57 |
cglobal transfer_8to16copy_mmx |
cglobal transfer_8to16copy_mmx |
58 |
cglobal transfer_16to8copy_mmx |
cglobal transfer_16to8copy_mmx |
59 |
cglobal transfer_8to16sub_mmx |
cglobal transfer_8to16sub_mmx |
60 |
|
cglobal transfer_8to16subro_mmx |
61 |
cglobal transfer_8to16sub2_mmx |
cglobal transfer_8to16sub2_mmx |
62 |
cglobal transfer_8to16sub2_xmm |
cglobal transfer_8to16sub2_xmm |
63 |
cglobal transfer_16to8add_mmx |
cglobal transfer_16to8add_mmx |
156 |
; * |
; * |
157 |
; *************************************************************************/ |
; *************************************************************************/ |
158 |
|
|
159 |
%macro COPY_8_TO_16_SUB 1 |
; when second argument == 1, reference (ebx) block is to current (eax) |
160 |
|
%macro COPY_8_TO_16_SUB 2 |
161 |
movq mm0, [eax] ; cur |
movq mm0, [eax] ; cur |
162 |
movq mm2, [eax+edx] |
movq mm2, [eax+edx] |
163 |
movq mm1, mm0 |
movq mm1, mm0 |
171 |
movq mm5, [ebx+edx] ; ref |
movq mm5, [ebx+edx] ; ref |
172 |
|
|
173 |
movq mm6, mm4 |
movq mm6, mm4 |
174 |
|
%if %2 == 1 |
175 |
movq [eax], mm4 |
movq [eax], mm4 |
176 |
movq [eax+edx], mm5 |
movq [eax+edx], mm5 |
177 |
|
%endif |
178 |
punpcklbw mm4, mm7 |
punpcklbw mm4, mm7 |
179 |
punpckhbw mm6, mm7 |
punpckhbw mm6, mm7 |
180 |
psubsw mm0, mm4 |
psubsw mm0, mm4 |
202 |
mov edx, [esp+4+16] ; Stride |
mov edx, [esp+4+16] ; Stride |
203 |
pxor mm7, mm7 |
pxor mm7, mm7 |
204 |
|
|
205 |
COPY_8_TO_16_SUB 0 |
COPY_8_TO_16_SUB 0, 1 |
206 |
COPY_8_TO_16_SUB 1 |
COPY_8_TO_16_SUB 1, 1 |
207 |
COPY_8_TO_16_SUB 2 |
COPY_8_TO_16_SUB 2, 1 |
208 |
COPY_8_TO_16_SUB 3 |
COPY_8_TO_16_SUB 3, 1 |
209 |
|
|
210 |
pop ebx |
pop ebx |
211 |
ret |
ret |
212 |
|
|
213 |
|
|
214 |
|
align 16 |
215 |
|
transfer_8to16subro_mmx: |
216 |
|
mov ecx, [esp + 4] ; Dst |
217 |
|
mov eax, [esp + 8] ; Cur |
218 |
|
push ebx |
219 |
|
mov ebx, [esp+4+12] ; Ref |
220 |
|
mov edx, [esp+4+16] ; Stride |
221 |
|
pxor mm7, mm7 |
222 |
|
|
223 |
|
COPY_8_TO_16_SUB 0, 0 |
224 |
|
COPY_8_TO_16_SUB 1, 0 |
225 |
|
COPY_8_TO_16_SUB 2, 0 |
226 |
|
COPY_8_TO_16_SUB 3, 0 |
227 |
|
|
228 |
|
pop ebx |
229 |
|
ret |
230 |
|
|
231 |
|
|
232 |
;=========================================================================== |
;=========================================================================== |
233 |
; |
; |
234 |
; void transfer_8to16sub2_mmx(int16_t * const dct, |
; void transfer_8to16sub2_mmx(int16_t * const dct, |