|
@@ -68,6 +68,19 @@
|
|
|
.text
|
|
|
.align 6
|
|
|
|
|
|
+function idct_row4_pld_neon
|
|
|
+ pld [r0]
|
|
|
+ add r3, r0, r1, lsl #2
|
|
|
+ pld [r0, r1]
|
|
|
+ pld [r0, r1, lsl #1]
|
|
|
+ pld [r3, -r1]
|
|
|
+ pld [r3]
|
|
|
+ pld [r3, r1]
|
|
|
+ add r3, r3, r1, lsl #1
|
|
|
+ pld [r3]
|
|
|
+ pld [r3, r1]
|
|
|
+ .endfunc
|
|
|
+
|
|
|
function idct_row4_neon
|
|
|
vmov.i32 q15, #(1<<(ROW_SHIFT-1))
|
|
|
vld1.64 {d2-d5}, [r2,:128]!
|
|
@@ -252,7 +265,7 @@ idct_coeff_neon:
|
|
|
function ff_simple_idct_put_neon, export=1
|
|
|
idct_start r2
|
|
|
|
|
|
- bl idct_row4_neon
|
|
|
+ bl idct_row4_pld_neon
|
|
|
bl idct_row4_neon
|
|
|
add r2, r2, #-128
|
|
|
bl idct_col4_neon
|
|
@@ -307,7 +320,7 @@ function idct_col4_add8_neon
|
|
|
function ff_simple_idct_add_neon, export=1
|
|
|
idct_start r2
|
|
|
|
|
|
- bl idct_row4_neon
|
|
|
+ bl idct_row4_pld_neon
|
|
|
bl idct_row4_neon
|
|
|
add r2, r2, #-128
|
|
|
bl idct_col4_neon
|