aosp12/external/libavc/encoder/arm/ime_distortion_metrics_a9q.s

1364 lines
44 KiB
ArmAsm

@/******************************************************************************
@ *
@ * Copyright (C) 2015 The Android Open Source Project
@ *
@ * Licensed under the Apache License, Version 2.0 (the "License");
@ * you may not use this file except in compliance with the License.
@ * You may obtain a copy of the License at:
@ *
@ * http://www.apache.org/licenses/LICENSE-2.0
@ *
@ * Unless required by applicable law or agreed to in writing, software
@ * distributed under the License is distributed on an "AS IS" BASIS,
@ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
@ * See the License for the specific language governing permissions and
@ * limitations under the License.
@ *
@ *****************************************************************************
@ * Originally developed and contributed by Ittiam Systems Pvt. Ltd, Bangalore
@*/
@**
@**
@******************************************************************************
@*
@*
@* @brief
@* This file contains definitions of routines that compute distortion
@* between two macro/sub blocks of identical dimensions
@*
@* @author
@* Ittiam
@*
@* @par List of Functions:
@* - ime_compute_sad_16x16_a9q()
@* - ime_compute_sad_16x16_fast_a9q()
@* - ime_compute_sad_16x8_a9q()
@* - ime_compute_sad_16x16_ea8_a9q()
@* - ime_calculate_sad2_prog_a9q()
@* - ime_calculate_sad3_prog_a9q()
@* - ime_calculate_sad4_prog_a9q()
@* - ime_sub_pel_compute_sad_16x16_a9q()
@* - ime_compute_satqd_16x16_lumainter_a9q()
@* -
@* @remarks
@* None
@*
@*******************************************************************************
@
@**
@******************************************************************************
@*
@* @brief computes distortion (SAD) between 2 16x16 blocks (fast mode)
@*
@* @par Description
@* This functions computes SAD between 2 16x16 blocks. There is a provision
@* for early exit if the up-to computed SAD exceeds maximum allowed SAD. To
@* compute the distortion of the entire block set u4_max_sad to USHRT_MAX.
@*
@* @param[in] pu1_src
@* UWORD8 pointer to the source
@*
@* @param[out] pu1_dst
@* UWORD8 pointer to the destination
@*
@* @param[in] src_strd
@* integer source stride
@*
@* @param[in] dst_strd
@* integer destination stride
@*
@* @param[in] i4_max_sad
@* integer maximum allowed distortion
@*
@* @param[in] pi4_mb_distortion
@* integer evaluated sad
@*
@* @remarks
@*
@******************************************************************************
@*
.text
.p2align 2
.global ime_compute_sad_16x16_fast_a9q
ime_compute_sad_16x16_fast_a9q:
stmfd sp!, {r12, lr}
vpush {d8-d15}
lsl r2, r2, #1
lsl r3, r3, #1
@for bringing buffer2 into cache..., dummy load instructions
@LDR r12,[r1]
vld1.8 {d4, d5}, [r0], r2
vld1.8 {d6, d7}, [r1], r3
mov r12, #6
vld1.8 {d8, d9}, [r0], r2
vabdl.u8 q0, d6, d4
vabdl.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
loop_sad_16x16_fast:
vld1.8 {d4, d5}, [r0], r2
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vld1.8 {d6, d7}, [r1], r3
subs r12, #2
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d6, d4
vabal.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
bne loop_sad_16x16_fast
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vadd.i16 q0, q0, q1
vadd.i16 d0, d1, d0
vpop {d8-d15}
ldr r12, [sp, #12]
vpaddl.u16 d0, d0
vpaddl.u32 d0, d0
vshl.u32 d0, d0, #1
vst1.32 {d0[0]}, [r12]
ldmfd sp!, {r12, pc}
@**
@******************************************************************************
@*
@* @brief computes distortion (SAD) between 2 16x8 blocks
@*
@*
@* @par Description
@* This functions computes SAD between 2 16x8 blocks. There is a provision
@* for early exit if the up-to computed SAD exceeds maximum allowed SAD. To
@* compute the distortion of the entire block set u4_max_sad to USHRT_MAX.
@*
@* @param[in] pu1_src
@* UWORD8 pointer to the source
@*
@* @param[out] pu1_dst
@* UWORD8 pointer to the destination
@*
@* @param[in] src_strd
@* integer source stride
@*
@* @param[in] dst_strd
@* integer destination stride
@*
@* @param[in] u4_max_sad
@* integer maximum allowed distortion
@*
@* @param[in] pi4_mb_distortion
@* integer evaluated sad
@*
@* @remarks
@*
@******************************************************************************
@*
@
.global ime_compute_sad_16x8_a9q
ime_compute_sad_16x8_a9q:
stmfd sp!, {r12, lr}
@for bringing buffer2 into cache..., dummy load instructions
@LDR r12,[r1]
vld1.8 {d4, d5}, [r0], r2
vld1.8 {d6, d7}, [r1], r3
mov r12, #6
vpush {d8-d15}
vld1.8 {d8, d9}, [r0], r2
vabdl.u8 q0, d6, d4
vabdl.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
loop_sad_16x8:
vld1.8 {d4, d5}, [r0], r2
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vld1.8 {d6, d7}, [r1], r3
subs r12, #2
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d6, d4
vabal.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
bne loop_sad_16x8
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vadd.i16 q0, q0, q1
vadd.i16 d0, d1, d0
vpop {d8-d15}
ldr r12, [sp, #12]
vpaddl.u16 d0, d0
vpaddl.u32 d0, d0
vst1.32 {d0[0]}, [r12]
ldmfd sp!, {r12, pc}
@**
@******************************************************************************
@*
@* @brief computes distortion (SAD) between 2 16x16 blocks with early exit
@*
@* @par Description
@* This functions computes SAD between 2 16x16 blocks. There is a provision
@* for early exit if the up-to computed SAD exceeds maximum allowed SAD. To
@* compute the distortion of the entire block set u4_max_sad to USHRT_MAX.
