@/******************************************************************************
@ *
@ * 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
@*/
@**
@*******************************************************************************
@* @file
@* ih264_resi_trans_quant_a9.s
@*
@* @brief
@* Contains function definitions for residual and forward trans
@*
@* @author
@* Ittiam
@*
@* @par List of Functions:
@* ih264_resi_trans_quant_4x4_a9
@* ih264_resi_trans_quant_8x8_a9
@* ih264_resi_trans_quant_chroma_4x4_a9
@* ih264_hadamard_quant_4x4_a9
@* ih264_hadamard_quant_2x2_uv_a9
@*
@* @remarks
@* None
@*
@*******************************************************************************
.text
.p2align 2
@*****************************************************************************
@*
@* Function Name : ih264_resi_trans_quant_4x4_a9
@* Description : This function does cf4 of H264
@*
@* Arguments : R0 :pointer to src buffer
@ R1 :pointer to pred buffer
@ R2 :pointer to dst buffer
@ R3 :source stride
@ STACK : pred stride,
@ dst stride,
@ pointer to scaling matrix,
@ pointer to threshold matrix,
@ qbits,
@ rounding factor,
@ pointer to store nnz
@ pointer to store non quantized dc value
@ Values Returned : NONE
@
@ Register Usage :
@ Stack Usage : 40 bytes
@ Cycles : Around
@ Interruptiaility : Interruptable
@
@ Known Limitations
@ \Assumptions :
@
@ Revision History :
@ DD MM YYYY Author(s) Changes
@ 1 12 2013 100633 First version
@ 20 1 2014 100633 Changes the API, Optimization
@
@*****************************************************************************
.global ih264_resi_trans_quant_4x4_a9
ih264_resi_trans_quant_4x4_a9:
@R0 :pointer to src buffer
@R1 :pointer to pred buffer
@R2 :pointer to dst buffer
@R3 :Source stride
@STACk :pred stride
@ :scale matirx,
@ :threshold matrix
@ :qbits
@ :round factor
@ :nnz
push {r4-r12, lr} @push all the variables first
add r11, sp, #40 @decrement stack pointer,to accomodate two variables
ldmfd r11, {r4-r10} @load the strides into registers
@R0 :pointer to src buffer
@R1 :pointer to pred buffer
@R2 :pointer to dst buffer
@R3 :Source stride
@R4 :Pred stride
@R5 :scale matirx,
@R6 :threshold matrix
@R7 :qbits
@R8 :round factor
@R9 :nnz
vpush {d8-d15}
mov r11, #0
sub r7, r11, r7 @Negate the qbit value for usiing LSL
@------------Fucntion Loading done----------------;
vld1.u8 d30, [r0], r3 @load first 8 pix src row 1
vld1.u8 d31, [r1], r4 @load first 8 pix pred row 1
vld1.u8 d28, [r0], r3 @load first 8 pix src row 2
vld1.u8 d29, [r1], r4 @load first 8 pix pred row 2
vld1.u8 d26, [r0], r3 @load first 8 pix src row 3
vld1.u8 d27, [r1], r4 @load first 8 pix pred row 3
vsubl.u8 q0, d30, d31 @find residue row 1
vld1.u8 d24, [r0], r3 @load first 8 pix src row 4
vld1.u8 d25, [r1], r4 @load first 8 pix pred row 4
vsubl.u8 q1, d28, d29 @find residue row 2
vsubl.u8 q2, d26, d27 @find residue row 3
vsubl.u8 q3, d24, d25 @find residue row 4
vtrn.16 d0, d2 @T12
vtrn.16 d4, d6 @T23
vtrn.32 d0, d4 @T13
vtrn.32 d2, d6 @T14
vadd.s16 d8 , d0, d6 @x0 = x4+x7
vadd.s16 d9 , d2, d4 @x1 = x5+x6
