31 |
|
|
32 |
%macro cglobal 1 |
%macro cglobal 1 |
33 |
%ifdef PREFIX |
%ifdef PREFIX |
34 |
|
%ifdef MARK_FUNCS |
35 |
|
global _%1:function %1.endfunc-%1 |
36 |
|
%define %1 _%1:function %1.endfunc-%1 |
37 |
|
%else |
38 |
global _%1 |
global _%1 |
39 |
%define %1 _%1 |
%define %1 _%1 |
40 |
|
%endif |
41 |
|
%else |
42 |
|
%ifdef MARK_FUNCS |
43 |
|
global %1:function %1.endfunc-%1 |
44 |
%else |
%else |
45 |
global %1 |
global %1 |
46 |
%endif |
%endif |
47 |
|
%endif |
48 |
%endmacro |
%endmacro |
49 |
|
|
50 |
%macro cextern 1 |
%macro cextern 1 |
61 |
;============================================================================= |
;============================================================================= |
62 |
|
|
63 |
%ifdef FORMAT_COFF |
%ifdef FORMAT_COFF |
64 |
SECTION .rodata data |
SECTION .rodata |
65 |
%else |
%else |
66 |
SECTION .rodata data align=16 |
SECTION .rodata align=16 |
67 |
%endif |
%endif |
68 |
|
|
69 |
mmx_one: |
mmx_one: |
85 |
%assign quant quant+1 |
%assign quant quant+1 |
86 |
%endrep |
%endrep |
87 |
|
|
|
;----------------------------------------------------------------------------- |
|
|
; intra matrix |
|
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
cextern intra_matrix |
|
|
cextern intra_matrix_fix |
|
|
|
|
|
;----------------------------------------------------------------------------- |
|
|
; inter matrix |
|
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
cextern inter_matrix |
|
|
cextern inter_matrix_fix |
|
|
|
|
|
|
|
88 |
%define VM18P 3 |
%define VM18P 3 |
89 |
%define VM18Q 4 |
%define VM18Q 4 |
90 |
|
|
129 |
times 4 dw 0 |
times 4 dw 0 |
130 |
|
|
131 |
;============================================================================= |
;============================================================================= |
132 |
|
; rounding |
133 |
|
;============================================================================= |
134 |
|
|
135 |
|
mmx_rounding: |
136 |
|
dw (1<<13) |
137 |
|
dw 0 |
138 |
|
dw (1<<13) |
139 |
|
dw 0 |
140 |
|
|
141 |
|
;============================================================================= |
142 |
; Code |
; Code |
143 |
;============================================================================= |
;============================================================================= |
144 |
|
|
149 |
cglobal dequant_mpeg_intra_mmx |
cglobal dequant_mpeg_intra_mmx |
150 |
cglobal dequant_mpeg_inter_mmx |
cglobal dequant_mpeg_inter_mmx |
151 |
|
|
152 |
|
|
153 |
|
%macro QUANT_MMX 1 |
154 |
|
movq mm0, [eax + 16*(%1)] ; data |
155 |
|
movq mm2, [ecx + 16*(%1) + 128] ; intra_matrix_rec |
156 |
|
movq mm4, [eax + 16*(%1) + 8] ; data |
157 |
|
movq mm6, [ecx + 16*(%1) + 128 + 8] ; intra_matrix_rec |
158 |
|
|
159 |
|
movq mm1, mm0 |
160 |
|
movq mm5, mm4 |
161 |
|
|
162 |
|
pmullw mm0, mm2 ; low results |
163 |
|
pmulhw mm1, mm2 ; high results |
164 |
|
pmullw mm4, mm6 ; low results |