@*
@* @param[in] pu1_src
@* UWORD8 pointer to the source
@*
@* @param[out] pu1_dst
@* UWORD8 pointer to the destination
@*
@* @param[in] src_strd
@* integer source stride
@*
@* @param[in] dst_strd
@* integer destination stride
@*
@* @param[in] i4_max_sad
@* integer maximum allowed distortion
@*
@* @param[in] pi4_mb_distortion
@* integer evaluated sad
@*
@* @remarks
@*
@******************************************************************************
@*
.global ime_compute_sad_16x16_ea8_a9q
ime_compute_sad_16x16_ea8_a9q:
stmfd sp!, {r5-r7, lr}
lsl r2, r2, #1
lsl r3, r3, #1
@for bringing buffer2 into cache..., dummy load instructions
@LDR r12,[r1]
vld1.8 {d4, d5}, [r0], r2
vld1.8 {d6, d7}, [r1], r3
mov r5, #6
ldrd r6, r7, [sp, #16]
vpush {d8-d15}
vld1.8 {d8, d9}, [r0], r2
vabdl.u8 q0, d6, d4
vabdl.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
@r6 = i4_max_sad, r7 = pi4_mb_distortion
loop_sad_16x16_ea8_1:
vld1.8 {d4, d5}, [r0], r2
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vld1.8 {d6, d7}, [r1], r3
subs r5, #2
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d6, d4
vabal.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
bne loop_sad_16x16_ea8_1
vabal.u8 q0, d10, d8
sub r0, r0, r2, lsl #3
vabal.u8 q1, d11, d9
sub r1, r1, r3, lsl #3
vadd.i16 q6, q0, q1
add r0, r0, r2, asr #1
vadd.i16 d12, d12, d13
add r1, r1, r3, asr #1
vpaddl.u16 d12, d12
vld1.8 {d4, d5}, [r0], r2
vld1.8 {d6, d7}, [r1], r3
vpaddl.u32 d12, d12
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d6, d4
vabal.u8 q1, d7, d5
vst1.32 {d12[0]}, [r7]
ldr r5, [r7]
cmp r5, r6
bgt end_func_16x16_ea8
vld1.8 {d10, d11}, [r1], r3
mov r5, #6
loop_sad_16x16_ea8_2:
vld1.8 {d4, d5}, [r0], r2
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vld1.8 {d6, d7}, [r1], r3
subs r5, #2
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d6, d4
vabal.u8 q1, d7, d5
vld1.8 {d10, d11}, [r1], r3
bne loop_sad_16x16_ea8_2
vabal.u8 q0, d10, d8
vabal.u8 q1, d11, d9
vadd.i16 q0, q0, q1
vadd.i16 d0, d1, d0
vpaddl.u16 d0, d0
vpaddl.u32 d0, d0
vst1.32 {d0[0]}, [r7]
end_func_16x16_ea8:
vpop {d8-d15}
ldmfd sp!, {r5-r7, pc}
@*
@//---------------------------------------------------------------------------
@// Function Name : Calculate_Mad2_prog()
@//
@// Detail Description : This function find the sad values of 4 Progressive MBs
@// at one shot
@//
@// Platform : CortexA8/NEON .
@//
@//-----------------------------------------------------------------------------
@*
.global ime_calculate_sad2_prog_a9q
ime_calculate_sad2_prog_a9q:
@ r0 = ref1 <UWORD8 *>
@ r1 = ref2 <UWORD8 *>
@ r2 = src <UWORD8 *>
@ r3 = RefBufferWidth <UWORD32>
@ stack = CurBufferWidth <UWORD32>, psad <UWORD32 *>
stmfd sp!, {r4-r5, lr}
ldr r4, [sp, #8] @ load src stride to r4
mov r5, #14
vpush {d8-d15}
@Row 1
vld1.8 {d0, d1}, [r2], r4 @ load src Row 1
vld1.8 {d2, d3}, [r0], r3 @ load ref1 Row 1
vld1.8 {d4, d5}, [r1], r3 @ load ref2 Row 1
@Row 2
vld1.8 {d6, d7}, [r2], r4 @ load src Row 2
vabdl.u8 q6, d2, d0
vabdl.u8 q7, d3, d1
vld1.8 {d8, d9}, [r0], r3 @ load ref1 Row 2
vabdl.u8 q8, d4, d0
vabdl.u8 q9, d5, d1
vld1.8 {d10, d11}, [r1], r3 @ load ref2 Row 2
loop_sad2_prog:
subs r5, #2
@Row 1
vld1.8 {d0, d1}, [r2], r4 @ load src Row 1
vabal.u8 q6, d8, d6
vabal.u8 q7, d9, d7
vld1.8 {d2, d3}, [r0], r3 @ load ref1 Row 1
vabal.u8 q8, d10, d6
vabal.u8 q9, d11, d7
vld1.8 {d4, d5}, [r1], r3 @ load ref2 Row 1
@Row 2
vld1.8 {d6, d7}, [r2], r4 @ load src Row 2
vabal.u8 q6, d2, d0
vabal.u8 q7, d3, d1
vld1.8 {d8, d9}, [r0], r3 @ load ref1 Row 2
vabal.u8 q8, d4, d0
vabal.u8 q9, d5, d1
vld1.8 {d10, d11}, [r1], r3 @ load ref2 Row 2
bne loop_sad2_prog
vabal.u8 q6, d8, d6
vabal.u8 q7, d9, d7
vabal.u8 q8, d10, d6
vabal.u8 q9, d11, d7
@ Compute SAD
vadd.u16 q6, q6, q7 @ Q6 : sad_ref1
vadd.u16 q8, q8, q9 @ Q8 : sad_ref2
vadd.u16 d12, d12, d13
ldr r5, [sp, #16] @ loading pi4_sad to r5
vadd.u16 d16, d16, d17
vpadd.u16 d12, d12, d16
vpaddl.u16 d12, d12
vst1.64 {d12}, [r5]!
vpop {d8-d15}
ldmfd sp!, {r4-r5, pc}
@*
@//---------------------------------------------------------------------------
@// Function Name : Calculate_Mad3_prog()
@//
@// Detail Description : This function find the sad values of 4 Progressive MBs
@// at one shot
@//
@// Platform : CortexA8/NEON .
