Add push-pop for Neon D8-D15 registers
Bug: 68320413
Test: hevcdec with testbench updates to validate D8-D15
Change-Id: Ib62e68b69b016b1ded12cb00d1bb580de8a744c2
(cherry picked from commit 6acb36d25de9b2ac19aa4c34d6821c42e9a84c00)
diff --git a/decoder/arm/ihevcd_fmt_conv_420sp_to_rgba8888.s b/decoder/arm/ihevcd_fmt_conv_420sp_to_rgba8888.s
index a9a75cb..caf7123 100644
--- a/decoder/arm/ihevcd_fmt_conv_420sp_to_rgba8888.s
+++ b/decoder/arm/ihevcd_fmt_conv_420sp_to_rgba8888.s
@@ -68,13 +68,13 @@
@* *
@* Register Usage : R0 - R14 *
@* *
-@* Stack Usage : 40 Bytes *
+@* Stack Usage : 104 Bytes *
@* *
@* Interruptibility : Interruptible *
@* *
@* Known Limitations *
@* Assumptions: Image Width: Assumed to be multiple of 16 and *
-@* greater than or equal to 16 *
+@* greater than or equal to 16 *
@* Image Height: Assumed to be even. *
@* *
@* Revision History : *
@@ -82,6 +82,7 @@
@* 07 06 2010 Varshita Draft *
@* 07 06 2010 Naveen Kr T Completed *
@* 05 08 2013 Naveen K P Modified for HEVC *
+@* 30 10 2018 Saurabh Sood Store D registers to stack *
@*****************************************************************************/
.global ihevcd_fmt_conv_420sp_to_rgba8888_a9q
.type ihevcd_fmt_conv_420sp_to_rgba8888_a9q, function
@@ -89,7 +90,7 @@
@// push the registers on the stack
STMFD SP!,{R4-R12,LR}
-
+ VPUSH {d8-d15}
@//R0 - Y PTR
@//R1 - UV PTR
@@ -131,12 +132,12 @@
@//D0 HAS C1-C2-C3-C4
@// load other parameters from stack
- LDR R5,[sp,#40]
+ LDR R5,[sp,#104]
@LDR R4,[sp,#44]
- LDR R6,[sp,#44]
- LDR R7,[sp,#48]
+ LDR R6,[sp,#108]
+ LDR R7,[sp,#112]
@LDR R8,[sp,#52]
- LDR R9,[sp,#52]
+ LDR R9,[sp,#116]
@// calculate offsets, offset = stride - width
SUB R10,R6,R3 @// luma offset
@@ -445,10 +446,8 @@
BNE LABEL_YUV420SP_TO_RGB8888_HEIGHT_LOOP
@//POP THE REGISTERS
+ VPOP {d8-d15}
LDMFD SP!,{R4-R12,PC}
-
-
.section .note.GNU-stack,"",%progbits
-
diff --git a/decoder/arm/ihevcd_itrans_recon_dc_chroma.s b/decoder/arm/ihevcd_itrans_recon_dc_chroma.s
index 6732ce0..9184b80 100644
--- a/decoder/arm/ihevcd_itrans_recon_dc_chroma.s
+++ b/decoder/arm/ihevcd_itrans_recon_dc_chroma.s
@@ -58,9 +58,9 @@
push {r0-r11,lr}
- ldr r4,[sp,#0x34] @loads log2_trans_size
- ldr r5,[sp,#0x38] @ loads i2_coeff_value
-
+ vpush {d8-d15}
+ ldr r4,[sp,#0x74] @loads log2_trans_size
+ ldr r5,[sp,#0x78] @ loads i2_coeff_value
mov r10,#1
lsl r4,r10,r4 @ trans_size = (1 << log2_trans_size)@
mov r6,#64 @ 1 << (shift1 - 1)@
@@ -188,6 +188,7 @@
vst1.u32 {d8},[r1]
end_loops_chroma:
+ vpop {d8-d15}
pop {r0-r11,pc}
diff --git a/decoder/arm/ihevcd_itrans_recon_dc_luma.s b/decoder/arm/ihevcd_itrans_recon_dc_luma.s
index 8aee84c..da1f1ad 100644
--- a/decoder/arm/ihevcd_itrans_recon_dc_luma.s
+++ b/decoder/arm/ihevcd_itrans_recon_dc_luma.s
@@ -59,9 +59,9 @@
push {r0-r11,lr}
- ldr r4,[sp,#0x34] @loads log2_trans_size
- ldr r5,[sp,#0x38] @ loads i2_coeff_value
-
+ vpush {d8-d15}
+ ldr r4,[sp,#0x74] @loads log2_trans_size
+ ldr r5,[sp,#0x78] @ loads i2_coeff_value
mov r10,#1
lsl r4,r10,r4 @ trans_size = (1 << log2_trans_size)@
mov r6,#64 @ 1 << (shift1 - 1)@
@@ -182,6 +182,7 @@
vst1.u32 {d5[0]},[r1]
end_loops:
+ vpop {d8-d15}
pop {r0-r11,pc}