165 |
|
pmulhw mm5, mm6 ; high results |
166 |
|
|
167 |
|
movq mm2, mm0 |
168 |
|
movq mm6, mm4 |
169 |
|
|
170 |
|
punpckhwd mm0, mm1 |
171 |
|
punpcklwd mm2, mm1 |
172 |
|
punpckhwd mm4, mm5 |
173 |
|
punpcklwd mm6, mm5 |
174 |
|
|
175 |
|
paddd mm2, mm7 |
176 |
|
paddd mm0, mm7 |
177 |
|
paddd mm6, mm7 |
178 |
|
paddd mm4, mm7 |
179 |
|
|
180 |
|
psrad mm2, 14 |
181 |
|
psrad mm0, 14 |
182 |
|
psrad mm6, 14 |
183 |
|
psrad mm4, 14 |
184 |
|
|
185 |
|
packssdw mm2, mm0 |
186 |
|
packssdw mm6, mm4 |
187 |
|
|
188 |
|
movq [edi + 16*(%1)], mm2 |
189 |
|
movq [edi + 16*(%1)+8], mm6 |
190 |
|
%endmacro |
191 |
|
|
192 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
193 |
; |
; |
194 |
; uint32_t quant_mpeg_intra_mmx(int16_t * coeff, |
; uint32_t quant_mpeg_intra_mmx(int16_t * coeff, |
195 |
; const int16_t const * data, |
; const int16_t const * data, |
196 |
; const uint32_t quant, |
; const uint32_t quant, |
197 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar, |
198 |
|
; const uint16_t *mpeg_matrices); |
199 |
; |
; |
200 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
201 |
|
|
202 |
ALIGN 16 |
ALIGN 16 |
203 |
quant_mpeg_intra_mmx: |
quant_mpeg_intra_mmx: |
204 |
|
|
|
push ecx |
|
|
push esi |
|
205 |
push edi |
push edi |
206 |
|
movq mm7, [mmx_rounding] |
207 |
|
|
208 |
mov edi, [esp + 12 + 4] ; coeff |
mov eax, [esp + 4 + 8] ; data |
209 |
mov esi, [esp + 12 + 8] ; data |
mov ecx, [esp + 4 + 20] ; mpeg_quant_matrices |
210 |
mov eax, [esp + 12 + 12] ; quant |
mov edi, [esp + 4 + 4] ; coeff |
211 |
|
|
212 |
movq mm5, [quantd + eax * 8 - 8] ; quantd -> mm5 |
QUANT_MMX(0) |
213 |
|
QUANT_MMX(1) |
214 |
xor ecx, ecx |
QUANT_MMX(2) |
215 |
cmp al, 1 |
QUANT_MMX(3) |
216 |
jz near .q1loop |
QUANT_MMX(4) |
217 |
|
QUANT_MMX(5) |
218 |
cmp al, 2 |
QUANT_MMX(6) |
219 |
jz near .q2loop |
QUANT_MMX(7) |
220 |
|
|
221 |
movq mm7, [mmx_div + eax * 8 - 8] ; multipliers[quant] -> mm7 |
; calculate DC |
222 |
|
movsx eax, word [eax] ; data[0] |
223 |
ALIGN 16 |
mov ecx, [esp + 4 + 16] ; dcscalar |
|
.loop |
|
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
|
|
movq mm3, [esi + 8*ecx + 8] ; |
|
|
pxor mm1, mm1 ; mm1 = 0 |
|
|
pxor mm4, mm4 |
|
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
|
|
pcmpgtw mm4, mm3 |
|
|
pxor mm0, mm1 ; mm0 = |mm0| |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; displace |
|
|
psubw mm3, mm4 ; |
|
|
psllw mm0, 4 ; level << 4 |
|
|
psllw mm3, 4 |
|
|
movq mm2, [intra_matrix + 8*ecx] |
|
|
psrlw mm2, 1 ; intra_matrix[i]>>1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [intra_matrix + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
|
|
pmulhw mm3, mm2 |
|
|
paddw mm0, mm5 ; + quantd |
|
|
paddw mm3, mm5 |
|
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
|
|
pmulhw mm3, mm7 ; |
|
|
psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 |
|
|
psrlw mm3, 1 |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; undisplace |