@//
@//-----------------------------------------------------------------------------
@*
.global ime_calculate_sad3_prog_a9q
ime_calculate_sad3_prog_a9q:
@ r0 = ref1 <UWORD8 *>
@ r1 = ref2 <UWORD8 *>
@ r2 = ref3 <UWORD8 *>
@ r3 = src <UWORD8 *>
@ stack = RefBufferWidth <UWORD32>, CurBufferWidth <UWORD32>, psad <UWORD32 *>
stmfd sp!, {r4-r6, lr}
ldrd r4, r5, [sp, #16] @ load ref stride to r4, src stride to r5
mov r6, #14
vpush {d8-d15}
@Row 1
vld1.8 {d0, d1}, [r3], r5 @ load src Row 1
vld1.8 {d2, d3}, [r0], r4 @ load ref1 Row 1
vld1.8 {d4, d5}, [r1], r4 @ load ref2 Row 1
vabdl.u8 q8, d2, d0
vabdl.u8 q9, d3, d1
vld1.8 {d6, d7}, [r2], r4 @ load ref3 Row 1
vabdl.u8 q10, d4, d0
vabdl.u8 q11, d5, d1
@Row 2
vld1.8 {d8, d9}, [r3], r5 @ load src Row 1
vabdl.u8 q12, d6, d0
vabdl.u8 q13, d7, d1
vld1.8 {d10, d11}, [r0], r4 @ load ref1 Row 1
vld1.8 {d12, d13}, [r1], r4 @ load ref2 Row 1
vabal.u8 q8, d10, d8
vabal.u8 q9, d11, d9
vld1.8 {d14, d15}, [r2], r4 @ load ref3 Row 1
vabal.u8 q10, d12, d8
vabal.u8 q11, d13, d9
loop_sad3_prog:
@Row 1
vld1.8 {d0, d1}, [r3], r5 @ load src Row 1
vabal.u8 q12, d14, d8
vabal.u8 q13, d15, d9
vld1.8 {d2, d3}, [r0], r4 @ load ref1 Row 1
vld1.8 {d4, d5}, [r1], r4 @ load ref2 Row 1
vabal.u8 q8, d2, d0
vabal.u8 q9, d3, d1
vld1.8 {d6, d7}, [r2], r4 @ load ref3 Row 1
vabal.u8 q10, d4, d0
vabal.u8 q11, d5, d1
@Row 2
vld1.8 {d8, d9}, [r3], r5 @ load src Row 1
vabal.u8 q12, d6, d0
vabal.u8 q13, d7, d1
vld1.8 {d10, d11}, [r0], r4 @ load ref1 Row 1
subs r6, #2
vld1.8 {d12, d13}, [r1], r4 @ load ref2 Row 1
vabal.u8 q8, d10, d8
vabal.u8 q9, d11, d9
vld1.8 {d14, d15}, [r2], r4 @ load ref3 Row 1
vabal.u8 q10, d12, d8
vabal.u8 q11, d13, d9
bne loop_sad3_prog
vabal.u8 q12, d14, d8
vabal.u8 q13, d15, d9
@ Compute SAD
vadd.u16 q8, q8, q9 @ Q8 : sad_ref1
vadd.u16 q10, q10, q11 @ Q10 : sad_ref2
vadd.u16 q12, q12, q13 @ Q12 : sad_ref3
vadd.u16 d16, d16, d17
vadd.u16 d20, d20, d21
vadd.u16 d24, d24, d25
vpadd.u16 d16, d16, d20
vpadd.u16 d24, d24, d24
ldr r6, [sp, #24] @ loading pi4_sad to r6
vpaddl.u16 d16, d16
vpaddl.u16 d24, d24
vst1.64 {d16}, [r6]!
vst1.32 {d24[0]}, [r6]
vpop {d8-d15}
ldmfd sp!, {r4-r6, pc}
@**
@******************************************************************************
@*
@* @brief computes distortion (SAD) for sub-pel motion estimation
@*
@* @par Description
@* This functions computes SAD for all the 8 half pel points
@*
@* @param[out] pi4_sad
@* integer evaluated sad
@* pi4_sad[0] - half x
@* pi4_sad[1] - half x - 1
@* pi4_sad[2] - half y
@* pi4_sad[3] - half y - 1
@* pi4_sad[4] - half xy
@* pi4_sad[5] - half xy - 1
@* pi4_sad[6] - half xy - strd
@* pi4_sad[7] - half xy - 1 - strd
@*
@* @remarks
@*
@******************************************************************************
@*
.text
.p2align 2
.global ime_sub_pel_compute_sad_16x16_a9q
ime_sub_pel_compute_sad_16x16_a9q:
stmfd sp!, {r4-r11, lr} @store register values to stack
ldr r9, [sp, #36]
ldr r10, [sp, #40]
vpush {d8-d15}
sub r4, r1, #1 @ x left
sub r5, r2, r10 @ y top
sub r6, r3, #1 @ xy left
sub r7, r3, r10 @ xy top
sub r8, r7, #1 @ xy top-left
mov r11, #15
@for bringing buffer2 into cache..., dummy load instructions
@ LDR r12,[r1]
@ LDR r12,[sp,#12]
vld1.8 {d0, d1}, [r0], r9 @ src
vld1.8 {d2, d3}, [r5], r10 @ y top LOAD
vld1.8 {d4, d5}, [r7], r10 @ xy top LOAD
vld1.8 {d6, d7}, [r8], r10 @ xy top-left LOAD
vabdl.u8 q6, d2, d0 @ y top ABS1
vabdl.u8 q7, d4, d0 @ xy top ABS1
vld1.8 {d8, d9}, [r1], r10 @ x LOAD
vabdl.u8 q8, d6, d0 @ xy top-left ABS1
vabdl.u8 q9, d8, d0 @ x ABS1
vld1.8 {d10, d11}, [r4], r10 @ x left LOAD
vabal.u8 q6, d3, d1 @ y top ABS2
vabal.u8 q7, d5, d1 @ xy top ABS2
vld1.8 {d2, d3}, [r2], r10 @ y LOAD
vabal.u8 q8, d7, d1 @ xy top-left ABS2
vabal.u8 q9, d9, d1 @ x ABS2
