21 |
; * along with this program ; if not, write to the Free Software |
; * along with this program ; if not, write to the Free Software |
22 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
; * |
; * |
24 |
; * $Id: quantize_mpeg_mmx.asm,v 1.5 2004-08-29 10:02:38 edgomez Exp $ |
; * $Id: quantize_mpeg_mmx.asm,v 1.6 2006-07-10 08:09:59 syskin Exp $ |
25 |
; * |
; * |
26 |
; *************************************************************************/ |
; *************************************************************************/ |
27 |
|
|
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, |
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 |
push ebx |
movq mm7, [mmx_rounding] |
|
|
|
|
mov edi, [esp + 16 + 4] ; coeff |
|
|
mov esi, [esp + 16 + 8] ; data |
|
|
mov eax, [esp + 16 + 12] ; quant |
|
|
mov ebx, [esp + 16 + 20] ; mpeg_quant_matrices |
|
|
|
|
|
movq mm5, [quantd + eax * 8 - 8] ; quantd -> mm5 |
|
|
|
|
|
xor ecx, ecx |
|
|
cmp al, 1 |
|
|
jz near .q1loop |
|
|
|
|
|
cmp al, 2 |
|
|
jz near .q2loop |
|
|
|
|
|
movq mm7, [mmx_div + eax * 8 - 8] ; multipliers[quant] -> mm7 |
|
|
|
|
|
ALIGN 16 |
|
|
.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, [ebx + 8*ecx] |
|
|
psrlw mm2, 1 ; intra_matrix[i]>>1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [ebx + 256 + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [ebx + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [ebx + 256 + 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 |
|
207 |
|
|
208 |
.done |
mov eax, [esp + 4 + 8] ; data |
209 |
; caclulate data[0] // (int32_t)dcscalar) |
mov ecx, [esp + 4 + 20] ; mpeg_quant_matrices |
210 |
mov ecx, [esp + 16 + 16] ; dcscalar |
mov edi, [esp + 4 + 4] ; coeff |
211 |
|
|
212 |
|
QUANT_MMX(0) |
213 |
|
QUANT_MMX(1) |
214 |
|
QUANT_MMX(2) |
215 |
|
QUANT_MMX(3) |
216 |
|
QUANT_MMX(4) |
217 |
|
QUANT_MMX(5) |
218 |
|
QUANT_MMX(6) |
219 |
|
QUANT_MMX(7) |
220 |
|
|
221 |
|
; calculate DC |
222 |
|
movsx eax, word [eax] ; data[0] |
223 |
|
mov ecx, [esp + 4 + 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 |
|
|
|
pop ebx |
|
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, [ebx + 8*ecx] |
|
|
psrlw mm2, 1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [ebx + 256 + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [ebx + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [ebx + 256 + 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 |
|
245 |
|
|
246 |
|
|
|
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, [ebx + 8*ecx] |
|
|
psrlw mm2, 1 |
|
|
paddw mm0, mm2 |
|
|
movq mm2, [ebx + 256 + ecx*8] |
|
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
|
|
movq mm2, [ebx + 8*ecx + 8] |
|
|
psrlw mm2, 1 |
|
|
paddw mm3, mm2 |
|
|
movq mm2, [ebx + 256 + 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 |
|
|
.endfunc |
|
|
|
|
247 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
248 |
; |
; |
249 |
; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, |
; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, |