xmm0, [esi + eax*8] |
|
|
movdqa xmm3, [esi + eax*8 + 16] |
|
885 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
886 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
887 |
pcmpgtw xmm1, xmm0 |
pcmpgtw xmm1, xmm0 ; sign(c) |
888 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4, xmm3 |
889 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
890 |
pxor xmm5, xmm5 |
pxor xmm5, xmm5 |
891 |
pcmpeqw xmm2, xmm0 |
pcmpeqw xmm2, xmm0 ; c is zero |
892 |
pcmpeqw xmm5, xmm3 |
pcmpeqw xmm5, xmm3 |
893 |
pandn xmm2, xmm6 |
pandn xmm2, xmm6 ; offset = isZero ? 0 : quant_add |
894 |
pandn xmm5, xmm6 |
pandn xmm5, xmm6 |
895 |
pxor xmm0, xmm1 |
pxor xmm0, xmm1 ; negate if negative |
896 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
897 |
psubw xmm0, xmm1 |
psubw xmm0, xmm1 |
898 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
899 |
pmullw xmm0, xmm7 |
pmullw xmm0, xmm7 ; *= 2Q |
900 |
pmullw xmm3, xmm7 |
pmullw xmm3, xmm7 |
901 |
paddw xmm0, xmm2 |
paddw xmm0, xmm2 ; + offset |
902 |
paddw xmm3, xmm5 |
paddw xmm3, xmm5 |
903 |
pxor xmm0, xmm1 |
paddw xmm0, xmm1 ; negate back |
904 |
pxor xmm3, xmm4 |
paddw xmm3, xmm4 |
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
905 |
|
|
906 |
%ifdef SATURATE |
; saturates to +2047 |
907 |
movdqu xmm2, [sse2_pos_2047] |
movdqa xmm2, [sse2_2047] |
|
movdqu xmm4, [sse2_neg_2048] |
|
908 |
pminsw xmm0, xmm2 |
pminsw xmm0, xmm2 |
|
pminsw xmm3, xmm2 |
|
|
pmaxsw xmm0, xmm4 |
|
|
pmaxsw xmm3, xmm4 |
|
|
%endif |
|
|
|
|
|
movdqa [edi + eax*8], xmm0 |
|
|
movdqa [edi + eax*8 + 16], xmm3 |
|
|
|
|
909 |
add eax, 4 |
add eax, 4 |
910 |
cmp eax, 16 |
pminsw xmm3, xmm2 |
|
jnz near .das2_loop |
|
911 |
|
|
912 |
mov ax, [esi] ; ax = data[0] |
pxor xmm0, xmm1 |
913 |
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
pxor xmm3, xmm4 |
914 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
915 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
916 |
|
jnz near .loop |
917 |
|
|
918 |
%ifdef SATURATE |
; deal with DC |
919 |
cmp ax, -2048 |
movd mm0, [ecx] |
920 |
jl .das2_set_n2048 |
pmullw mm0, [esp+16] ; dcscalar |
921 |
cmp ax, 2047 |
movq mm2, [mmx_32767_minus_2047] |
922 |
jg .das2_set_2047 |
paddsw mm0, mm2 |
923 |
%endif |
psubsw mm0, mm2 |
924 |
mov [edi], ax |
movq mm2, [mmx_32768_minus_2048] |
925 |
|
psubsw mm0, mm2 |
926 |
|
paddsw mm0, mm2 |
927 |
|
movd eax, mm0 |
928 |
|
mov [edx], ax |
929 |
|
|
|
pop edi |
|
|
pop esi |
|
930 |
ret |
ret |
931 |
|
|
|
%ifdef SATURATE |
|
|
align 16 |
|
|
.das2_set_n2048 |
|
|
mov word [edi], -2048 |
|
|
pop edi |
|
|
pop esi |
|
|
ret |
|
932 |
|
|
933 |
align 16 |
|
|
.das2_set_2047 |
|
|
mov word [edi], 2047 |
|
|
pop edi |
|
|
pop esi |
|
|
ret |
|
|
%endif |
|
934 |
|
|
935 |
;=========================================================================== |
;=========================================================================== |
936 |
; |
; |
1060 |
; const uint32_t quant); |
; const uint32_t quant); |
1061 |
; |
; |
1062 |
;=========================================================================== |