vld1.8 {d4, d5}, [r3], r10 @ xy LOAD
vabdl.u8 q10, d10, d0 @ x left ABS1
vabdl.u8 q11, d2, d0 @ y ABS1
vld1.8 {d6, d7}, [r6], r10 @ xy left LOAD
vabdl.u8 q12, d4, d0 @ xy ABS1
vabdl.u8 q13, d6, d0 @ xy left ABS1
loop_sub_pel_16x16:
vabal.u8 q10, d11, d1 @ x left ABS2
vabal.u8 q11, d3, d1 @ y ABS2
subs r11, #1
vabal.u8 q12, d5, d1 @ xy ABS2
vabal.u8 q13, d7, d1 @ xy left ABS2
vld1.8 {d0, d1}, [r0], r9 @ src
vabal.u8 q6, d2, d0 @ y top ABS1
vabal.u8 q7, d4, d0 @ xy top ABS1
vld1.8 {d8, d9}, [r1], r10 @ x LOAD
vabal.u8 q8, d6, d0 @ xy top-left ABS1
vabal.u8 q9, d8, d0 @ x ABS1
vld1.8 {d10, d11}, [r4], r10 @ x left LOAD
vabal.u8 q6, d3, d1 @ y top ABS2
vabal.u8 q7, d5, d1 @ xy top ABS2
vld1.8 {d2, d3}, [r2], r10 @ y LOAD
vabal.u8 q8, d7, d1 @ xy top-left ABS2
vabal.u8 q9, d9, d1 @ x ABS2
vld1.8 {d4, d5}, [r3], r10 @ xy LOAD
vabal.u8 q10, d10, d0 @ x left ABS1
vabal.u8 q11, d2, d0 @ y ABS1
vld1.8 {d6, d7}, [r6], r10 @ xy left LOAD
vabal.u8 q12, d4, d0 @ xy ABS1
vabal.u8 q13, d6, d0 @ xy left ABS1
bne loop_sub_pel_16x16
vabal.u8 q10, d11, d1 @ x left ABS2
vabal.u8 q11, d3, d1 @ y ABS2
vabal.u8 q12, d5, d1 @ xy ABS2
vabal.u8 q13, d7, d1 @ xy left ABS2
vadd.i16 d0, d18, d19 @ x
vadd.i16 d3, d12, d13 @ y top
vadd.i16 d6, d14, d15 @ xy top
vadd.i16 d5, d26, d27 @ xy left
vadd.i16 d1, d20, d21 @ x left
vadd.i16 d2, d22, d23 @ y
vadd.i16 d4, d24, d25 @ xy
vadd.i16 d7, d16, d17 @ xy top left
vpadd.i16 d0, d0, d1
vpadd.i16 d2, d2, d3
vpadd.i16 d4, d4, d5
vpadd.i16 d6, d6, d7
vpaddl.u16 d0, d0
vpaddl.u16 d2, d2
vpop {d8-d15}
ldr r11, [sp, #44]
vpaddl.u16 d4, d4
vpaddl.u16 d6, d6
vst1.32 {d0}, [r11]!
vst1.32 {d2}, [r11]!
vst1.32 {d4}, [r11]!
vst1.32 {d6}, [r11]!
ldmfd sp!, {r4-r11, pc} @Restoring registers from stack
@**
@******************************************************************************
@*
@* @brief computes distortion (SAD) between 2 16x16 blocks
@*
@* @par Description
@* This functions computes SAD between 2 16x16 blocks. There is a provision
@* for early exit if the up-to computed SAD exceeds maximum allowed SAD. To
@* compute the distortion of the entire block set u4_max_sad to USHRT_MAX.
@*
@* @param[in] pu1_src
@* UWORD8 pointer to the source
@*
@* @param[out] pu1_dst
@* UWORD8 pointer to the destination
@*
@* @param[in] src_strd
@* integer source stride
@*
@* @param[in] dst_strd
@* integer destination stride
@*
@* @param[in] i4_max_sad
@* integer maximum allowed distortion
@*
@* @param[in] pi4_mb_distortion
@* integer evaluated sad
@*
@* @remarks
@*
@******************************************************************************
@*
.text
.p2align 2
.global ime_compute_sad_16x16_a9q
ime_compute_sad_16x16_a9q:
@STMFD sp!,{r12,lr}
stmfd sp!, {r12, r14} @store register values to stack
@for bringing buffer2 into cache..., dummy load instructions
@ LDR r12,[r1]
@ LDR r12,[sp,#12]
vld1.8 {d4, d5}, [r0], r2
vld1.8 {d6, d7}, [r1], r3
vpush {d8-d15}
mov r12, #14
vld1.8 {d8, d9}, [r0], r2
vabdl.u8 q0, d4, d6
vld1.8 {d10, d11}, [r1], r3
vabdl.u8 q1, d5, d7
loop_sad_16x16:
vld1.8 {d4, d5}, [r0], r2
vabal.u8 q0, d8, d10
vld1.8 {d6, d7}, [r1], r3
vabal.u8 q1, d9, d11
vld1.8 {d8, d9}, [r0], r2
vabal.u8 q0, d4, d6
subs r12, #2
vld1.8 {d10, d11}, [r1], r3
vabal.u8 q1, d5, d7
bne loop_sad_16x16
vabal.u8 q0, d8, d10
vabal.u8 q1, d9, d11
vadd.i16 q0, q0, q1
vadd.i16 d0, d1, d0
vpop {d8-d15}
ldr r12, [sp, #12]
vpaddl.u16 d0, d0
vpaddl.u32 d0, d0
vst1.32 {d0[0]}, [r12]
ldmfd sp!, {r12, pc} @Restoring registers from stack
@*
@//---------------------------------------------------------------------------
@// Function Name : Calculate_Mad4_prog()
@//
@// Detail Description : This function find the sad values of 4 Progressive MBs
@// at one shot
@//
@// Platform : CortexA8/NEON .