|
|
psubw mm3, mm4 ; |
|
|
|
|
|
movq [edi + 8*ecx], mm0 |
|
|
movq [edi + 8*ecx + 8], mm3 |
|
|
|
|
|
add ecx,2 |
|
|
cmp ecx,16 |
|
|
jnz near .loop |
|
|
|
|
|
.done |
|
|
; caclulate data[0] // (int32_t)dcscalar) |
|
|
mov ecx, [esp + 12 + 16] ; dcscalar |
|
224 |
mov edx, ecx |
mov edx, ecx |
|
movsx eax, word [esi] ; data[0] |
|
225 |
shr edx, 1 ; edx = dcscalar /2 |
shr edx, 1 ; edx = dcscalar /2 |
226 |
cmp eax, 0 |
mov edi, edx |
227 |
jg .gtzero |
neg edi |
228 |
|
|
229 |
|
cmp eax, 0 |
230 |
|
cmovg edx, edi |
231 |
sub eax, edx |
sub eax, edx |
232 |
jmp short .mul |
|
233 |
.gtzero |
mov edi, [esp + 4 + 4] ; coeff again |
234 |
add eax, edx |
|
|
.mul |
|
235 |
cdq ; expand eax -> edx:eax |
cdq ; expand eax -> edx:eax |
236 |
idiv ecx ; eax = edx:eax / dcscalar |
idiv ecx ; eax = edx:eax / dcscalar |
237 |
|
|
238 |
mov [edi], ax ; coeff[0] = ax |
mov [edi], ax ; coeff[0] = ax |
239 |
|
|
240 |
pop edi |
pop edi |
|
pop esi |
|
|
pop ecx |
|
241 |
|
|
242 |
xor eax, eax ; return(0); |
xor eax, eax ; return(0); |
243 |
ret |
ret |
244 |
|
|
|
ALIGN 16 |
|
|
.q1loop |
|
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
|
|
movq mm3, [esi + 8*ecx + 8] ; |
|
|
pxor mm1, mm1 ; mm1 = 0 |
|
|
pxor mm4, mm4 ; |
|
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
|
|
pcmpgtw mm4, mm3 ; |
|
|
pxor mm0, mm1 ; mm0 = |mm0| |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; displace |
|
|
psubw mm3, mm4 ; |
|
|
psllw mm0, 4 |
|
|
psllw mm3, 4 |
|
|
movq mm2, [intra_matrix + 8*ecx] |
|
|
psrlw mm2, 1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [intra_matrix + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
|
|
pmulhw mm3, mm2 |
|
|
paddw mm0, mm5 |
|
|
paddw mm3, mm5 |
|
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
|
|
psrlw mm3, 1 ; |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; undisplace |
|
|
psubw mm3, mm4 ; |
|
|
movq [edi + 8*ecx], mm0 |
|
|
movq [edi + 8*ecx + 8], mm3 |
|
|
|
|
|
add ecx, 2 |
|
|
cmp ecx, 16 |
|
|
jnz near .q1loop |
|
|
jmp near .done |
|
|
|
|
|
|
|
|
ALIGN 16 |
|
|
.q2loop |
|
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
|
|
movq mm3, [esi + 8*ecx + 8] ; |
|
|
pxor mm1, mm1 ; mm1 = 0 |
|
|
pxor mm4, mm4 ; |
|
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
|
|
pcmpgtw mm4, mm3 ; |
|
|
pxor mm0, mm1 ; mm0 = |mm0| |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; displace |
|
|
psubw mm3, mm4 ; |
|
|
psllw mm0, 4 |
|
|
psllw mm3, 4 |
|
|
movq mm2, [intra_matrix + 8*ecx] |
|
|
psrlw mm2, 1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [intra_matrix + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
|
|
pmulhw mm3, mm2 |
|
|
paddw mm0, mm5 |
|
|
paddw mm3, mm5 |
|
|
psrlw mm0, 2 ; mm0 >>= 1 (/4) |
|
|
psrlw mm3, 2 ; |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; undisplace |
|
|
psubw mm3, mm4 ; |
|
|
movq [edi + 8*ecx], mm0 |
|
|
movq [edi + 8*ecx + 8], mm3 |
