OpenBLAS/kernel/arm/sgemm_kernel_4x2_vfp.S

/***************************************************************************
Copyright (c) 2013, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

/**************************************************************************************
* 2013/11/28 Saar
* 	 BLASTEST 		: OK
* 	 CTEST			: OK
* 	 TEST			: OK
*
**************************************************************************************/

#define ASSEMBLER
#include "common.h"

#define STACKSIZE 256

#define	OLD_M	r0
#define	OLD_N	r1
#define	OLD_K	r2
#define	OLD_A	r3
#define OLD_ALPHA s0

/******************************************************
* [fp, #-128] - [fp, #-64] is reserved
* for store and restore of floating point
* registers
*******************************************************/

#define LDC	[fp, #-252 ]
#define M	[fp, #-256 ]
#define N	[fp, #-260 ]
#define K	[fp, #-264 ]
#define A	[fp, #-268 ]

#define ALPHA	[fp, #-280]

#define B	[fp, #4 ]
#define C	[fp, #8 ]
#define OLD_LDC	[fp, #12 ]

#define I	r0
#define J	r1
#define L	r2

#define	AO	r5
#define	BO	r6

#define	CO1	r8
#define	CO2	r9

#define K1	r7
#define BC	r12

#define A_PRE	96
#define B_PRE	96
#define C_PRE	64

/**************************************************************************************
* Macro definitions
**************************************************************************************/

.macro INIT4x2

	vsub.f32		s8 , s8 , s8
	vmov.f32		s9, s8
	vmov.f32		s10, s8
	vmov.f32		s11, s8
	vmov.f32		s12, s8
	vmov.f32		s13, s8
	vmov.f32		s14, s8
	vmov.f32		s15, s8

.endm



.macro KERNEL4x2_SUB

	fldmias	AO! , { s0 - s3 }
	fldmias	BO! , { s4 - s5 }

	fmacs	s8  , s0,  s4
	fmacs	s9  , s1,  s4
	fmacs	s10  , s2,  s4
	fmacs	s11  , s3,  s4

	fmacs	s12  , s0,  s5
	fmacs	s13  , s1,  s5
	fmacs	s14  , s2,  s5
	fmacs	s15  , s3,  s5

.endm

.macro SAVE4x2

	ldr	r3  , LDC
	add	CO2 , CO1, r3

	flds		s0, ALPHA

	flds	s4 , [CO1]
	flds	s5 , [CO1, #4 ]
	flds	s6 , [CO1, #8 ]
	flds	s7 , [CO1, #12 ]

	fmacs	s4 , s0 , s8
	fmacs	s5 , s0 , s9
	fmacs	s6 , s0 , s10
	fmacs	s7 , s0 , s11

	fsts	s4 , [CO1]
	fsts	s5 , [CO1, #4 ]
	fsts	s6 , [CO1, #8 ]
	fsts	s7 , [CO1, #12 ]

	flds	s4 , [CO2]
	flds	s5 , [CO2, #4 ]
	flds	s6 , [CO2, #8 ]
	flds	s7 , [CO2, #12 ]

	fmacs	s4 , s0 , s12
	fmacs	s5 , s0 , s13
	fmacs	s6 , s0 , s14
	fmacs	s7 , s0 , s15

	fsts	s4 , [CO2]
	fsts	s5 , [CO2, #4 ]
	fsts	s6 , [CO2, #8 ]
	fsts	s7 , [CO2, #12 ]

	add	CO1, CO1, #16

.endm


/******************************************************************************/

.macro INIT2x2

	vsub.f32		s8 , s8 , s8
	vmov.f32		s9, s8
	vmov.f32		s12, s8
	vmov.f32		s13, s8

.endm

.macro KERNEL2x2_SUB

	flds	s4 , [ BO ]
	flds	s5 , [ BO, #4 ]

	flds	s0 , [ AO ]
	flds	s1 , [ AO, #4 ]