@//
@//-----------------------------------------------------------------------------
@*
.global ime_calculate_sad4_prog_a9q
ime_calculate_sad4_prog_a9q:
@ r0 = temp_frame <UWORD8 *>
@ r1 = buffer_ptr <UWORD8 *>
@ r2 = RefBufferWidth <UWORD32>
@ r3 = CurBufferWidth <UWORD32>
@ stack = psad <UWORD32 *> {at 0x34}
stmfd sp!, {r4-r7, lr}
@UWORD8 *left_ptr = temp_frame - 1;
@UWORD8 *right_ptr = temp_frame + 1;
@UWORD8 *top_ptr = temp_frame - RefBufferWidth;
@UWORD8 *bot_ptr = temp_frame + RefBufferWidth;
mov r7, #14
sub r4, r0, #0x01 @r4 = left_ptr
add r5, r0, #0x1 @r5 = right_ptr
sub r6, r0, r2 @r6 = top_ptr
add r0, r0, r2 @r0 = bot_ptr
@r1 = buffer_ptr
vpush {d8-d15}
@D0:D1 : buffer
@D2:D3 : top
@D4:D5 : left
@D6:D7 : right
@D8:D9 : bottom
@Row 1
vld1.8 {d0, d1}, [r1], r3 @ load src Row 1
vld1.8 {d2, d3}, [r6], r2 @ load top Row 1
vld1.8 {d4, d5}, [r4], r2 @ load left Row 1
vabdl.u8 q5, d2, d0
vld1.8 {d6, d7}, [r5], r2 @ load right Row 1
vabdl.u8 q6, d3, d1
vabdl.u8 q7, d0, d4
vld1.8 {d8, d9}, [r0], r2 @ load bottom Row 1
vabdl.u8 q8, d1, d5
@Row 2
vabdl.u8 q9, d0, d6
vld1.8 {d26, d27}, [r1], r3 @ load src Row 2
vabdl.u8 q10, d1, d7
vabdl.u8 q11, d0, d8
vld1.8 {d2, d3}, [r6], r2 @ load top Row 2
vabdl.u8 q12, d1, d9
loop_sad4_prog:
vabal.u8 q5, d26, d2
vld1.8 {d4, d5}, [r4], r2 @ load left Row 2
vabal.u8 q6, d27, d3
vabal.u8 q7, d26, d4
vld1.8 {d6, d7}, [r5], r2 @ load right Row 2
vabal.u8 q8, d27, d5
vabal.u8 q9, d26, d6
vld1.8 {d8, d9}, [r0], r2 @ load bottom Row 2
vabal.u8 q10, d27, d7
@Row 1
vabal.u8 q11, d26, d8
vld1.8 {d0, d1}, [r1], r3 @ load src Row 1
vabal.u8 q12, d27, d9
vld1.8 {d2, d3}, [r6], r2 @ load top Row 1
subs r7, #2
vld1.8 {d4, d5}, [r4], r2 @ load left Row 1
vabal.u8 q5, d0, d2
vld1.8 {d6, d7}, [r5], r2 @ load right Row 1
vabal.u8 q6, d1, d3
vabal.u8 q7, d0, d4
vld1.8 {d8, d9}, [r0], r2 @ load bottom Row 1
vabal.u8 q8, d1, d5
@Row 2
vabal.u8 q9, d0, d6
vld1.8 {d26, d27}, [r1], r3 @ load src Row 2
vabal.u8 q10, d1, d7
vabal.u8 q11, d0, d8
vld1.8 {d2, d3}, [r6], r2 @ load top Row 2
vabal.u8 q12, d1, d9
bne loop_sad4_prog
vabal.u8 q5, d26, d2
vld1.8 {d4, d5}, [r4], r2 @ load left Row 2
vabal.u8 q6, d27, d3
vabal.u8 q7, d26, d4
vld1.8 {d6, d7}, [r5], r2 @ load right Row 2
vabal.u8 q8, d27, d5
vabal.u8 q9, d26, d6
vld1.8 {d8, d9}, [r0], r2 @ load bottom Row 2
vabal.u8 q10, d27, d7
vabal.u8 q11, d26, d8
vabal.u8 q12, d27, d9
@;Q5:Q6 : sad_top
@;Q7:Q8 : sad_left
@;Q9:Q10 : sad_right
@;Q11:Q12 : sad_bot
vadd.u16 q5, q5, q6
vadd.u16 q7, q7, q8
vadd.u16 q9, q9, q10
vadd.u16 q11, q11, q12
@; Free :-
@; Q6,Q8,Q10,Q12
@;Q5 -> D10:D11
@;Q7 -> D14:D15
@;Q9 -> D18:D19
@;Q11 -> D22:D23
vadd.u16 d10, d10, d11
vadd.u16 d14, d14, d15
vadd.u16 d18, d18, d19
vadd.u16 d22, d22, d23
@;D10 : sad_top
@;D14 : sad_left
@;D18 : sad_right
@;D22 : sad_bot
vpaddl.u16 d11, d10
vpaddl.u16 d15, d14
vpaddl.u16 d19, d18
vpaddl.u16 d23, d22
@;D11 : sad_top
@;D15 : sad_left
@;D19 : sad_right
@;D23 : sad_bot
vpaddl.u32 d10, d11
vpaddl.u32 d22, d23
vpaddl.u32 d14, d15
vpaddl.u32 d18, d19
@;D10 : sad_top
@;D14 : sad_left
@;D18 : sad_right
@;D22 : sad_bot
ldr r4, [sp, #84] @;Can be rearranged
vsli.64 d10, d22, #32
vsli.64 d14, d18, #32
vst1.64 {d14}, [r4]!
vst1.64 {d10}, [r4]!
vpop {d8-d15}
ldmfd sp!, {r4-r7, pc}
@*****************************************************************************
@*
@* Function Name : ime_compute_satqd_16x16_lumainter_a9
@* Description : This fucntion computes SAD for a 16x16 block.
