diff options
Diffstat (limited to 'media/libstagefright/codecs/on2/h264dec/omxdl/arm_neon/vc/m4p10/src_gcc/omxVCM4P10_FilterDeblockingLuma_VerEdge_I_s.S')
-rw-r--r-- | media/libstagefright/codecs/on2/h264dec/omxdl/arm_neon/vc/m4p10/src_gcc/omxVCM4P10_FilterDeblockingLuma_VerEdge_I_s.S | 157 |
1 files changed, 157 insertions, 0 deletions
diff --git a/media/libstagefright/codecs/on2/h264dec/omxdl/arm_neon/vc/m4p10/src_gcc/omxVCM4P10_FilterDeblockingLuma_VerEdge_I_s.S b/media/libstagefright/codecs/on2/h264dec/omxdl/arm_neon/vc/m4p10/src_gcc/omxVCM4P10_FilterDeblockingLuma_VerEdge_I_s.S new file mode 100644 index 0000000..66cc32e --- /dev/null +++ b/media/libstagefright/codecs/on2/h264dec/omxdl/arm_neon/vc/m4p10/src_gcc/omxVCM4P10_FilterDeblockingLuma_VerEdge_I_s.S @@ -0,0 +1,157 @@ +/* + * (c) Copyright 2007-2008 ARM Limited. All Rights Reserved. + * + */ + + .eabi_attribute 24, 1 + .eabi_attribute 25, 1 + + .arm + .fpu neon + .text + + .global omxVCM4P10_FilterDeblockingLuma_VerEdge_I + .func omxVCM4P10_FilterDeblockingLuma_VerEdge_I +omxVCM4P10_FilterDeblockingLuma_VerEdge_I: + PUSH {r4-r12,lr} + VPUSH {d8-d15} + ADD r7,r2,#1 + ADD r8,r3,#1 + VLD1.8 {d0[]},[r2] + SUB r0,r0,#4 + VLD1.8 {d2[]},[r3] + LDR r4,[sp,#0x6c] + LDR r5,[sp,#0x68] + MOV r6,#0 + VMOV.I8 d14,#0 + VMOV.I8 d15,#0x1 + MOV r9,#0x11000000 + ADD r11,r1,r1 +L0x38: + LDRH r12,[r4],#4 + CMP r12,#0 + BEQ L0x160 + ADD r10,r0,r1 + VLD1.8 {d7},[r0],r11 + VLD1.8 {d8},[r10],r11 + VLD1.8 {d5},[r0],r11 + VZIP.8 d7,d8 + VLD1.8 {d10},[r10],r11 + VLD1.8 {d6},[r0],r11 + VZIP.8 d5,d10 + VLD1.8 {d9},[r10],r11 + VLD1.8 {d4},[r0],r11 + VLD1.8 {d11},[r10],r11 + VZIP.8 d6,d9 + VZIP.16 d8,d10 + VZIP.8 d4,d11 + SUB r0,r0,r1,LSL #3 + VZIP.16 d7,d5 + VZIP.16 d9,d11 + VZIP.16 d6,d4 + VTRN.32 d8,d9 + VTRN.32 d5,d4 + VTRN.32 d10,d11 + VTRN.32 d7,d6 + VABD.U8 d13,d4,d8 + VABD.U8 d12,d5,d4 + VABD.U8 d18,d9,d8 + VABD.U8 d19,d6,d4 + TST r12,#0xff + VCGT.U8 d16,d0,d13 + VMAX.U8 d12,d18,d12 + VABD.U8 d17,d10,d8 + VMOVEQ.32 d16[0],r6 + TST r12,#0xff00 + VCGT.U8 d19,d2,d19 + VCGT.U8 d12,d2,d12 + VMOVEQ.32 d16[1],r6 + VCGT.U8 d17,d2,d17 + VAND d16,d16,d12 + TST r12,#4 + VAND d12,d16,d17 + VAND d17,d16,d19 + BNE L0x17c + BL armVCM4P10_DeblockingLumabSLT4_unsafe + VZIP.8 d7,d6 + VZIP.8 d30,d29 + VZIP.8 d24,d25 + VZIP.8 d10,d11 + VZIP.16 d7,d30 + ADD r10,r0,r1 + VZIP.16 d24,d10 + VZIP.16 d25,d11 + VZIP.16 d6,d29 + VTRN.32 d7,d24 + VTRN.32 d30,d10 + VTRN.32 d6,d25 + VTRN.32 d29,d11 + VST1.8 {d7},[r0],r11 + VST1.8 {d24},[r10],r11 + VST1.8 {d30},[r0],r11 + VST1.8 {d10},[r10],r11 + VST1.8 {d6},[r0],r11 + VST1.8 {d25},[r10],r11 + ADDS r9,r9,r9 + VST1.8 {d29},[r0],r11 + ADD r5,r5,#2 + VST1.8 {d11},[r10],r1 + SUB r0,r0,r1,LSL #3 + VLD1.8 {d0[]},[r7] + ADD r0,r0,#4 + VLD1.8 {d2[]},[r8] + BCC L0x38 + B L0x1f0 +L0x160: + ADD r0,r0,#4 + ADDS r9,r9,r9 + VLD1.8 {d0[]},[r7] + ADD r5,r5,#4 + VLD1.8 {d2[]},[r8] + BCC L0x38 + B L0x1f0 +L0x17c: + BL armVCM4P10_DeblockingLumabSGE4_unsafe + VZIP.8 d7,d31 + VZIP.8 d30,d29 + VZIP.8 d24,d25 + VZIP.8 d28,d11 + VZIP.16 d7,d30 + ADD r10,r0,r1 + VZIP.16 d24,d28 + VZIP.16 d25,d11 + VZIP.16 d31,d29 + VTRN.32 d7,d24 + VTRN.32 d30,d28 + VTRN.32 d31,d25 + VTRN.32 d29,d11 + VST1.8 {d7},[r0],r11 + VST1.8 {d24},[r10],r11 + VST1.8 {d30},[r0],r11 + VST1.8 {d28},[r10],r11 + VST1.8 {d31},[r0],r11 + VST1.8 {d25},[r10],r11 + ADDS r9,r9,r9 + VST1.8 {d29},[r0],r11 + ADD r5,r5,#4 + VST1.8 {d11},[r10],r11 + SUB r0,r0,r1,LSL #3 + VLD1.8 {d0[]},[r7] + ADD r0,r0,#4 + VLD1.8 {d2[]},[r8] + BCC L0x38 +L0x1f0: + SUB r4,r4,#0xe + SUB r5,r5,#0xe + SUB r0,r0,#0x10 + VLD1.8 {d0[]},[r2] + ADD r0,r0,r1,LSL #3 + VLD1.8 {d2[]},[r3] + BNE L0x38 + MOV r0,#0 + VPOP {d8-d15} + POP {r4-r12,pc} + .endfunc + + .end + |