	fmacs	s8  , s0,  s4
	fmacs	s9  , s1,  s4

	fmacs	s12  , s0,  s5
	fmacs	s13  , s1,  s5

	add	AO , AO, #8
	add	BO , BO, #8

.endm

.macro SAVE2x2

	ldr	r3  , LDC
	add	CO2 , CO1, r3

	flds		s0, ALPHA

	flds	s4 , [CO1]
	flds	s5 , [CO1, #4 ]

	fmacs	s4 , s0 , s8
	fmacs	s5 , s0 , s9

	fsts	s4 , [CO1]
	fsts	s5 , [CO1, #4 ]

	flds	s4 , [CO2]
	flds	s5 , [CO2, #4 ]

	fmacs	s4 , s0 , s12
	fmacs	s5 , s0 , s13

	fsts	s4 , [CO2]
	fsts	s5 , [CO2, #4 ]

	add	CO1, CO1, #8

.endm


/******************************************************************************/

.macro INIT1x2

	vsub.f32		s8 , s8 , s8
	vmov.f32		s12, s8

.endm

.macro KERNEL1x2_SUB

	flds	s4 , [ BO ]
	flds	s5 , [ BO, #4 ]

	flds	s0 , [ AO ]

	fmacs	s8  , s0,  s4

	fmacs	s12  , s0,  s5

	add	AO , AO, #4
	add	BO , BO, #8

.endm

.macro SAVE1x2

	ldr	r3  , LDC
	add	CO2 , CO1, r3

	flds		s0, ALPHA

	flds	s4 , [CO1]

	fmacs	s4 , s0 , s8

	fsts	s4 , [CO1]

	flds	s4 , [CO2]

	fmacs	s4 , s0 , s12

	fsts	s4 , [CO2]

	add	CO1, CO1, #4

.endm



/******************************************************************************/

.macro INIT4x1

	vsub.f32		s8 , s8 , s8
	vmov.f32		s9, s8
	vmov.f32		s10, s8
	vmov.f32		s11, s8

.endm



.macro KERNEL4x1_SUB

	flds	s4 , [ BO ]

	flds	s0 , [ AO ]
	flds	s1 , [ AO, #4 ]
	flds	s2 , [ AO, #8 ]
	flds	s3 , [ AO, #12 ]

	fmacs	s8  , s0,  s4
	fmacs	s9  , s1,  s4
	fmacs	s10 , s2,  s4
	fmacs	s11 , s3,  s4

	add	AO , AO, #16
	add	BO , BO, #4

.endm

.macro SAVE4x1

	flds		s0, ALPHA

	flds	s4 , [CO1]
	flds	s5 , [CO1, #4 ]
	flds	s6 , [CO1, #8 ]
	flds	s7 , [CO1, #12 ]

	fmacs	s4 , s0 , s8
	fmacs	s5 , s0 , s9
	fmacs	s6 , s0 , s10
	fmacs	s7 , s0 , s11

	fsts	s4 , [CO1]
	fsts	s5 , [CO1, #4 ]
	fsts	s6 , [CO1, #8 ]
	fsts	s7 , [CO1, #12 ]

	add	CO1, CO1, #16

.endm

/******************************************************************************/

.macro INIT2x1

	vsub.f32		s8 , s8 , s8
	vmov.f32		s9 , s8

.endm

.macro KERNEL2x1_SUB

	flds	s4 , [ BO ]

	flds	s0 , [ AO ]
	flds	s1 , [ AO, #4 ]

	fmacs	s8  , s0,  s4
	fmacs	s9  , s1,  s4

	add	AO , AO, #8
	add	BO , BO, #4

.endm

.macro SAVE2x1

	flds		s0, ALPHA

	flds	s4 , [CO1]
	flds	s5 , [CO1, #4 ]

	fmacs	s4 , s0 , s8
	fmacs	s5 , s0 , s9

	fsts	s4 , [CO1]
	fsts	s5 , [CO1, #4 ]

	add	CO1, CO1, #8

.endm


/******************************************************************************/

.macro INIT1x1

	vsub.f32		s8 , s8 , s8

.endm

.macro KERNEL1x1_SUB

	flds	s4 , [ BO ]

	flds	s0 , [ AO ]

	fmacs	s8  , s0,  s4

	add	AO , AO, #4
	add	BO , BO, #4

.endm

.macro SAVE1x1

	flds		s0, ALPHA

	flds	s4 , [CO1]

	fmacs	s4 , s0 , s8

	fsts	s4 , [CO1]

	add	CO1, CO1, #4

.endm


/**************************************************************************************
* End of macro definitions
**************************************************************************************/