@ : It also computes if any 4x4 block will have a nonzero coefficent after transform and quant
@
@ Arguments : R0 :pointer to src buffer
@ R1 :pointer to est buffer
@ R2 :source stride
@ R3 :est stride
@ STACk :Threshold,distotion,is_nonzero
@*
@* Values Returned : NONE
@*
@* Register Usage : R0-R11
@* Stack Usage :
@* Cycles : Around
@* Interruptiaility : Interruptable
@*
@* Known Limitations
@* \Assumptions :
@*
@* Revision History :
@* DD MM YYYY Author(s) Changes
@* 14 04 2014 Harinarayanan K K First version
@*
@*****************************************************************************
.global ime_compute_satqd_16x16_lumainter_a9q
ime_compute_satqd_16x16_lumainter_a9q:
@R0 :pointer to src buffer
@R1 :pointer to est buffer
@R2 :Source stride
@R3 :Pred stride
@R4 :Threshold pointer
@R5 :Distortion,ie SAD
@R6 :is nonzero
push {r4-r12, lr} @push all the variables first
@ADD SP,SP,#40 ;decrement stack pointer,to accomodate two variables
ldr r4, [sp, #40] @load the threshold address
vpush {d8-d15}
mov r8, #8 @Number of 4x8 blocks to be processed
mov r10, #0 @Sad
mov r7, #0 @Nonzero info
@----------------------------------------------------
vld1.u8 d30, [r0], r2 @I load 8 pix src row 1
vld1.u8 d31, [r1], r3 @I load 8 pix pred row 1
vld1.u8 d28, [r0], r2 @I load 8 pix src row 2
vld1.u8 d29, [r1], r3 @I load 8 pix pred row 2
vld1.u8 d26, [r0], r2 @I load 8 pix src row 3
vabdl.u8 q0, d30, d31 @I Abs diff r1 blk 12
vld1.u8 d27, [r1], r3 @I load 8 pix pred row 3
vld1.u8 d24, [r0], r2 @I load 8 pix src row 4
vld1.u8 d25, [r1], r3 @I load 8 pix pred row 4
vabdl.u8 q1, d28, d29 @I Abs diff r1 blk 12
vld1.u16 {q11}, [r4] @I load the threhold
vabdl.u8 q2, d26, d27 @I Abs diff r1 blk 12
vabdl.u8 q3, d24, d25 @I Abs diff r1 blk 12
core_loop:
@S1 S2 S3 S4 A1 A2 A3 A4
@S5 S6 S7 S8 A5 A6 A7 A8
@S9 S10 S11 S12 A9 A10 A11 A12
@S13 S14 S15 S16 A13 A14 A15 A16
ands r11, r8, #1 @II See if we are at even or odd block
vadd.u16 q4 , q0, q3 @I Add r1 r4
lsl r11, r2, #2 @II Move back src 4 rows
subeq r0, r0, r11 @II Move back src 4 rows if we are at even block
vadd.u16 q5 , q1, q2 @I Add r2 r3
addeq r0, r0, #8 @II Move src 8 cols forward if we are at even block
lsl r11, r3, #2 @II Move back pred 4 rows
vtrn.16 d8 , d10 @I trnspse 1
subeq r1, r1, r11 @II Move back pred 4 rows if we are at even block
addeq r1, r1, #8 @II Move pred 8 cols forward if we are at even block
vtrn.16 d9 , d11 @I trnspse 2
subne r0, r0, #8 @II Src 8clos back for odd rows
subne r1, r1, #8 @II Pred 8 cols back for odd rows
vtrn.32 d10, d11 @I trnspse 4
vtrn.32 d8 , d9 @I trnspse 3
vswp d10, d11 @I rearrange so that the q4 and q5 add properly
@D8 S1 S4 A1 A4
@D9 S2 S3 A2 A3
@D11 S1 S4 A1 A4
@D10 S2 S3 A2 A3
vadd.s16 q6, q4, q5 @I Get s1 s4
vld1.u8 d30, [r0], r2 @II load first 8 pix src row 1
vtrn.s16 d12, d13 @I Get s2 s3
@D12 S1 S4 A1 A4
@D13 S2 S3 A2 A3
vshl.s16 q7, q6 , #1 @I si = si<<1
vld1.u8 d31, [r1], r3 @II load first 8 pix pred row 1
vpadd.s16 d16, d12, d13 @I (s1 + s4) (s2 + s3)
vld1.u8 d28, [r0], r2 @II load first 8 pix src row 2
@ D16 S14 A14 S23 A23
vrev32.16 d0, d16 @I
vuzp.s16 d16, d0 @I
@D16 S14 S23 A14 A23
vadd.s16 d17, d12, d13 @I (s1 + s2) (s3 + s4)
vld1.u8 d29, [r1], r3 @II load first 8 pix pred row 2
@D17 S12 S34 A12 A34
vrev32.16 q9, q7 @I Rearrange si's
@Q9 Z4,Z1,Y4,Y1,Z3,Z2,Y3,Y2
@D12 S1 S4 A1 A4
@D19 Z3 Z2 Y3 Y2
vsub.s16 d8, d12, d19 @I (s1 - (s3<<1)) (s4 - (s2<<1))
vld1.u8 d26, [r0], r2 @II load first 8 pix src row 3
@D13 S2 S3 A2 A3
@D18 Z4 Z1 Y4 Y1
vsub.s16 d9, d13, d18 @I (s2 - (s4<<1)) (s3 - (s1<<1))
vld1.u8 d27, [r1], r3 @II load first 8 pix pred row 3
@Q10 S8 S5 A8 A5 S7 S4 A7 A4
@D16 S14 S23 A14 A23
vpadd.s16 d10, d16, d17 @I Get sad by adding s1 s2 s3 s4
vld1.u8 d24, [r0], r2 @II load first 8 pix src row 4
@D22 SAD1 SAD2 junk junk
@Q8 S2 S1 A2 A1 S6 S3 A6 A3
@Q10 S8 S5 A8 A5 S7 S4 A7 A4
vtrn.32 q8, q4 @I Rearrange to make ls of each block togather