vsub.s16 d10, d2, d4 @x2 = x5-x6
vsub.s16 d11, d0, d6 @x3 = x4-x7
vshl.s16 d12, d10, #1 @U_SHIFT(x2,1,shft)
vshl.s16 d13, d11, #1 @U_SHIFT(x3,1,shft)
vadd.s16 d14, d8, d9 @x4 = x0 + x1;
vsub.s16 d16, d8, d9 @x6 = x0 - x1;
vadd.s16 d15, d13, d10 @x5 = U_SHIFT(x3,1,shft) + x2;
vsub.s16 d17, d11, d12 @x7 = x3 - U_SHIFT(x2,1,shft);
@taking transpose again so as to make do vert transform
vtrn.16 d14, d15 @T12
vtrn.16 d16, d17 @T23
vtrn.32 d14, d16 @T13
vtrn.32 d15, d17 @T24
@let us do vertical transform
@same code as horiz
vadd.s16 d18, d14, d17 @x0 = x4+x7
vadd.s16 d19, d15, d16 @x1 = x5+x6
vsub.s16 d20, d15, d16 @x2 = x5-x6
vsub.s16 d21, d14, d17 @x3 = x4-x7
vshl.s16 d22, d20, #1 @U_SHIFT(x2,1,shft)
vshl.s16 d23, d21, #1 @U_SHIFT(x3,1,shft)
vdup.s32 q4, r8 @Load rounding value row 1
vadd.s16 d24, d18, d19 @x5 = x0 + x1;
vsub.s16 d26, d18, d19 @x7 = x0 - x1;
vadd.s16 d25, d23, d20 @x6 = U_SHIFT(x3,1,shft) + x2;
vsub.s16 d27, d21, d22 @x8 = x3 - U_SHIFT(x2,1,shft);
vdup.s32 q10, r7 @Load qbit values
vst1.s16 d24[0], [r10] @Store the dc value to alternate dc sddress
@core tranform is done for 4x8 block 1
vld1.s16 {q14-q15}, [r5] @load the scaling values
vabs.s16 q0, q12 @Abs val of row 1 blk 1
vabs.s16 q1, q13 @Abs val of row 2 blk 1
vmov.s32 q5, q4 @copy round fact for row 2
vmov.s32 q6, q4 @copy round fact for row 2
vclt.s16 q2, q12, #0 @Get the sign of row 1 blk 1
vmov.s32 q7, q4 @copy round fact for row 2
vclt.s16 q3, q13, #0 @Get the sign of row 2 blk 1
vmlal.s16 q4, d0, d28 @Multiply and add row 1
vmlal.s16 q5, d1, d29 @Multiply and add row 2
vmlal.s16 q6, d2, d30 @Multiply and add row 3
vmlal.s16 q7, d3, d31 @Multiply and add row 4
vshl.s32 q11, q4, q10 @Shift row 1
vshl.s32 q12, q5, q10 @Shift row 2
vshl.s32 q13, q6, q10 @Shift row 3
vshl.s32 q14, q7, q10 @Shift row 4
vmovn.s32 d30, q11 @Narrow row 1
vmovn.s32 d31, q12 @Narrow row 2
vmovn.s32 d0 , q13 @Narrow row 3
vmovn.s32 d1 , q14 @Narrow row 4
vneg.s16 q1, q15 @Get negative
vneg.s16 q4, q0 @Get negative
vceq.s16 q5, q15, #0 @I compare with zero row 1 and 2 blk 1
vceq.s16 q6, q0 , #0 @I compare with zero row 1 and 2 blk 1
vbsl.s16 q2, q1, q15 @Restore sign of row 1 and 2
vbsl.s16 q3, q4, q0 @Restore sign of row 3 and 4
vmovn.u16 d14, q5 @I Narrow the comparison for row 1 and 2 blk 1
vmovn.u16 d15, q6 @I Narrow the comparison for row 1 and 2 blk 2
vshr.u8 q8, q7, #7 @I Reduce comaparison bit to a signle bit row 1 and 2 blk 1 and 2 [ keep the value for later use ]
vpadd.u8 d18, d16, d17 @I pair add nnz 1
vpadd.u8 d20, d18, d19 @I Pair add nnz 2
vpadd.u8 d22, d20, d21 @I Pair add nnz 3
vpadd.u8 d24, d22, d23 @I Pair add nnz4
vst1.s16 {q2-q3}, [r2] @Store blk
vmov.u8 d25, #16 @I Get max nnz
vsub.u8 d26, d25, d24 @I invert current nnz
vst1.u8 d26[0], [r9] @I Write nnz
vpop {d8-d15}
pop {r4-r12, pc}
@*****************************************************************************
@*
@* Function Name : ih264_resi_trans_quant_chroma_4x4_a9
@* Description : This function does residue calculation, forward transform
@* and quantization for 4x4 chroma block.