	PROLOGUE

	.align 5

	push	{r4 - r9, fp}
	add	fp, sp, #24
	sub	sp, sp, #STACKSIZE				// reserve stack

	str	OLD_M, M
	str	OLD_N, N
	str	OLD_K, K
	str	OLD_A, A
	vstr	OLD_ALPHA, ALPHA

	sub	r3, fp, #128
	vstm	r3, { s8 - s15} 				// store floating point registers

	ldr	r3, OLD_LDC
	lsl	r3, r3, #2					// ldc = ldc * 4
	str	r3, LDC

	ldr	K1, K
	ldr	BC, B

	ldr	J, N
	asrs	J, J, #1					// J = J / 2
	ble	sgemm_kernel_L1_BEGIN


/*********************************************************************************************/

sgemm_kernel_L2_BEGIN:

	ldr	CO1, C						// CO1 = C
	ldr	r4 , LDC
	lsl	r4 , r4 , #1					// LDC * 2
	add	r3 , r4, CO1
	str	r3 , C						// store C

	ldr	AO, A						// AO = A

sgemm_kernel_L2_M4_BEGIN:

	ldr	I, M
	asrs	I, I, #2					// I = I / 4
	ble	sgemm_kernel_L2_M2_BEGIN

sgemm_kernel_L2_M4_20:

	INIT4x2

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L2_M4_40
	.align 5

sgemm_kernel_L2_M4_22:

	pld [ AO, #A_PRE ]
	pld [ BO, #B_PRE ]
	KERNEL4x2_SUB
	KERNEL4x2_SUB
	pld [ AO, #A_PRE ]
	KERNEL4x2_SUB
	KERNEL4x2_SUB

	pld [ AO, #A_PRE ]
	pld [ BO, #B_PRE ]
	KERNEL4x2_SUB
	KERNEL4x2_SUB
	pld [ AO, #A_PRE ]
	KERNEL4x2_SUB
	KERNEL4x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M4_22


sgemm_kernel_L2_M4_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L2_M4_100

sgemm_kernel_L2_M4_42:

	KERNEL4x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M4_42

sgemm_kernel_L2_M4_100:

	SAVE4x2

sgemm_kernel_L2_M4_END:

	subs	I, I, #1
	bgt	sgemm_kernel_L2_M4_20


sgemm_kernel_L2_M2_BEGIN:

	ldr	I, M
	tst	I , #3
	ble	sgemm_kernel_L2_END

	tst	I, #2					// I = I / 2
	ble	sgemm_kernel_L2_M1_BEGIN

sgemm_kernel_L2_M2_20:

	INIT2x2

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L2_M2_40

sgemm_kernel_L2_M2_22:

	KERNEL2x2_SUB
	KERNEL2x2_SUB
	KERNEL2x2_SUB
	KERNEL2x2_SUB

	KERNEL2x2_SUB
	KERNEL2x2_SUB
	KERNEL2x2_SUB
	KERNEL2x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M2_22


sgemm_kernel_L2_M2_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L2_M2_100

sgemm_kernel_L2_M2_42:

	KERNEL2x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M2_42

sgemm_kernel_L2_M2_100:

	SAVE2x2

sgemm_kernel_L2_M2_END:


sgemm_kernel_L2_M1_BEGIN:

	tst	I, #1					// I = I % 2
	ble	sgemm_kernel_L2_END

sgemm_kernel_L2_M1_20:

	INIT1x2

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L2_M1_40

sgemm_kernel_L2_M1_22:
	KERNEL1x2_SUB
	KERNEL1x2_SUB
	KERNEL1x2_SUB
	KERNEL1x2_SUB

	KERNEL1x2_SUB
	KERNEL1x2_SUB
	KERNEL1x2_SUB
	KERNEL1x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M1_22


sgemm_kernel_L2_M1_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L2_M1_100

sgemm_kernel_L2_M1_42:

	KERNEL1x2_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L2_M1_42

sgemm_kernel_L2_M1_100:

	SAVE1x2


sgemm_kernel_L2_END:

	mov	r3, BC
	mov	r4, K1
	lsl	r4, r4, #3					// k * 2 * 4
	add	r3, r3, r4					// B = B + K * 2 * 4
	mov	BC, r3

	subs	J , #1						// j--
	bgt	sgemm_kernel_L2_BEGIN

/*********************************************************************************************/

sgemm_kernel_L1_BEGIN:

	ldr	J , N
	tst	J , #1
	ble	sgemm_kernel_L999


	ldr	CO1, C						// CO1 = C
	ldr	r4 , LDC
	add	r3 , r4, CO1
	str	r3 , C						// store C

	ldr	AO, A						// AO = A



sgemm_kernel_L1_M4_BEGIN:

	ldr	I, M
	asrs	I, I, #2					// I = I / 4
	ble	sgemm_kernel_L1_M2_BEGIN

sgemm_kernel_L1_M4_20:

	INIT4x1

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L1_M4_40
	.align 5

sgemm_kernel_L1_M4_22:
	KERNEL4x1_SUB
	KERNEL4x1_SUB
	KERNEL4x1_SUB
	KERNEL4x1_SUB

	KERNEL4x1_SUB
	KERNEL4x1_SUB
	KERNEL4x1_SUB
	KERNEL4x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M4_22


sgemm_kernel_L1_M4_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L1_M4_100

sgemm_kernel_L1_M4_42:

	KERNEL4x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M4_42

sgemm_kernel_L1_M4_100:

	SAVE4x1

sgemm_kernel_L1_M4_END:

	subs	I, I, #1
	bgt	sgemm_kernel_L1_M4_20


sgemm_kernel_L1_M2_BEGIN:

	ldr	I, M
	tst	I , #3
	ble	sgemm_kernel_L1_END

	tst	I, #2					// I = I / 2
	ble	sgemm_kernel_L1_M1_BEGIN

sgemm_kernel_L1_M2_20:

	INIT2x1

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L1_M2_40

sgemm_kernel_L1_M2_22:

	KERNEL2x1_SUB
	KERNEL2x1_SUB
	KERNEL2x1_SUB
	KERNEL2x1_SUB

	KERNEL2x1_SUB
	KERNEL2x1_SUB
	KERNEL2x1_SUB
	KERNEL2x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M2_22


sgemm_kernel_L1_M2_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L1_M2_100

sgemm_kernel_L1_M2_42:

	KERNEL2x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M2_42

sgemm_kernel_L1_M2_100:

	SAVE2x1

sgemm_kernel_L1_M2_END:


sgemm_kernel_L1_M1_BEGIN:

	tst	I, #1					// I = I % 2
	ble	sgemm_kernel_L1_END

sgemm_kernel_L1_M1_20:

	INIT1x1

	mov	BO, BC
	asrs	L , K1, #3					// L = L / 8
	ble	sgemm_kernel_L1_M1_40

sgemm_kernel_L1_M1_22:

	KERNEL1x1_SUB
	KERNEL1x1_SUB
	KERNEL1x1_SUB
	KERNEL1x1_SUB

	KERNEL1x1_SUB
	KERNEL1x1_SUB
	KERNEL1x1_SUB
	KERNEL1x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M1_22


sgemm_kernel_L1_M1_40:

	ands	L , K1, #7					// L = L % 8
	ble	sgemm_kernel_L1_M1_100

sgemm_kernel_L1_M1_42:

	KERNEL1x1_SUB

	subs	L, L, #1
	bgt	sgemm_kernel_L1_M1_42

sgemm_kernel_L1_M1_100:

	SAVE1x1


sgemm_kernel_L1_END:


sgemm_kernel_L999:

	sub	r3, fp, #128
	vldm	r3, { s8 - s15}					// restore floating point registers

	movs	r0, #0						// set return value
	sub	sp, fp, #24
	pop	{r4 - r9, fp}
	bx	lr

	EPILOGUE