|
|
|
|
|
add ecx,2 |
|
|
cmp ecx,16 |
|
|
jnz near .q2loop |
|
|
jmp near .done |
|
245 |
|
|
246 |
|
|
247 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
248 |
; |
; |
249 |
; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, |
; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, |
250 |
; const int16_t const * data, |
; const int16_t const * data, |
251 |
; const uint32_t quant); |
; const uint32_t quant, |
252 |
|
; const uint16_t *mpeg_matrices); |
253 |
; |
; |
254 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
255 |
|
|
259 |
push ecx |
push ecx |
260 |
push esi |
push esi |
261 |
push edi |
push edi |
262 |
|
push ebx |
263 |
|
|
264 |
mov edi, [esp + 12 + 4] ; coeff |
mov edi, [esp + 16 + 4] ; coeff |
265 |
mov esi, [esp + 12 + 8] ; data |
mov esi, [esp + 16 + 8] ; data |
266 |
mov eax, [esp + 12 + 12] ; quant |
mov eax, [esp + 16 + 12] ; quant |
267 |
|
mov ebx, [esp + 16 + 16] ; mpeg_quant_matrices |
268 |
|
|
269 |
xor ecx, ecx |
xor ecx, ecx |
270 |
|
|
292 |
psubw mm3, mm4 ; |
psubw mm3, mm4 ; |
293 |
psllw mm0, 4 |
psllw mm0, 4 |
294 |
psllw mm3, 4 |
psllw mm3, 4 |
295 |
movq mm2, [inter_matrix + 8*ecx] |
movq mm2, [ebx + 512 + 8*ecx] |
296 |
psrlw mm2, 1 |
psrlw mm2, 1 |
297 |
paddw mm0, mm2 |
paddw mm0, mm2 |
298 |
movq mm2, [inter_matrix_fix + ecx*8] |
movq mm2, [ebx + 768 + ecx*8] |
299 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
300 |
movq mm2, [inter_matrix + 8*ecx + 8] |
movq mm2, [ebx + 512 + 8*ecx + 8] |
301 |
psrlw mm2, 1 |
psrlw mm2, 1 |
302 |
paddw mm3, mm2 |
paddw mm3, mm2 |
303 |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
movq mm2, [ebx + 768 + ecx*8 + 8] |
304 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
305 |
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
306 |
pmulhw mm3, mm7 ; |
pmulhw mm3, mm7 ; |
326 |
paddd mm0, mm5 |
paddd mm0, mm5 |
327 |
movd eax, mm0 ; return sum |
movd eax, mm0 ; return sum |
328 |
|
|
329 |
|
pop ebx |
330 |
pop edi |
pop edi |
331 |
pop esi |
pop esi |
332 |
pop ecx |
pop ecx |
347 |
psubw mm3, mm4 ; |
psubw mm3, mm4 ; |
348 |
psllw mm0, 4 |
psllw mm0, 4 |
349 |
psllw mm3, 4 |
psllw mm3, 4 |
350 |
movq mm2, [inter_matrix + 8*ecx] |
movq mm2, [ebx + 512 + 8*ecx] |
351 |
psrlw mm2, 1 |
psrlw mm2, 1 |
352 |
paddw mm0, mm2 |
paddw mm0, mm2 |
353 |
movq mm2, [inter_matrix_fix + ecx*8] |
movq mm2, [ebx + 768 + ecx*8] |
354 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
355 |
movq mm2, [inter_matrix + 8*ecx + 8] |
movq mm2, [ebx + 512 + 8*ecx + 8] |
356 |
psrlw mm2, 1 |
psrlw mm2, 1 |
357 |
paddw mm3, mm2 |
paddw mm3, mm2 |
358 |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
movq mm2, [ebx + 768 + ecx*8 + 8] |
359 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
360 |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
361 |
psrlw mm3, 1 ; |
psrlw mm3, 1 ; |
374 |
|
|
375 |
jmp .done |
jmp .done |
376 |
|
|
|
|
|
377 |
ALIGN 16 |
ALIGN 16 |
378 |