;=========================================================================== |
1063 |
|
align ALIGN |
|
align 16 |
|
1064 |
cglobal dequant_inter_sse2 |
cglobal dequant_inter_sse2 |
1065 |
dequant_inter_sse2 |
dequant_inter_sse2 |
1066 |
|
mov edx, [esp + 4] ; data |
1067 |
push esi |
mov ecx, [esp + 8] ; coeff |
1068 |
push edi |
mov eax, [esp + 12] ; quant |
|
|
|
|
mov edi, [esp + 8 + 4] ; data |
|
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
1069 |
movq mm6, [mmx_add + eax * 8 - 8] |
movq mm6, [mmx_add + eax * 8 - 8] |
1070 |
movq mm7, [mmx_mul + eax * 8 - 8] |
movq mm7, [mmx_mul + eax * 8 - 8] |
|
|
|
1071 |
movq2dq xmm6, mm6 |
movq2dq xmm6, mm6 |
1072 |
movq2dq xmm7, mm7 |
movq2dq xmm7, mm7 |
1073 |
movlhps xmm6, xmm6 |
movlhps xmm6, xmm6 |
1074 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
1075 |
|
mov eax, -16 |
1076 |
|
|
1077 |
xor eax, eax |
align ALIGN |
1078 |
|
.loop |
1079 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
1080 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
1081 |
|
|
|
align 16 |
|
|
.des2_loop |
|
|
movdqa xmm0, [esi + eax*8] ; xmm0 = [coeff] |
|
|
movdqa xmm3, [esi + eax*8 + 16] |
|
1082 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
1083 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
1084 |
pcmpgtw xmm1, xmm0 |
pcmpgtw xmm1, xmm0 ; sign(c) |
1085 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4, xmm3 |
1086 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
1087 |
pxor xmm5, xmm5 |
pxor xmm5, xmm5 |
1088 |
pcmpeqw xmm2, xmm0 |
pcmpeqw xmm2, xmm0 ; c is zero |
1089 |
pcmpeqw xmm5, xmm3 |
pcmpeqw xmm5, xmm3 |
1090 |
pandn xmm2, xmm6 |
pandn xmm2, xmm6 |
1091 |
pandn xmm5, xmm6 |
pandn xmm5, xmm6 |
1092 |
pxor xmm0, xmm1 |
pxor xmm0, xmm1 ; negate if negative |
1093 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
1094 |
psubw xmm0, xmm1 |
psubw xmm0, xmm1 |
1095 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
1096 |
pmullw xmm0, xmm7 |
pmullw xmm0, xmm7 ; *= 2Q |
1097 |
pmullw xmm3, xmm7 |
pmullw xmm3, xmm7 |
1098 |
paddw xmm0, xmm2 |
paddw xmm0, xmm2 ; + offset |
1099 |
paddw xmm3, xmm5 |
paddw xmm3, xmm5 |
|
pxor xmm0, xmm1 |
|
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
|
|
|
|
%ifdef SATURATE |
|
|
movdqu xmm2, [sse2_pos_2047] |
|
|
movdqu xmm4, [sse2_neg_2048] |
|
|
pminsw xmm0, xmm2 |
|
|
pminsw xmm3, xmm2 |
|
|
pmaxsw xmm0, xmm4 |
|
|
pmaxsw xmm3, xmm4 |
|
|
%endif |
|
1100 |
|
|
1101 |
movdqa [edi + eax*8], xmm0 |
paddw xmm0, xmm1 ; start restoring sign |
1102 |
movdqa [edi + eax*8 + 16], xmm3 |
paddw xmm3, xmm4 |
1103 |
|
|
1104 |
|
; saturates to +2047 |
1105 |
|
movdqa xmm2, [sse2_2047] |
1106 |
|
pminsw xmm0, xmm2 |
1107 |
add eax, 4 |
add eax, 4 |
1108 |
cmp eax, 16 |
pminsw xmm3, xmm2 |
|
jnz near .des2_loop |
|
1109 |
|
|
1110 |
pop edi |
pxor xmm0, xmm1 ; finish restoring sign |
1111 |
pop esi |
pxor xmm3, xmm4 |
1112 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
1113 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
1114 |
|
jnz near .loop |
1115 |
|
|
1116 |
ret |
ret |