Lines Matching refs:MM
33 #undef MM
43 #define MM "%%xmm"
58 #define MM "%%mm"
144 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
145 SPREADW(MM"3")
146 "pxor "MM"7, "MM"7 \n\t" // 0
147 "pxor "MM"4, "MM"4 \n\t" // 0
148 MOVQ" (%2), "MM"5 \n\t" // qmat[0]
149 "pxor "MM"6, "MM"6 \n\t"
150 "psubw (%3), "MM"6 \n\t" // -bias[0]
154 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
155 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
156 "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
157 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
158 "por "MM"0, "MM"4 \n\t"
159 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
160 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
161 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
162 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
163 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
164 "pandn "MM"1, "MM"0 \n\t"
165 PMAXW(MM"0", MM"3")
168 PMAX(MM"3", MM"0")
169 "movd "MM"3, %%"FF_REG_a" \n\t"
179 "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
180 SPREADW(MM"3")
181 "pxor "MM"7, "MM"7 \n\t" // 0
182 "pxor "MM"4, "MM"4 \n\t" // 0
186 MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
187 SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
188 MOVQ" (%3, %%"FF_REG_a"), "MM"6 \n\t" // bias[0]
189 "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
190 MOVQ" (%2, %%"FF_REG_a"), "MM"5 \n\t" // qmat[i]
191 "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
192 "por "MM"0, "MM"4 \n\t"
193 RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
194 MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
195 "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
196 MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
197 MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
198 "pandn "MM"1, "MM"0 \n\t"
199 PMAXW(MM"0", MM"3")
202 PMAX(MM"3", MM"0")
203 "movd "MM"3, %%"FF_REG_a" \n\t"
213 "movd %1, "MM"1 \n\t" // max_qcoeff
214 SPREADW(MM"1")
215 "psubusw "MM"1, "MM"4 \n\t"
216 "packuswb "MM"4, "MM"4 \n\t"
218 "packsswb "MM"4, "MM"4 \n\t"
220 "movd "MM"4, %0 \n\t" // *overflow