@Q8 S2 S1 S8 S5 S6 S3 S7 S4
@Q10 A2 A1 A8 A5 A6 A3 A7 A4
ldrh r11, [r4, #16] @I Load the threshold for DC val blk 1
vdup.s16 q6, d10[0] @I Get the sad blk 1
vabdl.u8 q0, d30, d31 @II Abs diff r1 blk 12
vshl.s16 q7, q6, #1 @I sad_2 = sad_1<<1
vmov.s16 r9, d10[0] @I Get the sad for block 1
vsub.s16 q9, q7, q8 @I Add to the lss
vmov.s16 r5, d10[1] @I Get the sad for block 2
vcle.s16 q7, q11, q9 @I Add to the lss
vld1.u8 d25, [r1], r3 @II load first 8 pix pred row 4
vdup.s16 q15, d10[1] @I Get the sad blk 1
vabdl.u8 q1, d28, d29 @II Abs diff r1 blk 12
vshl.s16 q14, q15, #1 @I sad_2 = sad_1<<1
vsub.s16 q3, q14, q4 @I Add to the lss
vcle.s16 q15, q11, q3 @I Add to the lss
ADD R10, R10, R9 @I Add to the global sad blk 1
vtrn.u8 q15, q7 @I get all comparison bits to one reg
vabdl.u8 q2, d26, d27 @II Abs diff r1 blk 12
ADD R10, R10, R5 @I Add to the global sad blk 2
vshr.u8 q14, q15, #7 @I Shift the bits so that no overflow occurs
cmp r11, r9
movle r7, #0xf @I If not met mark it by mvoing non zero val to R7 blk 1 ;I Compare with threshold blk 1
vadd.u8 d28, d28, d29 @I Add the bits
cmp r11, r5 @I Compare with threshold blk 2
movle r7, #0xf @I If not met mark it by mvoing non zero val to R7 blk 2
vpadd.u8 d28, d28, d29 @I Add the bits
vmov.u32 r11, d28[0] @I Since a set bit now represents a unstatisofrd contifon store it in r11
vabdl.u8 q3, d24, d25 @II Abs diff r1 blk 12
orr r7, r7, r11 @I get the guy to r11
sub r8, r8, #1 @I Decremrnt block count
cmp r7, #0 @I If we have atlest one non zero block
bne compute_sad_only @I if a non zero block is der,From now on compute sad only
cmp r8, #1 @I See if we are at the last block
bne core_loop @I If the blocks are zero, lets continue the satdq
@EPILOUGE for core loop
@S1 S2 S3 S4 A1 A2 A3 A4
@S5 S6 S7 S8 A5 A6 A7 A8
@S9 S10 S11 S12 A9 A10 A11 A12
@S13 S14 S15 S16 A13 A14 A15 A16
vadd.u16 q4 , q0, q3 @Add r1 r4
vadd.u16 q5 , q1, q2 @Add r2 r3
@D8 S1 S2 S2 S1
@D10 S4 S3 S3 S4
@D9 A1 A2 A2 A1
@D11 A4 A3 A3 A4
vtrn.16 d8 , d10 @I trnspse 1
vtrn.16 d9 , d11 @I trnspse 2
vtrn.32 d8 , d9 @I trnspse 3
vtrn.32 d10, d11 @I trnspse 4
vswp d10, d11 @I rearrange so that the q4 and q5 add properly
@D8 S1 S4 A1 A4
@D9 S2 S3 A2 A3
@D11 S1 S4 A1 A4
@D10 S2 S3 A2 A3
vadd.s16 q6, q4, q5 @Get s1 s4
vtrn.s16 d12, d13 @Get s2 s3
@D12 S1 S4 A1 A4
@D13 S2 S3 A2 A3
vshl.s16 q7, q6 , #1 @si = si<<1
vmov.s16 r9, d10[0] @Get the sad for block 1
vpadd.s16 d16, d12, d13 @(s1 + s4) (s2 + s3)
vmov.s16 r5, d10[1] @Get the sad for block 2
@D16 S14 A14 S23 A23
vrev32.16 d30, d16 @
vuzp.s16 d16, d30 @
@D16 S14 S23 A14 A23
vadd.s16 d17, d12, d13 @(s1 + s2) (s3 + s4)
@D17 S12 S34 A12 A34
vrev32.16 q9, q7 @Rearrange si's
@Q9 Z4,Z1,Y4,Y1,Z3,Z2,Y3,Y2
@D12 S1 S4 A1 A4
@D19 Z3 Z2 Y3 Y2
vsub.s16 d8, d12, d19 @(s1 - (s3<<1)) (s4 - (s2<<1))
@D13 S2 S3 A2 A3
@D18 Z4 Z1 Y4 Y1
vsub.s16 d9, d13, d18 @(s2 - (s4<<1)) (s3 - (s1<<1))
@Q10 S8 S5 A8 A5 S7 S4 A7 A4
@D16 S14 S23 A14 A23
vpadd.s16 d10, d16, d17 @I Get sad by adding s1 s2 s3 s4
@D22 SAD1 SAD2 junk junk
vmov.u16 r9, d10[0] @Get the sad for block 1
vmov.u16 r5, d10[1] @Get the sad for block 2
@Q8 S2 S1 A2 A1 S6 S3 A6 A3
@Q10 S8 S5 A8 A5 S7 S4 A7 A4
ldrh r11, [r4, #16] @Load the threshold for DC val blk 1
vtrn.32 q8, q4 @Rearrange to make ls of each block togather
ADD R10, R10, R9 @Add to the global sad blk 1
@Q8 S2 S1 S8 S5 S6 S3 S7 S4
@Q10 A2 A1 A8 A5 A6 A3 A7 A4
vld1.u16 {q11}, [r4] @load the threhold
ADD R10, R10, R5 @Add to the global sad blk 2
vdup.u16 q6, d10[0] @Get the sad blk 1
cmp r11, r9 @Compare with threshold blk 1
vshl.u16 q7, q6, #1 @sad_2 = sad_1<<1
vsub.s16 q9, q7, q8 @Add to the lss
vcle.s16 q15, q11, q9 @Add to the lss
movle r7, #0xf @If not met mark it by mvoing non zero val to R7 blk 1
cmp r11, r5 @Compare with threshold blk 2
vdup.u16 q14, d10[1] @Get the sad blk 1
vshl.u16 q13, q14, #1 @sad_2 = sad_1<<1
vsub.s16 q12, q13, q4 @Add to the lss