@*
@* Arguments : R0 :pointer to src buffer
@ R1 :pointer to pred buffer
@ R2 :pointer to dst buffer
@ R3 :source stride
@ STACK : pred stride,
@ dst stride,
@ pointer to scaling matrix,
@ pointer to threshold matrix,
@ qbits,
@ rounding factor,
@ pointer to store nnz
@ pointer to store unquantized dc values
@ Values Returned : NONE
@
@ Register Usage :
@ Stack Usage : 40 bytes
@ Cycles : Around
@ Interruptiaility : Interruptable
@
@ Known Limitations
@ \Assumptions :
@
@ Revision History :
@ DD MM YYYY Author(s) Changes
@ 11 2 2015 100664 First version
@
@*****************************************************************************
.global ih264_resi_trans_quant_chroma_4x4_a9
ih264_resi_trans_quant_chroma_4x4_a9:
@R0 :pointer to src buffer
@R1 :pointer to pred buffer
@R2 :pointer to dst buffer
@R3 :Source stride
@STACk :pred stride
@ :scale matirx,
@ :threshold matrix
@ :qbits
@ :round factor
@ :nnz
@ :pu1_dc_alt_addr
push {r4-r12, lr} @push all the variables first
add r11, sp, #40 @decrement stack pointer,to accomodate two variables
ldmfd r11, {r4-r10} @load the strides into registers
@R0 :pointer to src buffer
@R1 :pointer to pred buffer
@R2 :pointer to dst buffer
@R3 :Source stride
@R4 :Pred stride
@R5 :scale matirx,
@R6 :threshold matrix
@R7 :qbits
@R8 :round factor
@R9 :nnz
vpush {d8-d15}
mov r11, #0
sub r7, r11, r7 @Negate the qbit value for usiing LSL
@------------Fucntion Loading done----------------;
vld2.u8 {d10, d11}, [r0], r3 @load first 8 pix src row 1
vld2.u8 {d11, d12}, [r1], r4 @load first 8 pix pred row 1
vld2.u8 {d28, d29}, [r0], r3 @load first 8 pix src row 2
vld2.u8 {d29, d30}, [r1], r4 @load first 8 pix pred row 2
vld2.u8 {d25, d26}, [r0], r3 @load first 8 pix src row 3
vld2.u8 {d26, d27}, [r1], r4 @load first 8 pix pred row 3
vsubl.u8 q0, d10, d11 @find residue row 1
vld2.u8 {d22, d23}, [r0], r3 @load first 8 pix src row 4
vld2.u8 {d23, d24}, [r1], r4 @load first 8 pix pred row 4
vsubl.u8 q1, d28, d29 @find residue row 2
vsubl.u8 q2, d25, d26 @find residue row 3
vsubl.u8 q3, d22, d23 @find residue row 4
vtrn.16 d0, d2 @T12
vtrn.16 d4, d6 @T23
vtrn.32 d0, d4 @T13
vtrn.32 d2, d6 @T14
vadd.s16 d8 , d0, d6 @x0 = x4+x7
vadd.s16 d9 , d2, d4 @x1 = x5+x6
vsub.s16 d10, d2, d4 @x2 = x5-x6
vsub.s16 d11, d0, d6 @x3 = x4-x7
vshl.s16 d12, d10, #1 @U_SHIFT(x2,1,shft)
vshl.s16 d13, d11, #1 @U_SHIFT(x3,1,shft)
vadd.s16 d14, d8, d9 @x4 = x0 + x1;
vsub.s16 d16, d8, d9 @x6 = x0 - x1;
vadd.s16 d15, d13, d10 @x5 = U_SHIFT(x3,1,shft) + x2;
vsub.s16 d17, d11, d12 @x7 = x3 - U_SHIFT(x2,1,shft);
@taking transpose again so as to make do vert transform
vtrn.16 d14, d15 @T12
vtrn.16 d16, d17 @T23
vtrn.32 d14, d16 @T13
vtrn.32 d15, d17 @T24
