arm: add softfp support in kernel/arm/*dot_vfp.S
This commit is contained in:
parent
0150fabdb6
commit
54915ce343
|
@ -47,6 +47,11 @@ DROTKERNEL = rot_vfp.S
|
||||||
CROTKERNEL = rot_vfp.S
|
CROTKERNEL = rot_vfp.S
|
||||||
ZROTKERNEL = rot_vfp.S
|
ZROTKERNEL = rot_vfp.S
|
||||||
|
|
||||||
|
SDOTKERNEL = sdot_vfp.S
|
||||||
|
DDOTKERNEL = ddot_vfp.S
|
||||||
|
CDOTKERNEL = cdot_vfp.S
|
||||||
|
ZDOTKERNEL = zdot_vfp.S
|
||||||
|
|
||||||
SGEMMKERNEL = ../generic/gemmkernel_4x2.c
|
SGEMMKERNEL = ../generic/gemmkernel_4x2.c
|
||||||
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
|
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
|
||||||
SGEMMINCOPY = sgemm_ncopy_4_vfp.S
|
SGEMMINCOPY = sgemm_ncopy_4_vfp.S
|
||||||
|
@ -86,11 +91,6 @@ ZGEMMOTCOPYOBJ = zgemm_otcopy.o
|
||||||
|
|
||||||
ifeq ($(ARM_ABI),hard)
|
ifeq ($(ARM_ABI),hard)
|
||||||
|
|
||||||
SDOTKERNEL = sdot_vfp.S
|
|
||||||
DDOTKERNEL = ddot_vfp.S
|
|
||||||
CDOTKERNEL = cdot_vfp.S
|
|
||||||
ZDOTKERNEL = zdot_vfp.S
|
|
||||||
|
|
||||||
SNRM2KERNEL = nrm2_vfp.S
|
SNRM2KERNEL = nrm2_vfp.S
|
||||||
DNRM2KERNEL = nrm2_vfp.S
|
DNRM2KERNEL = nrm2_vfp.S
|
||||||
CNRM2KERNEL = nrm2_vfp.S
|
CNRM2KERNEL = nrm2_vfp.S
|
||||||
|
|
|
@ -41,8 +41,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
#define N r0
|
#define N r0
|
||||||
#define X r1
|
#define X r1
|
||||||
#define INC_X r2
|
#define INC_X r2
|
||||||
#define OLD_Y r3
|
|
||||||
|
|
||||||
|
|
||||||
/******************************************************
|
/******************************************************
|
||||||
* [fp, #-128] - [fp, #-64] is reserved
|
* [fp, #-128] - [fp, #-64] is reserved
|
||||||
|
@ -50,7 +48,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
* registers
|
* registers
|
||||||
*******************************************************/
|
*******************************************************/
|
||||||
|
|
||||||
#define OLD_INC_Y [fp, #4 ]
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
#define OLD_RETURN_ADDR r0
|
||||||
|
#define OLD_N r1
|
||||||
|
#define OLD_X r2
|
||||||
|
#define OLD_INC_X r3
|
||||||
|
#define OLD_Y [fp, #0 ]
|
||||||
|
#define OLD_INC_Y [fp, #4 ]
|
||||||
|
#define RETURN_ADDR r8
|
||||||
|
#else
|
||||||
|
#define OLD_Y r3
|
||||||
|
#define OLD_INC_Y [fp, #0 ]
|
||||||
|
#endif
|
||||||
|
|
||||||
#define I r5
|
#define I r5
|
||||||
#define Y r6
|
#define Y r6
|
||||||
|
@ -179,7 +188,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
.align 5
|
.align 5
|
||||||
|
|
||||||
push {r4 - r9, fp}
|
push {r4 - r9, fp}
|
||||||
add fp, sp, #24
|
add fp, sp, #28
|
||||||
sub sp, sp, #STACKSIZE // reserve stack
|
sub sp, sp, #STACKSIZE // reserve stack
|
