@@ -153,8 +153,7 @@ memcpy_wst_asm:
153153 * Arg 3 ~ x3: C1/Primary channel offset(index) -- > ex. 0 for first ; alpha/beta follow
154154 * Arg 4 ~ x4: Alpha
155155 * Arg 5 ~ x5: Beta
156- * Arg 6 ~ x6: Gamma
157- * Arg 7 ~ x7: Threshold * /
156+ * Arg 6 ~ x6: Threshold * /
158157memcpy_deinterlace_wstb_asm:
159158
160159 dup v4.16b , w4 // duplicate alpha across 16x8bit lanes
@@ -165,8 +164,7 @@ memcpy_deinterlace_wstb_asm:
165164 clz v5.16b , v5.16b
166165 neg v5.16b , v5.16b
167166
168- dup v6.16b , w6 // duplicate gamma across 16x8bit lanes
169- dup v7.16b , w7 // duplicate threshold '' '
167+ dup v6.16b , w6 // duplicate threshold '' '
170168
171169 lsr x2 , x2 , # 4 // right shift the count by 4 bits (divide by 16 , 16 bytes = 128 bits)
172170
@@ -182,9 +180,8 @@ memcpy_deinterlace_wstb_asm:
182180 ushl v1.16b , v1.16b , v4.16b // weight alpha -- > shift right by the amount of leading 0 's from alpha (lshift by negative)
183181 ushl v2.16b , v2.16b , v5.16b // weight beta
184182 uqadd v3.16b , v1.16b , v2.16b // add weights
185- uqadd v3.16b , v3.16b , v6.16b // add gamma
186183 uqsub v3.16b , v0.16b , v3.16b // subtract result from primary
187- cmhi v3.16b , v3.16b , v7 .16b // threshold result
184+ cmhi v3.16b , v3.16b , v6 .16b // threshold result
188185
189186 str q3 , [ x1 ], # 16 // Load out of Q3
190187
@@ -199,9 +196,8 @@ memcpy_deinterlace_wstb_asm:
199196 ushl v2.16b , v2.16b , v4.16b // weight alpha
200197 ushl v0.16b , v0.16b , v5.16b // weight beta
201198 uqadd v3.16b , v2.16b , v0.16b // add weights
202- uqadd v3.16b , v3.16b , v6.16b // add gamma
203199 uqsub v3.16b , v1.16b , v3.16b // subtract result from primary
204- cmhi v3.16b , v3.16b , v7 .16b // threshold result
200+ cmhi v3.16b , v3.16b , v6 .16b // threshold result
205201
206202 str q3 , [ x1 ], # 16 // Load out of Q3
207203
@@ -216,9 +212,8 @@ memcpy_deinterlace_wstb_asm:
216212 ushl v0.16b , v0.16b , v4.16b // weight alpha
217213 ushl v1.16b , v1.16b , v5.16b // weight beta
218214 uqadd v3.16b , v0.16b , v1.16b // add weights
219- uqadd v3.16b , v3.16b , v6.16b // add gamma
220215 uqsub v3.16b , v2.16b , v3.16b // subtract result from primary
221- cmhi v3.16b , v3.16b , v7 .16b // threshold result
216+ cmhi v3.16b , v3.16b , v6 .16b // threshold result
222217
223218 str q3 , [ x1 ], # 16 // Load out of Q3
224219
0 commit comments