Merge pull request #2357 from chenxuqiang/dgemm_beta_zero

kernel/arm64/dgemm_beta.S: add beta == zero branch

kernel/arm64/dgemm_beta.S

@@ -80,6 +80,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	sp, sp, #(11*16)
 .endm
 
+.macro INIT_ZERO
+	fmul v0.2d, v0.2d, betaV0
+	fmul v1.2d, v1.2d, betaV0
+	fmul v2.2d, v2.2d, betaV0
+	fmul v3.2d, v3.2d, betaV0
+	fmul v4.2d, v4.2d, betaV0
+	fmul v5.2d, v5.2d, betaV0
+	fmul v6.2d, v6.2d, betaV0
+	fmul v7.2d, v7.2d, betaV0
+.endm
+
 /**************************************************************************************
 * End of macro definitions
 **************************************************************************************/
@@ -97,6 +108,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp  N, #0
 	ble	.Lgemm_beta_L999
 
+	fcmp BETA, #0.0
+	beq .Lgemm_beta_zero_01
+
 .Lgemm_beta_01:
 
 	lsl LDC, LDC, #3
@@ -180,4 +194,59 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	RESTORE_REGS
 	ret
 
+.Lgemm_beta_zero_01:
+	INIT_ZERO
+	lsl LDC, LDC, #3
+
+	.align 5
+.Lgemm_beta_zero_02:
+	mov A01, C00
+	add C00, C00, LDC
+
+	asr I, M, #4
+	cmp I, #0
+	ble .Lgemm_beta_zero_04
+
+	add A02, A01, #64
+
+	.align 5
+.Lgemm_beta_zero_03:
+	
+	st1 {v0.2d, v1.2d, v2.2d, v3.2d}, [A01]
+	add A01, A01, calc_size
+	st1 {v4.2d, v5.2d, v6.2d, v7.2d}, [A02]
+	add A02, A02, calc_size
+	
+	subs I, I, #1
+	bne .Lgemm_beta_zero_03
+
+	.align 5
+.Lgemm_beta_zero_04:
+
+	and I, M, #15
+	cmp I, #0
+	ble .Lgemm_beta_zero_06
+
+	.align 5
+.Lgemm_beta_zero_05:
+
+	str beta0, [A01]
+	add A01, A01, #8
+
+	subs I, I, #1
+	bne .Lgemm_beta_zero_05
+
+	.align 5
+.Lgemm_beta_zero_06:
+
+	subs N, N, #1
+	bne .Lgemm_beta_zero_02
+
+	.align 5
+.Lgemm_beta_zero_L999:
+
+	mov x0, #0
+	RESTORE_REGS
+	ret
+
 	EPILOGUE