.q2loop |
.q2loop |
379 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
388 |
psubw mm3, mm4 ; |
psubw mm3, mm4 ; |
389 |
psllw mm0, 4 |
psllw mm0, 4 |
390 |
psllw mm3, 4 |
psllw mm3, 4 |
391 |
movq mm2, [inter_matrix + 8*ecx] |
movq mm2, [ebx + 512 + 8*ecx] |
392 |
psrlw mm2, 1 |
psrlw mm2, 1 |
393 |
paddw mm0, mm2 |
paddw mm0, mm2 |
394 |
movq mm2, [inter_matrix_fix + ecx*8] |
movq mm2, [ebx + 768 + ecx*8] |
395 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
396 |
movq mm2, [inter_matrix + 8*ecx + 8] |
movq mm2, [ebx + 512 + 8*ecx + 8] |
397 |
psrlw mm2, 1 |
psrlw mm2, 1 |
398 |
paddw mm3, mm2 |
paddw mm3, mm2 |
399 |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
movq mm2, [ebx + 768 + ecx*8 + 8] |
400 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
401 |
psrlw mm0, 2 ; mm0 >>= 1 (/2) |
psrlw mm0, 2 ; mm0 >>= 1 (/2) |
402 |
psrlw mm3, 2 ; |
psrlw mm3, 2 ; |
414 |
jnz near .q2loop |
jnz near .q2loop |
415 |
|
|
416 |
jmp .done |
jmp .done |
417 |
|
.endfunc |
418 |
|
|
419 |
|
|
420 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
422 |
; uint32_t dequant_mpeg_intra_mmx(int16_t *data, |
; uint32_t dequant_mpeg_intra_mmx(int16_t *data, |
423 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
424 |
; const uint32_t quant, |
; const uint32_t quant, |
425 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar, |
426 |
|
; const uint16_t *mpeg_matrices); |
427 |
; |
; |
428 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
429 |
|
|
446 |
psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] |
psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] |
447 |
|
|
448 |
movq mm2, mm7 ; mm2 = quant |
movq mm2, mm7 ; mm2 = quant |
449 |
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant. |
pmullw mm2, [ebx + 8*eax + 8*16 ] ; matrix[i]*quant. |
450 |
|
|
451 |
movq mm6, mm2 |
movq mm6, mm2 |
452 |
pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) |
pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) |
469 |
ALIGN 16 |
ALIGN 16 |
470 |
dequant_mpeg_intra_mmx: |
dequant_mpeg_intra_mmx: |
471 |
|
|
472 |
mov edx, [esp+4] ; data |
push ebx |
473 |
mov ecx, [esp+8] ; coeff |
|
474 |
mov eax, [esp+12] ; quant |
mov edx, [esp + 4 + 4] ; data |
475 |
|
mov ecx, [esp + 4 + 8] ; coeff |
476 |
|
mov eax, [esp + 4 + 12] ; quant |
477 |
|
mov ebx, [esp + 4 + 20] ; mpeg_quant_matrices |
478 |
|
|
479 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
480 |
mov eax, -16 ; to keep ALIGNed, we regularly process coeff[0] |
mov eax, -16 ; to keep ALIGNed, we regularly process coeff[0] |
491 |
movq mm2, mm7 ; mm2 = quant |
movq mm2, mm7 ; mm2 = quant |
492 |
|
|
493 |
pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
494 |
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant |
pmullw mm2, [ebx + 8*eax + 8*16 ] ; matrix[i]*quant |
495 |
|
|
496 |
pxor mm0, mm1 ; negate if negative |
pxor mm0, mm1 ; negate if negative |
497 |
pxor mm3, mm4 ; negate if negative |