||||||
|
|
||||||
sub r4, fp, #128
|
sub r4, fp, #128
|
||||||
|
@ -191,8 +200,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
vmov s2, s0
|
vmov s2, s0
|
||||||
vmov s3, s0
|
vmov s3, s0
|
||||||
|
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
mov RETURN_ADDR, OLD_RETURN_ADDR
|
||||||
|
mov N, OLD_N
|
||||||
|
mov X, OLD_X
|
||||||
|
mov INC_X, OLD_INC_X
|
||||||
|
ldr Y, OLD_Y
|
||||||
|
ldr INC_Y, OLD_INC_Y
|
||||||
|
#else
|
||||||
mov Y, OLD_Y
|
mov Y, OLD_Y
|
||||||
ldr INC_Y, OLD_INC_Y
|
ldr INC_Y, OLD_INC_Y
|
||||||
|
#endif
|
||||||
|
|
||||||
cmp N, #0
|
cmp N, #0
|
||||||
ble cdot_kernel_L999
|
ble cdot_kernel_L999
|
||||||
|
@ -265,7 +283,6 @@ cdot_kernel_S10:
|
||||||
|
|
||||||
|
|
||||||
cdot_kernel_L999:
|
cdot_kernel_L999:
|
||||||
|
|
||||||
sub r3, fp, #128
|
sub r3, fp, #128
|
||||||
vldm r3, { s8 - s15} // restore floating point registers
|
vldm r3, { s8 - s15} // restore floating point registers
|
||||||
|
|
||||||
|
@ -276,8 +293,11 @@ cdot_kernel_L999:
|
||||||
vadd.f32 s0 , s0, s2
|
vadd.f32 s0 , s0, s2
|
||||||
vsub.f32 s1 , s1, s3
|
vsub.f32 s1 , s1, s3
|
||||||
#endif
|
#endif
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
vstm RETURN_ADDR, {s0 - s1}
|
||||||
|
#endif
|
||||||
|
|
||||||
sub sp, fp, #24
|
sub sp, fp, #28
|
||||||
pop {r4 - r9, fp}
|
pop {r4 - r9, fp}
|
||||||
bx lr
|
bx lr
|
||||||
|
|
||||||
|
|
|
@ -246,6 +246,9 @@ ddot_kernel_L999:
|
||||||
vldm r3, { d8 - d15} // restore floating point registers
|
vldm r3, { d8 - d15} // restore floating point registers
|
||||||
|
|
||||||
vadd.f64 d0 , d0, d1 // set return value
|
vadd.f64 d0 , d0, d1 // set return value
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
vmov r0, r1, d0
|
||||||
|
#endif
|
||||||
sub sp, fp, #24
|
sub sp, fp, #24
|
||||||
pop {r4 - r9, fp}
|
pop {r4 - r9, fp}
|
||||||
bx lr
|
bx lr
|
||||||
|
|
|
@ -329,20 +329,19 @@ sdot_kernel_L999:
|
||||||
vldm r3, { s8 - s15} // restore floating point registers
|
vldm r3, { s8 - s15} // restore floating point registers
|
||||||
|
|
||||||
#if defined(DSDOT)
|
#if defined(DSDOT)
|
||||||
|
|
||||||
vadd.f64 d0 , d0, d1 // set return value
|
vadd.f64 d0 , d0, d1 // set return value
|
||||||
|
#else
|
||||||
#ifdef ARM_SOFTFP_ABI
|
vadd.f32 s0 , s0, s1 // set return value
|
||||||
vmov r0, r1, d0
|
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
#if defined(DSDOT)
|
||||||
|
vmov r0, r1, d0
|
||||||
#else
|
#else
|
||||||
|
|
||||||
vadd.f32 s0 , s0, s1 // set return value
|
|
||||||
#ifdef ARM_SOFTFP_ABI
|
|
||||||
vmov r0, s0
|
vmov r0, s0
|
||||||
#endif
|
#endif
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
sub sp, fp, #24
|
sub sp, fp, #24
|
||||||