@let us do vertical transform
@same code as horiz
vadd.s16 d18, d14, d17 @x0 = x4+x7
vadd.s16 d19, d15, d16 @x1 = x5+x6
vsub.s16 d20, d15, d16 @x2 = x5-x6
vsub.s16 d21, d14, d17 @x3 = x4-x7
vshl.s16 d22, d20, #1 @U_SHIFT(x2,1,shft)
vshl.s16 d23, d21, #1 @U_SHIFT(x3,1,shft)
vdup.s32 q4, r8 @Load rounding value row 1
vadd.s16 d24, d18, d19 @x5 = x0 + x1;
vsub.s16 d26, d18, d19 @x7 = x0 - x1;
vadd.s16 d25, d23, d20 @x6 = U_SHIFT(x3,1,shft) + x2;
vsub.s16 d27, d21, d22 @x8 = x3 - U_SHIFT(x2,1,shft);
vdup.s32 q10, r7 @Load qbit values
vst1.s16 d24[0], [r10] @Store Unquantized dc value to dc alte address
@core tranform is done for 4x8 block 1
vld1.s16 {q14-q15}, [r5] @load the scaling values
vabs.s16 q0, q12 @Abs val of row 1 blk 1
vabs.s16 q1, q13 @Abs val of row 2 blk 1
vmov.s32 q5, q4 @copy round fact for row 2
vmov.s32 q6, q4 @copy round fact for row 2
vclt.s16 q2, q12, #0 @Get the sign of row 1 blk 1
vmov.s32 q7, q4 @copy round fact for row 2
vclt.s16 q3, q13, #0 @Get the sign of row 2 blk 1
vmlal.s16 q4, d0, d28 @Multiply and add row 1
vmlal.s16 q5, d1, d29 @Multiply and add row 2
vmlal.s16 q6, d2, d30 @Multiply and add row 3
vmlal.s16 q7, d3, d31 @Multiply and add row 4
vshl.s32 q11, q4, q10 @Shift row 1
vshl.s32 q12, q5, q10 @Shift row 2
vshl.s32 q13, q6, q10 @Shift row 3
vshl.s32 q14, q7, q10 @Shift row 4
vmovn.s32 d30, q11 @Narrow row 1
vmovn.s32 d31, q12 @Narrow row 2
vmovn.s32 d0 , q13 @Narrow row 3
vmovn.s32 d1 , q14 @Narrow row 4
vneg.s16 q1, q15 @Get negative
vneg.s16 q4, q0 @Get negative
vceq.s16 q5, q15, #0 @I compare with zero row 1 and 2 blk 1
vceq.s16 q6, q0 , #0 @I compare with zero row 1 and 2 blk 1
vbsl.s16 q2, q1, q15 @Restore sign of row 1 and 2
vbsl.s16 q3, q4, q0 @Restore sign of row 3 and 4
vmovn.u16 d14, q5 @I Narrow the comparison for row 1 and 2 blk 1
vmovn.u16 d15, q6 @I Narrow the comparison for row 1 and 2 blk 2
vshr.u8 q8, q7, #7 @I Reduce comaparison bit to a signle bit row 1 and 2 blk 1 and 2 [ keep the value for later use ]
vpadd.u8 d18, d16, d17 @I pair add nnz 1
vpadd.u8 d20, d18, d19 @I Pair add nnz 2
vpadd.u8 d22, d20, d21 @I Pair add nnz 3
vpadd.u8 d24, d22, d23 @I Pair add nnz4
vst1.s16 {q2-q3}, [r2] @Store blk
vmov.u8 d25, #16 @I Get max nnz
vsub.u8 d26, d25, d24 @I invert current nnz
vst1.u8 d26[0], [r9] @I Write nnz
vpop {d8-d15}
pop {r4-r12, pc}
@*****************************************************************************
@*
@* Function Name : ih264_hadamard_quant_4x4_a9
@* Description : This function does forward hadamard transform and
@* quantization for luma dc block
@*
@* Arguments : R0 :pointer to src buffer
@ R1 :pointer to dst buffer
@ R2 :pu2_scale_matrix
@ R2 :pu2_threshold_matrix
@ STACk : u4_qbits