vcle.s16 q14, q11, q12 @Add to the lss
movle r7, #0xf @If not met mark it by mvoing non zero val to R7 blk 2
vtrn.u8 q14, q15 @get all comparison bits to one reg
vshr.u8 q14, q14, #7 @Shift the bits so that no overflow occurs
vadd.u8 d28, d28, d29 @Add the bits
vpadd.u8 d28, d28, d29 @Add the bits
vmov.u32 r11, d28[0] @Since a set bit now represents a unstatisofrd contifon store it in r11
orr r7, r7, r11 @get the guy to r11
b funcend_sad_16x16 @Since all blocks ar processed nw, got to end
compute_sad_only: @This block computes SAD only, so will be lighter
@IT will start processign at n odd block
@It will compute sad for odd blok,
@and then for two blocks at a time
@The counter is r7, hence r7 blocks will be processed
and r11, r8, #1 @Get the last bit of counter
cmp r11, #0 @See if we are at even or odd block
@iif the blk is even we just have to set the pointer to the
@start of current row
lsleq r11, r2, #2 @I Move back src 4 rows
subeq r0, r0, r11 @I Move back src 4 rows if we are at even block
lsleq r11, r3, #2 @I Move back pred 4 rows
subeq r1, r1, r11 @I Move back pred 4 rows if we are at even block
@ADDEQ R8,R8,#2 ;Inc counter
beq skip_odd_blk @If the blk is odd we have to compute sad
vadd.u16 q4, q0, q1 @Add SAD of row1 and row2
vadd.u16 q5, q2, q3 @Add SAD of row3 and row4
vadd.u16 q6, q4, q5 @Add SAD of row 1-4
vadd.u16 d14, d12, d13 @Add Blk1 and blk2
vpadd.u16 d16, d14, d15 @Add col 1-2 and 3-4
vpadd.u16 d18, d16, d17 @Add col 12-34
vmov.u16 r9, d18[0] @Move sad to arm
ADD R10, R10, R9 @Add to the global sad
sub r8, r8, #1 @Dec counter
cmp r8, #0 @See if we processed last block
beq funcend_sad_16x16 @if lprocessed last block goto end of func
sub r0, r0, #8 @Since we processed od block move back src by 8 cols
sub r1, r1, #8 @Since we processed od block move back pred by 8 cols
skip_odd_blk:
vmov.s16 q0, #0 @Initialize the accumulator
vmov.s16 q1, #0 @Initialize the accumulator
vld1.u8 {q15}, [r0], r2 @load src r1
vld1.u8 {q14}, [r1], r3 @load pred r1
vld1.u8 {q13}, [r0], r2 @load src r2
vld1.u8 {q12}, [r1], r3 @load pred r2
vld1.u8 {q11}, [r0], r2 @load src r3
vld1.u8 {q10}, [r1], r3 @load pred r2
vld1.u8 {q9}, [r0], r2 @load src r4
vld1.u8 {q8}, [r1], r3 @load pred r4
cmp r8, #2
beq sad_epilouge
sad_loop:
vabal.u8 q0, d30, d28 @I accumulate Abs diff R1
vabal.u8 q1, d31, d29 @I accumulate Abs diff R1
vld1.u8 {q15}, [r0], r2 @II load r1 src
vabal.u8 q0, d26, d24 @I accumulate Abs diff R2
vld1.u8 {q14}, [r1], r3 @II load r1 pred
vabal.u8 q1, d27, d25 @I accumulate Abs diff R2
vld1.u8 {q13}, [r0], r2 @II load r3 src
vabal.u8 q0, d22, d20 @I accumulate Abs diff R3
vld1.u8 {q12}, [r1], r3 @II load r2 pred
vabal.u8 q1, d23, d21 @I accumulate Abs diff R3
vld1.u8 {q11}, [r0], r2 @II load r3 src
vabal.u8 q0, d18, d16 @I accumulate Abs diff R4
sub r8, r8, #2 @Since we processe 16 pix @a time, dec by 2
vld1.u8 {q10}, [r1], r3 @II load r3 pred
vabal.u8 q1, d19, d17 @I accumulate Abs diff R4
cmp r8, #2 @Check if last loop
vld1.u8 {q9}, [r0], r2 @II load r4 src
vld1.u8 {q8}, [r1], r3 @II load r4 pred
bne sad_loop @Go back to SAD computation
sad_epilouge:
vabal.u8 q0, d30, d28 @Accumulate Abs diff R1
vabal.u8 q1, d31, d29 @Accumulate Abs diff R1
vabal.u8 q0, d26, d24 @Accumulate Abs diff R2
vabal.u8 q1, d27, d25 @Accumulate Abs diff R2
vabal.u8 q0, d22, d20 @Accumulate Abs diff R3
vabal.u8 q1, d23, d21 @Aaccumulate Abs diff R3
vabal.u8 q0, d18, d16 @Accumulate Abs diff R4
vabal.u8 q1, d19, d17 @Accumulate Abs diff R4
vadd.u16 q2, q0, q1 @ADD two accumulators
vadd.u16 d6, d4, d5 @Add two blk sad
vpadd.u16 d8, d6, d7 @Add col 1-2 and 3-4 sad
vpadd.u16 d10, d8, d9 @Add col 12-34 sad
vmov.u16 r9, d10[0] @move SAD to ARM
ADD R10, R10, R9 @Add to the global sad
funcend_sad_16x16: @End of fucntion process
vpop {d8-d15}
ldr r5, [sp, #44]
ldr r6, [sp, #48]
str r7, [r6] @Store the is zero reg
str r10, [r5] @Store sad
@SUB SP,SP,#40
pop {r4-r12, pc}