Minor speedups to celt_pitch_xcorr_edsp().
This moves some stuff out of the main loop for the unaligned case.
diff --git a/celt/arm/celt_pitch_xcorr_arm.s b/celt/arm/celt_pitch_xcorr_arm.s
index 0ad8f5b..09917b1 100644
--- a/celt/arm/celt_pitch_xcorr_arm.s
+++ b/celt/arm/celt_pitch_xcorr_arm.s
@@ -371,20 +371,19 @@
LDRH r8, [r5], #2
BLE celt_pitch_xcorr_edsp_process1u_loop4_done
LDR r6, [r4], #4
- LDR r9, [r5], #4
- LDR r7, [r4], #4
+ MOV r8, r8, LSL #16
celt_pitch_xcorr_edsp_process1u_loop4
- SMLABB r14, r6, r8, r14 ; sum = MAC16_16(sum, x_0, y_0)
- SUBS r12, r12, #4 ; j-=4
+ LDR r9, [r5], #4
+ SMLABT r14, r6, r8, r14 ; sum = MAC16_16(sum, x_0, y_0)
+ LDR r7, [r4], #4
SMLATB r14, r6, r9, r14 ; sum = MAC16_16(sum, x_1, y_1)
- LDR r10, [r5], #4
+ LDR r8, [r5], #4
SMLABT r14, r7, r9, r14 ; sum = MAC16_16(sum, x_2, y_2)
+ SUBS r12, r12, #4 ; j-=4
+ SMLATB r14, r7, r8, r14 ; sum = MAC16_16(sum, x_3, y_3)
LDRGT r6, [r4], #4
- SMLATB r14, r7, r10, r14 ; sum = MAC16_16(sum, x_3, y_3)
- LDRGT r9, [r5], #4
- MOV r8, r10, LSR #16
- LDRGT r7, [r4], #4
BGT celt_pitch_xcorr_edsp_process1u_loop4
+ MOV r8, r8, LSR #16
celt_pitch_xcorr_edsp_process1u_loop4_done
ADDS r12, r12, #4
celt_pitch_xcorr_edsp_process1u_loop1
@@ -530,6 +529,7 @@
ADDS r12, r12, #1
LDRGEH r6, [r4], #2
LDRGEH r8, [r5], #2
+ ; Stall
SMLABBGE r14, r6, r8, r14 ; sum = MAC16_16(sum, *x, *y)
; maxcorr = max(maxcorr, sum)
CMP r0, r14