@ u4_round_factor
@ pu1_nnz
@ Values Returned : NONE
@
@ Register Usage :
@ Stack Usage : 0 bytes
@ Cycles : Around
@ Interruptiaility : Interruptable
@
@ Known Limitations
@ \Assumptions :
@
@ Revision History :
@ DD MM YYYY Author(s) Changes
@ 20 2 2015 100633 First version
@
@*****************************************************************************
@ih264_hadamard_quant_4x4_a9(WORD16 *pi2_src, WORD16 *pi2_dst,
@ const UWORD16 *pu2_scale_matrix,
@ const UWORD16 *pu2_threshold_matrix, UWORD32 u4_qbits,
@ UWORD32 u4_round_factor,UWORD8 *pu1_nnz
@ )
.global ih264_hadamard_quant_4x4_a9
ih264_hadamard_quant_4x4_a9:
@Registert usage
@ r0 : src
@ r1 : dst
@ r2 : *pu2_scale_matrix
@ r3 : *pu2_threshold_matrix
vld4.s16 {d0, d1, d2, d3}, [r0]! @Load 4x4 block
vpush {d8-d15}
vld1.u16 d30[0], [r2] @load pu2_scale_matrix[0]
vaddl.s16 q3, d0, d3 @x0 = x4 + x7;
vaddl.s16 q4, d1, d2 @x1 = x5 + x6;
vsubl.s16 q5, d1, d2 @x2 = x5 - x6;
vsubl.s16 q6, d0, d3 @x3 = x4 - x7;
vdup.u16 d30, d30[0] @pu2_scale_matrix[0]
vadd.s32 q7, q3, q4 @pi2_dst[0] = x0 + x1;
vadd.s32 q8, q6, q5 @pi2_dst[1] = x3 + x2;
add r3, sp, #68 @Get address of u4_round_factor
vsub.s32 q9, q3, q4 @pi2_dst[2] = x0 - x1;
vsub.s32 q10, q6, q5 @pi2_dst[3] = x3 - x2;
vtrn.s32 q7, q8 @transpose 4x4 block
vtrn.s32 q9, q10
vld1.s32 d0[0], [r3] @load u4_round_factor
vswp d15, d18
vswp d17, d20
add r3, sp, #64 @Get address of u4_qbits
vadd.s32 q11, q7, q10 @x0 = x4 + x7;
vadd.s32 q12, q8, q9 @x1 = x5 + x6;
vld1.s32 d31[0], [r3] @load u4_qbits
vsub.s32 q13, q8, q9 @x2 = x5 - x6;
vsub.s32 q14, q7, q10 @x3 = x4 - x7;
vdup.s32 q7, d0[0] @u4_round_factor
vadd.s32 q0, q11, q12 @(x0 + x1)
vadd.s32 q1, q14, q13 @(x3 + x2)
vsub.s32 q2, q11, q12 @(x0 - x1)
vsub.s32 q3, q14, q13 @(x3 - x2)
vdup.s32 q11, d31[0] @u4_round_factor
vshrn.s32 d0, q0, #1 @i4_value = (x0 + x1) >> 1;
vshrn.s32 d1, q1, #1 @i4_value = (x3 + x2) >> 1;
vshrn.s32 d2, q2, #1 @i4_value = (x0 - x1) >> 1;
vshrn.s32 d3, q3, #1 @i4_value = (x3 - x2) >> 1;
vabs.s16 q5, q0
vabs.s16 q6, q1
vmov.s32 q8, q7 @Get the round fact
vmov.s32 q9, q7
vmov.s32 q10, q7
vclt.s16 q3, q0, #0 @get the sign row 1,2
vclt.s16 q4, q1, #0
vneg.s32 q11, q11 @-u4_round_factor
vmlal.u16 q7, d10, d30
vmlal.u16 q8, d11, d30
vmlal.u16 q9, d12, d30
vmlal.u16 q10, d13, d30
vshl.u32 q7, q7, q11
vshl.u32 q8, q8, q11
vshl.u32 q9, q9, q11
vshl.u32 q10, q10, q11
vqmovn.u32 d22, q7
vqmovn.u32 d23, q8
vqmovn.u32 d24, q9
vqmovn.u32 d25, q10
vneg.s16 q13, q11
vneg.s16 q14, q12
vbsl.s16 q3, q13, q11
vbsl.s16 q4, q14, q12
vceq.s16 q5, q11, #0
vceq.s16 q6, q12, #0
vst1.s16 {q3}, [r1]!
vshrn.u16 d14, q5, #8
vshrn.u16 d15, q6, #8
ldr r3, [sp, #72] @Load *pu1_nnz
vshr.u8 q7, q7, #7
vst1.s16 {q4}, [r1]!