pop {r4 - r9, fp}
|
pop {r4 - r9, fp}
|
||||||
bx lr
|
bx lr
|
||||||
|
|
|
@ -41,8 +41,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
#define N r0
|
#define N r0
|
||||||
#define X r1
|
#define X r1
|
||||||
#define INC_X r2
|
#define INC_X r2
|
||||||
#define OLD_Y r3
|
|
||||||
|
|
||||||
|
|
||||||
/******************************************************
|
/******************************************************
|
||||||
* [fp, #-128] - [fp, #-64] is reserved
|
* [fp, #-128] - [fp, #-64] is reserved
|
||||||
|
@ -50,7 +48,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
* registers
|
* registers
|
||||||
*******************************************************/
|
*******************************************************/
|
||||||
|
|
||||||
#define OLD_INC_Y [fp, #4 ]
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
#define OLD_RETURN_ADDR r0
|
||||||
|
#define OLD_N r1
|
||||||
|
#define OLD_X r2
|
||||||
|
#define OLD_INC_X r3
|
||||||
|
#define OLD_Y [fp, #0 ]
|
||||||
|
#define OLD_INC_Y [fp, #4 ]
|
||||||
|
#define RETURN_ADDR r8
|
||||||
|
#else
|
||||||
|
#define OLD_Y r3
|
||||||
|
#define OLD_INC_Y [fp, #0 ]
|
||||||
|
#endif
|
||||||
|
|
||||||
#define I r5
|
#define I r5
|
||||||
#define Y r6
|
#define Y r6
|
||||||
|
@ -181,7 +190,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
.align 5
|
.align 5
|
||||||
|
|
||||||
push {r4 - r9, fp}
|
push {r4 - r9, fp}
|
||||||
add fp, sp, #24
|
add fp, sp, #28
|
||||||
sub sp, sp, #STACKSIZE // reserve stack
|
sub sp, sp, #STACKSIZE // reserve stack
|
||||||
|
|
||||||
sub r4, fp, #128
|
sub r4, fp, #128
|
||||||
|
@ -194,9 +203,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||||
vcvt.f64.f32 d2, s0
|
vcvt.f64.f32 d2, s0
|
||||||
vcvt.f64.f32 d3, s0
|
vcvt.f64.f32 d3, s0
|
||||||
|
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
mov RETURN_ADDR, OLD_RETURN_ADDR
|
||||||
|
mov N, OLD_N
|
||||||
|
mov X, OLD_X
|
||||||
|
mov INC_X, OLD_INC_X
|
||||||
|
ldr Y, OLD_Y
|
||||||
|
ldr INC_Y, OLD_INC_Y
|
||||||
|
#else
|
||||||
mov Y, OLD_Y
|
mov Y, OLD_Y
|
||||||
ldr INC_Y, OLD_INC_Y
|
ldr INC_Y, OLD_INC_Y
|
||||||
|
#endif
|
||||||
|
|
||||||
cmp N, #0
|
cmp N, #0
|
||||||
ble zdot_kernel_L999
|
ble zdot_kernel_L999
|
||||||
|
@ -280,8 +297,11 @@ zdot_kernel_L999:
|
||||||
vadd.f64 d0 , d0, d2
|
vadd.f64 d0 , d0, d2
|
||||||
vsub.f64 d1 , d1, d3
|
vsub.f64 d1 , d1, d3
|
||||||
#endif
|
#endif
|
||||||
|
#if !defined(__ARM_PCS_VFP)
|
||||||
|
vstm RETURN_ADDR, {d0 - d1}
|
||||||
|
#endif
|
||||||
|
|
||||||
sub sp, fp, #24
|
sub sp, fp, #28
|
||||||
pop {r4 - r9, fp}
|
pop {r4 - r9, fp}
|
||||||
bx lr
|
bx lr
|
||||||
|
|
||||||
|
|
Loading…
Reference in New Issue