pxor mm3, mm4 ; negate if negative |
507 |
pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
508 |
movq mm5, mm7 ; mm2 = quant |
movq mm5, mm7 ; mm2 = quant |
509 |
|
|
510 |
pmullw mm5, [intra_matrix + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
pmullw mm5, [ebx + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
511 |
|
|
512 |
movq mm6, mm5 |
movq mm6, mm5 |
513 |
add eax,2 ; z-flag will be tested later |
add eax,2 ; z-flag will be tested later |
539 |
|
|
540 |
; deal with DC |
; deal with DC |
541 |
movd mm0, [ecx] |
movd mm0, [ecx] |
542 |
pmullw mm0, [esp+16] ; dcscalar |
pmullw mm0, [esp + 4 + 16] ; dcscalar |
543 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
544 |
paddsw mm0, mm2 |
paddsw mm0, mm2 |
545 |
psubsw mm0, mm2 |
psubsw mm0, mm2 |
550 |
mov [edx], ax |
mov [edx], ax |
551 |
|
|
552 |
xor eax, eax |
xor eax, eax |
553 |
|
|
554 |
|
pop ebx |
555 |
|
|
556 |
ret |
ret |
557 |
|
.endfunc |
558 |
|
|
559 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
560 |
; |
; |
561 |
; uint32_t dequant_mpeg_inter_mmx(int16_t * data, |
; uint32_t dequant_mpeg_inter_mmx(int16_t * data, |
562 |
; const int16_t * const coeff, |
; const int16_t * const coeff, |
563 |
; const uint32_t quant); |
; const uint32_t quant, |
564 |
|
; const uint16_t *mpeg_matrices); |
565 |
; |
; |
566 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
567 |
|
|
573 |
ALIGN 16 |
ALIGN 16 |
574 |
dequant_mpeg_inter_mmx: |
dequant_mpeg_inter_mmx: |
575 |
|
|
576 |
mov edx, [esp+ 4] ; data |
push ebx |
577 |
mov ecx, [esp+ 8] ; coeff |
|
578 |
mov eax, [esp+12] ; quant |
mov edx, [esp + 4 + 4] ; data |
579 |
|
mov ecx, [esp + 4 + 8] ; coeff |
580 |
|
mov eax, [esp + 4 + 12] ; quant |
581 |
|
mov ebx, [esp + 4 + 16] ; mpeg_quant_matrices |
582 |
|
|
583 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
584 |
mov eax, -16 |
mov eax, -16 |
585 |
paddw mm7, mm7 ; << 1 |
paddw mm7, mm7 ; << 1 |
614 |
; we're short on register, here. Poor pairing... |
; we're short on register, here. Poor pairing... |
615 |
|
|
616 |
movq mm4, mm7 ; (matrix*quant) |
movq mm4, mm7 ; (matrix*quant) |
617 |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8] |
pmullw mm4, [ebx + 512 + 8*eax + 8*16 -2*8] |
618 |
movq mm5, mm4 |
movq mm5, mm4 |
619 |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
620 |
pmullw mm0, mm4 ; low of c*(matrix*quant) |
pmullw mm0, mm4 ; low of c*(matrix*quant) |
621 |
|
|
622 |
movq mm4, mm7 ; (matrix*quant) |
movq mm4, mm7 ; (matrix*quant) |
623 |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8 + 8] |
pmullw mm4, [ebx + 512 + 8*eax + 8*16 -2*8 + 8] |
624 |
|
|
625 |
pcmpgtw mm5, [zero] |
pcmpgtw mm5, [zero] |
626 |
paddusw mm0, mm5 |
paddusw mm0, mm5 |
664 |
xor word [edx + 2*63], ax |
xor word [edx + 2*63], ax |
665 |
|
|
666 |
xor eax, eax |
xor eax, eax |
667 |
|
|
668 |
|
pop ebx |
669 |
|
|
670 |
ret |
ret |
671 |
|
.endfunc |
672 |
|
|