vadd.u8 d16, d14, d15
vmov.u8 d20, #16
vpadd.u8 d17, d16, d16
vpadd.u8 d18, d17, d17
vpadd.u8 d19, d18, d18
vsub.u8 d20, d20, d19
vst1.u8 d20[0], [r3]
vpop {d8-d15}
bx lr
@*****************************************************************************
@*
@* Function Name : ih264_hadamard_quant_2x2_uv_a9
@* Description : This function does forward hadamard transform and
@* quantization for dc block of chroma for both planes
@*
@* Arguments : R0 :pointer to src buffer
@ R1 :pointer to dst buffer
@ R2 :pu2_scale_matrix
@ R2 :pu2_threshold_matrix
@ STACk : u4_qbits
@ u4_round_factor
@ pu1_nnz
@ Values Returned : NONE
@
@ Register Usage :
@ Stack Usage : 0 bytes
@ Cycles : Around
@ Interruptiaility : Interruptable
@
@ Known Limitations
@ \Assumptions :
@
@ Revision History :
@ DD MM YYYY Author(s) Changes
@ 20 2 2015 100633 First version
@
@*****************************************************************************
@ ih264_hadamard_quant_2x2_uv_a9(WORD16 *pi2_src, WORD16 *pi2_dst,
@ const UWORD16 *pu2_scale_matrix,
@ const UWORD16 *pu2_threshold_matrix, UWORD32 u4_qbits,
@ UWORD32 u4_round_factor,UWORD8 *pu1_nnz
@ )
.global ih264_hadamard_quant_2x2_uv_a9
ih264_hadamard_quant_2x2_uv_a9:
vpush {d8-d15}
vld2.s16 {d0-d1}, [r0] @load src
add r3, sp, #68 @Get address of u4_round_factor
vaddl.s16 q3, d0, d1 @x0 = x4 + x5;, x2 = x6 + x7;
vld1.u16 d30[0], [r2] @load pu2_scale_matrix[0]
vsubl.s16 q4, d0, d1 @x1 = x4 - x5; x3 = x6 - x7;
add r0, sp, #64 @Get affress of u4_qbits
vld1.s32 d28[0], [r3] @load u4_round_factor
vtrn.s32 q3, q4 @q1 -> x0 x1, q2 -> x2 x3
vadd.s32 q0, q3, q4 @ (x0 + x2) (x1 + x3) (y0 + y2); (y1 + y3);
vld1.s32 d24[0], [r0] @load u4_qbits
vsub.s32 q1, q3, q4 @ (x0 - x2) (x1 - x3) (y0 - y2); (y1 - y3);
vdup.u16 d30, d30[0] @pu2_scale_matrix
vabs.s32 q2, q0
vabs.s32 q3, q1
vdup.s32 q14, d28[0] @u4_round_factor
vmovl.u16 q15, d30 @pu2_scale_matrix
vclt.s32 q4, q0, #0 @get the sign row 1,2
vdup.s32 q12, d24[0] @u4_round_factor
vclt.s32 q5, q1, #0
vqmovn.u32 d8, q4
vqmovn.s32 d9, q5
vmov.s32 q13, q14 @Get the round fact
vneg.s32 q12, q12 @-u4_round_factor
vmla.u32 q13, q2, q15
vmla.u32 q14, q3, q15
vshl.u32 q13, q13, q12 @>>qbit
vshl.u32 q14, q14, q12 @>>qbit
vqmovn.u32 d10, q13
vqmovn.u32 d11, q14
vneg.s16 q6, q5
vbsl.s16 q4, q6, q5 @*sign
vtrn.s32 d8, d9
vceq.s16 q7, q4, #0 @Compute nnz
vshrn.u16 d14, q7, #8 @reduce nnz comparison to 1 bit
ldr r3, [sp, #72] @Load *pu1_nnz
vshr.u8 d14, d14, #7 @reduce nnz comparison to 1 bit
vmov.u8 d20, #4 @Since we add zeros, we need to subtract from 4 to get nnz
vpadd.u8 d17, d14, d14 @Sum up nnz
vst1.s16 {q4}, [r1]! @Store the block
vpadd.u8 d17, d17, d17 @Sum up nnz
vsub.u8 d20, d20, d17 @4- numzeros
vst1.u16 d20[0], [r3] @store nnz
vpop {d8-d15}
bx lr