From 1cc02b43377c4246c78c660e189e4900da5338d9 Mon Sep 17 00:00:00 2001 From: wernsaar Date: Mon, 28 Jul 2014 11:50:01 +0200 Subject: [PATCH] optimized sgemm kernel for haswell --- kernel/x86_64/sgemm_kernel_16x4_haswell.S | 3859 ++++++++++++++++++++++++++++- param.h | 3 +- 2 files changed, 3756 insertions(+), 106 deletions(-) diff --git a/kernel/x86_64/sgemm_kernel_16x4_haswell.S b/kernel/x86_64/sgemm_kernel_16x4_haswell.S index 1f9f886..d88add0 100644 --- a/kernel/x86_64/sgemm_kernel_16x4_haswell.S +++ b/kernel/x86_64/sgemm_kernel_16x4_haswell.S @@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. **********************************************************************************/ /********************************************************************* -* 2013/11/13 Saar +* 2014/07/28 Saar * BLASTEST : OK * CTEST : OK * TEST : OK @@ -40,12 +40,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * A_PR1 512 * B_PR1 512 * -* +* +* 2014/07/28 Saar * Performance at 9216x9216x9216: -* 1 thread: 86 GFLOPS (SANDYBRIDGE: 59) (MKL: 83) -* 2 threads: 157 GFLOPS (SANDYBRIDGE: 116) (MKL: 155) -* 3 threads: 235 GFLOPS (SANDYBRIDGE: 165) (MKL: 230) -* 4 threads: 288 GFLOPS (SANDYBRIDGE: 223) (MKL: 267) +* 1 thread: 102 GFLOPS (SANDYBRIDGE: 59) (MKL: 83) +* 2 threads: 195 GFLOPS (SANDYBRIDGE: 116) (MKL: 155) +* 3 threads: 281 GFLOPS (SANDYBRIDGE: 165) (MKL: 230) +* 4 threads: 366 GFLOPS (SANDYBRIDGE: 223) (MKL: 267) * *********************************************************************/ @@ -69,6 +70,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define CO1 %r15 #define K %r12 #define BI %rbp +#define BO2 %rbp #define SP %rbx #define BO1 %rdi @@ -90,7 +92,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #endif +#if defined(OS_WINDOWS) #define L_BUFFER_SIZE 8192 +#else +#define L_BUFFER_SIZE 12288 +#endif #define Ndiv6 24(%rsp) #define Nmod6 32(%rsp) @@ -146,6 +152,373 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define B_PR1 512 /******************************************************************************************* +* 6 lines of N +*******************************************************************************************/ + +.macro KERNEL16x6_SUB + vmovups -16 * SIZE(AO), %ymm0 + vmovups -8 * SIZE(AO), %ymm1 + vbroadcastss -4 * SIZE(BO), %ymm2 + vbroadcastss -3 * SIZE(BO), %ymm3 + prefetcht0 A_PR1(AO) + + VFMADD231PS_( %ymm4,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm5,%ymm2,%ymm1 ) + VFMADD231PS_( %ymm6,%ymm3,%ymm0 ) + VFMADD231PS_( %ymm7,%ymm3,%ymm1 ) + + vbroadcastss -2 * SIZE(BO), %ymm2 + vbroadcastss -1 * SIZE(BO), %ymm3 + VFMADD231PS_( %ymm8,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm9,%ymm2,%ymm1 ) + VFMADD231PS_( %ymm10,%ymm3,%ymm0 ) + VFMADD231PS_( %ymm11,%ymm3,%ymm1 ) + + vbroadcastss 0 * SIZE(BO), %ymm2 + vbroadcastss 1 * SIZE(BO), %ymm3 + VFMADD231PS_( %ymm12,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm13,%ymm2,%ymm1 ) + VFMADD231PS_( %ymm14,%ymm3,%ymm0 ) + VFMADD231PS_( %ymm15,%ymm3,%ymm1 ) + + addq $6*SIZE, BO + addq $16*SIZE, AO + decq %rax +.endm + +.macro SAVE16x6 + + vbroadcastss ALPHA, %ymm0 + + vmulps %ymm0 , %ymm4 , %ymm4 + vmulps %ymm0 , %ymm5 , %ymm5 + vmulps %ymm0 , %ymm6 , %ymm6 + vmulps %ymm0 , %ymm7 , %ymm7 + vmulps %ymm0 , %ymm8 , %ymm8 + vmulps %ymm0 , %ymm9 , %ymm9 + vmulps %ymm0 , %ymm10, %ymm10 + vmulps %ymm0 , %ymm11, %ymm11 + vmulps %ymm0 , %ymm12, %ymm12 + vmulps %ymm0 , %ymm13, %ymm13 + vmulps %ymm0 , %ymm14, %ymm14 + vmulps %ymm0 , %ymm15, %ymm15 + + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %ymm4,%ymm4 + vaddps 8 * SIZE(CO1), %ymm5,%ymm5 + + vaddps (CO1, LDC), %ymm6,%ymm6 + vaddps 8 * SIZE(CO1, LDC), %ymm7,%ymm7 + + vaddps (CO1, LDC,2), %ymm8,%ymm8 + vaddps 8 * SIZE(CO1, LDC,2), %ymm9,%ymm9 + + vaddps (CO2), %ymm10,%ymm10 + vaddps 8 * SIZE(CO2), %ymm11,%ymm11 + + vaddps (CO2, LDC), %ymm12,%ymm12 + vaddps 8 * SIZE(CO2, LDC), %ymm13,%ymm13 + + vaddps (CO2, LDC,2), %ymm14,%ymm14 + vaddps 8 * SIZE(CO2, LDC,2), %ymm15,%ymm15 + +#endif + + vmovups %ymm4 , (CO1) + vmovups %ymm5 , 8 * SIZE(CO1) + + vmovups %ymm6 , (CO1, LDC) + vmovups %ymm7 , 8 * SIZE(CO1, LDC) + + vmovups %ymm8 , (CO1, LDC,2) + vmovups %ymm9 , 8 * SIZE(CO1, LDC,2) + + vmovups %ymm10, (CO2) + vmovups %ymm11, 8 * SIZE(CO2) + + vmovups %ymm12, (CO2, LDC) + vmovups %ymm13, 8 * SIZE(CO2, LDC) + + vmovups %ymm14, (CO2, LDC,2) + vmovups %ymm15, 8 * SIZE(CO2, LDC,2) + +.endm + + + +/*******************************************************************************************/ + +.macro KERNEL8x6_SUB + vmovups -16 * SIZE(AO), %ymm0 + vbroadcastss -4 * SIZE(BO), %ymm2 + vbroadcastss -3 * SIZE(BO), %ymm3 + + VFMADD231PS_( %ymm4,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm6,%ymm3,%ymm0 ) + + vbroadcastss -2 * SIZE(BO), %ymm2 + vbroadcastss -1 * SIZE(BO), %ymm3 + VFMADD231PS_( %ymm8,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm10,%ymm3,%ymm0 ) + + vbroadcastss 0 * SIZE(BO), %ymm2 + vbroadcastss 1 * SIZE(BO), %ymm3 + VFMADD231PS_( %ymm12,%ymm2,%ymm0 ) + VFMADD231PS_( %ymm14,%ymm3,%ymm0 ) + + addq $6*SIZE, BO + addq $8*SIZE, AO + decq %rax +.endm + +.macro SAVE8x6 + + vbroadcastss ALPHA, %ymm0 + + vmulps %ymm0 , %ymm4 , %ymm4 + vmulps %ymm0 , %ymm6 , %ymm6 + vmulps %ymm0 , %ymm8 , %ymm8 + vmulps %ymm0 , %ymm10, %ymm10 + vmulps %ymm0 , %ymm12, %ymm12 + vmulps %ymm0 , %ymm14, %ymm14 + + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %ymm4,%ymm4 + vaddps (CO1, LDC), %ymm6,%ymm6 + vaddps (CO1, LDC,2), %ymm8,%ymm8 + vaddps (CO2), %ymm10,%ymm10 + vaddps (CO2, LDC), %ymm12,%ymm12 + vaddps (CO2, LDC,2), %ymm14,%ymm14 + +#endif + + vmovups %ymm4 , (CO1) + vmovups %ymm6 , (CO1, LDC) + vmovups %ymm8 , (CO1, LDC,2) + vmovups %ymm10, (CO2) + vmovups %ymm12, (CO2, LDC) + vmovups %ymm14, (CO2, LDC,2) + +.endm + + + +/*******************************************************************************************/ + +.macro KERNEL4x6_SUB + vmovups -16 * SIZE(AO), %xmm0 + vbroadcastss -4 * SIZE(BO), %xmm2 + vbroadcastss -3 * SIZE(BO), %xmm3 + + VFMADD231PS_( %xmm4,%xmm2,%xmm0 ) + VFMADD231PS_( %xmm6,%xmm3,%xmm0 ) + + vbroadcastss -2 * SIZE(BO), %xmm2 + vbroadcastss -1 * SIZE(BO), %xmm3 + VFMADD231PS_( %xmm8,%xmm2,%xmm0 ) + VFMADD231PS_( %xmm10,%xmm3,%xmm0 ) + + vbroadcastss 0 * SIZE(BO), %xmm2 + vbroadcastss 1 * SIZE(BO), %xmm3 + VFMADD231PS_( %xmm12,%xmm2,%xmm0 ) + VFMADD231PS_( %xmm14,%xmm3,%xmm0 ) + + addq $6*SIZE, BO + addq $4*SIZE, AO + decq %rax +.endm + +.macro SAVE4x6 + + vbroadcastss ALPHA, %xmm0 + + vmulps %xmm0 , %xmm4 , %xmm4 + vmulps %xmm0 , %xmm6 , %xmm6 + vmulps %xmm0 , %xmm8 , %xmm8 + vmulps %xmm0 , %xmm10, %xmm10 + vmulps %xmm0 , %xmm12, %xmm12 + vmulps %xmm0 , %xmm14, %xmm14 + + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %xmm4,%xmm4 + vaddps (CO1, LDC), %xmm6,%xmm6 + vaddps (CO1, LDC,2), %xmm8,%xmm8 + vaddps (CO2), %xmm10,%xmm10 + vaddps (CO2, LDC), %xmm12,%xmm12 + vaddps (CO2, LDC,2), %xmm14,%xmm14 + +#endif + + vmovups %xmm4 , (CO1) + vmovups %xmm6 , (CO1, LDC) + vmovups %xmm8 , (CO1, LDC,2) + vmovups %xmm10, (CO2) + vmovups %xmm12, (CO2, LDC) + vmovups %xmm14, (CO2, LDC,2) + +.endm + + +/*******************************************************************************************/ + +.macro KERNEL2x6_SUB + vmovss -16 * SIZE(AO), %xmm0 + vmovss -15 * SIZE(AO), %xmm1 + vmovss -4 * SIZE(BO), %xmm2 + vmovss -3 * SIZE(BO), %xmm3 + + VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm5,%xmm2,%xmm1 ) + VFMADD231SS_( %xmm6,%xmm3,%xmm0 ) + VFMADD231SS_( %xmm7,%xmm3,%xmm1 ) + + vmovss -2 * SIZE(BO), %xmm2 + vmovss -1 * SIZE(BO), %xmm3 + VFMADD231SS_( %xmm8,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm9,%xmm2,%xmm1 ) + VFMADD231SS_( %xmm10,%xmm3,%xmm0 ) + VFMADD231SS_( %xmm11,%xmm3,%xmm1 ) + + vmovss 0 * SIZE(BO), %xmm2 + vmovss 1 * SIZE(BO), %xmm3 + VFMADD231SS_( %xmm12,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm13,%xmm2,%xmm1 ) + VFMADD231SS_( %xmm14,%xmm3,%xmm0 ) + VFMADD231SS_( %xmm15,%xmm3,%xmm1 ) + + addq $6*SIZE, BO + addq $2*SIZE, AO + decq %rax +.endm + +.macro SAVE2x6 + + vmovss ALPHA, %xmm0 + + vmulss %xmm0 , %xmm4 , %xmm4 + vmulss %xmm0 , %xmm5 , %xmm5 + vmulss %xmm0 , %xmm6 , %xmm6 + vmulss %xmm0 , %xmm7 , %xmm7 + vmulss %xmm0 , %xmm8 , %xmm8 + vmulss %xmm0 , %xmm9 , %xmm9 + vmulss %xmm0 , %xmm10, %xmm10 + vmulss %xmm0 , %xmm11, %xmm11 + vmulss %xmm0 , %xmm12, %xmm12 + vmulss %xmm0 , %xmm13, %xmm13 + vmulss %xmm0 , %xmm14, %xmm14 + vmulss %xmm0 , %xmm15, %xmm15 + + +#if !defined(TRMMKERNEL) + + vaddss (CO1), %xmm4,%xmm4 + vaddss 1 * SIZE(CO1), %xmm5,%xmm5 + + vaddss (CO1, LDC), %xmm6,%xmm6 + vaddss 1 * SIZE(CO1, LDC), %xmm7,%xmm7 + + vaddss (CO1, LDC,2), %xmm8,%xmm8 + vaddss 1 * SIZE(CO1, LDC,2), %xmm9,%xmm9 + + vaddss (CO2), %xmm10,%xmm10 + vaddss 1 * SIZE(CO2), %xmm11,%xmm11 + + vaddss (CO2, LDC), %xmm12,%xmm12 + vaddss 1 * SIZE(CO2, LDC), %xmm13,%xmm13 + + vaddss (CO2, LDC,2), %xmm14,%xmm14 + vaddss 1 * SIZE(CO2, LDC,2), %xmm15,%xmm15 + +#endif + + vmovss %xmm4 , (CO1) + vmovss %xmm5 , 1 * SIZE(CO1) + + vmovss %xmm6 , (CO1, LDC) + vmovss %xmm7 , 1 * SIZE(CO1, LDC) + + vmovss %xmm8 , (CO1, LDC,2) + vmovss %xmm9 , 1 * SIZE(CO1, LDC,2) + + vmovss %xmm10, (CO2) + vmovss %xmm11, 1 * SIZE(CO2) + + vmovss %xmm12, (CO2, LDC) + vmovss %xmm13, 1 * SIZE(CO2, LDC) + + vmovss %xmm14, (CO2, LDC,2) + vmovss %xmm15, 1 * SIZE(CO2, LDC,2) + +.endm + + +/*******************************************************************************************/ + +.macro KERNEL1x6_SUB + vmovss -16 * SIZE(AO), %xmm0 + vmovss -4 * SIZE(BO), %xmm2 + vmovss -3 * SIZE(BO), %xmm3 + + VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm6,%xmm3,%xmm0 ) + + vmovss -2 * SIZE(BO), %xmm2 + vmovss -1 * SIZE(BO), %xmm3 + VFMADD231SS_( %xmm8,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm10,%xmm3,%xmm0 ) + + vmovss 0 * SIZE(BO), %xmm2 + vmovss 1 * SIZE(BO), %xmm3 + VFMADD231SS_( %xmm12,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm14,%xmm3,%xmm0 ) + + addq $6*SIZE, BO + addq $1*SIZE, AO + decq %rax +.endm + +.macro SAVE1x6 + + vmovss ALPHA, %xmm0 + + vmulss %xmm0 , %xmm4 , %xmm4 + vmulss %xmm0 , %xmm6 , %xmm6 + vmulss %xmm0 , %xmm8 , %xmm8 + vmulss %xmm0 , %xmm10, %xmm10 + vmulss %xmm0 , %xmm12, %xmm12 + vmulss %xmm0 , %xmm14, %xmm14 + +#if !defined(TRMMKERNEL) + + vaddss (CO1), %xmm4,%xmm4 + vaddss (CO1, LDC), %xmm6,%xmm6 + vaddss (CO1, LDC,2), %xmm8,%xmm8 + vaddss (CO2), %xmm10,%xmm10 + vaddss (CO2, LDC), %xmm12,%xmm12 + vaddss (CO2, LDC,2), %xmm14,%xmm14 + +#endif + + vmovss %xmm4 , (CO1) + vmovss %xmm6 , (CO1, LDC) + vmovss %xmm8 , (CO1, LDC,2) + vmovss %xmm10, (CO2) + vmovss %xmm12, (CO2, LDC) + vmovss %xmm14, (CO2, LDC,2) + +.endm + + +/*******************************************************************************************/ + + +/******************************************************************************************* * 4 lines of N *******************************************************************************************/ @@ -343,17 +716,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #if !defined(TRMMKERNEL) - vaddps (CO1), %xmm4,%xmm4 - vaddps 1 * SIZE(CO1), %xmm5,%xmm5 + vaddss (CO1), %xmm4,%xmm4 + vaddss 1 * SIZE(CO1), %xmm5,%xmm5 - vaddps (CO1, LDC), %xmm6,%xmm6 - vaddps 1 * SIZE(CO1, LDC), %xmm7,%xmm7 + vaddss (CO1, LDC), %xmm6,%xmm6 + vaddss 1 * SIZE(CO1, LDC), %xmm7,%xmm7 - vaddps (CO2), %xmm8,%xmm8 - vaddps 1 * SIZE(CO2), %xmm9,%xmm9 + vaddss (CO2), %xmm8,%xmm8 + vaddss 1 * SIZE(CO2), %xmm9,%xmm9 - vaddps (CO2, LDC), %xmm10,%xmm10 - vaddps 1 * SIZE(CO2, LDC), %xmm11,%xmm11 + vaddss (CO2, LDC), %xmm10,%xmm10 + vaddss 1 * SIZE(CO2, LDC), %xmm11,%xmm11 #endif @@ -400,10 +773,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #if !defined(TRMMKERNEL) - vaddps (CO1), %xmm4,%xmm4 - vaddps (CO1, LDC), %xmm6,%xmm6 - vaddps (CO2), %xmm8,%xmm8 - vaddps (CO2, LDC), %xmm10,%xmm10 + vaddss (CO1), %xmm4,%xmm4 + vaddss (CO1, LDC), %xmm6,%xmm6 + vaddss (CO2), %xmm8,%xmm8 + vaddss (CO2, LDC), %xmm10,%xmm10 #endif @@ -556,11 +929,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #if !defined(TRMMKERNEL) - vaddps (CO1), %xmm4,%xmm4 - vaddps 1 * SIZE(CO1), %xmm5,%xmm5 + vaddss (CO1), %xmm4,%xmm4 + vaddss 1 * SIZE(CO1), %xmm5,%xmm5 - vaddps (CO1, LDC), %xmm6,%xmm6 - vaddps 1 * SIZE(CO1, LDC), %xmm7,%xmm7 + vaddss (CO1, LDC), %xmm6,%xmm6 + vaddss 1 * SIZE(CO1, LDC), %xmm7,%xmm7 #endif @@ -594,8 +967,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #if !defined(TRMMKERNEL) - vaddps (CO1), %xmm4,%xmm4 - vaddps (CO1, LDC), %xmm6,%xmm6 + vaddss (CO1), %xmm4,%xmm4 + vaddss (CO1, LDC), %xmm6,%xmm6 #endif @@ -621,141 +994,3418 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. addq $ 16, %rax .endm -.macro SAVE16x1 +.macro SAVE16x1 + + vbroadcastss ALPHA, %ymm0 + + vmulps %ymm0 , %ymm4 , %ymm4 + vmulps %ymm0 , %ymm5 , %ymm5 + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %ymm4,%ymm4 + vaddps 8 * SIZE(CO1), %ymm5,%ymm5 + +#endif + + vmovups %ymm4 , (CO1) + vmovups %ymm5 , 8 * SIZE(CO1) + +.endm + + +/*******************************************************************************************/ + +.macro KERNEL8x1_SUB + vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0 + vbroadcastss -4 * SIZE(BO, BI, SIZE), %ymm2 + VFMADD231PS_( %ymm4,%ymm2,%ymm0 ) + addq $ 1 , BI + addq $ 8 , %rax +.endm + +.macro SAVE8x1 + + vbroadcastss ALPHA, %ymm0 + + vmulps %ymm0 , %ymm4 , %ymm4 + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %ymm4,%ymm4 + +#endif + + vmovups %ymm4 , (CO1) + +.endm + + + +/*******************************************************************************************/ + +.macro KERNEL4x1_SUB + vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0 + vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm2 + VFMADD231PS_( %xmm4,%xmm2,%xmm0 ) + addq $ 1 , BI + addq $ 4 , %rax +.endm + +.macro SAVE4x1 + + vbroadcastss ALPHA, %xmm0 + + vmulps %xmm0 , %xmm4 , %xmm4 + +#if !defined(TRMMKERNEL) + + vaddps (CO1), %xmm4,%xmm4 + +#endif + + vmovups %xmm4 , (CO1) + +.endm + + +/*******************************************************************************************/ + +.macro KERNEL2x1_SUB + vmovss -16 * SIZE(AO, %rax, SIZE), %xmm0 + vmovss -15 * SIZE(AO, %rax, SIZE), %xmm1 + vmovss -4 * SIZE(BO, BI, SIZE), %xmm2 + VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) + VFMADD231SS_( %xmm5,%xmm2,%xmm1 ) + addq $ 1 , BI + addq $ 2 , %rax +.endm + +.macro SAVE2x1 + + vmovss ALPHA, %xmm0 + + vmulss %xmm0 , %xmm4 , %xmm4 + vmulss %xmm0 , %xmm5 , %xmm5 + +#if !defined(TRMMKERNEL) + + vaddss (CO1), %xmm4,%xmm4 + vaddss 1 * SIZE(CO1), %xmm5,%xmm5 + +#endif + + vmovss %xmm4 , (CO1) + vmovss %xmm5 , 1 * SIZE(CO1) + +.endm + + +/*******************************************************************************************/ + +.macro KERNEL1x1_SUB + vmovss -16 * SIZE(AO, %rax, SIZE), %xmm0 + vmovss -4 * SIZE(BO, BI, SIZE), %xmm2 + VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) + addq $ 1 , BI + addq $ 1 , %rax +.endm + +.macro SAVE1x1 + + vmovss ALPHA, %xmm0 + + vmulss %xmm0 , %xmm4 , %xmm4 + +#if !defined(TRMMKERNEL) + + vaddss (CO1), %xmm4,%xmm4 + +#endif + + vmovss %xmm4 , (CO1) + +.endm + + +/*******************************************************************************************/ + +#if !defined(TRMMKERNEL) + +/************************************************************************************* +* GEMM Kernel +*************************************************************************************/ + + + PROLOGUE + PROFCODE + + subq $STACKSIZE, %rsp + movq %rbx, (%rsp) + movq %rbp, 8(%rsp) + movq %r12, 16(%rsp) + movq %r13, 24(%rsp) + movq %r14, 32(%rsp) + movq %r15, 40(%rsp) + + vzeroupper + +#ifdef WINDOWS_ABI + movq %rdi, 48(%rsp) + movq %rsi, 56(%rsp) + movups %xmm6, 64(%rsp) + movups %xmm7, 80(%rsp) + movups %xmm8, 96(%rsp) + movups %xmm9, 112(%rsp) + movups %xmm10, 128(%rsp) + movups %xmm11, 144(%rsp) + movups %xmm12, 160(%rsp) + movups %xmm13, 176(%rsp) + movups %xmm14, 192(%rsp) + movups %xmm15, 208(%rsp) + + movq ARG1, OLD_M + movq ARG2, OLD_N + movq ARG3, OLD_K + movq OLD_A, A + movq OLD_B, B + movq OLD_C, C + movq OLD_LDC, LDC +#ifdef TRMMKERNEL + vmovsd OLD_OFFSET, %xmm12 +#endif + vmovaps %xmm3, %xmm0 + +#else + movq STACKSIZE + 8(%rsp), LDC +#ifdef TRMMKERNEL + movsd STACKSIZE + 16(%rsp), %xmm12 +#endif + +#endif + + movq %rsp, SP # save old stack + subq $128 + L_BUFFER_SIZE, %rsp + andq $-4096, %rsp # align stack + + STACK_TOUCH + + cmpq $0, OLD_M + je .L999 + + cmpq $0, OLD_N + je .L999 + + cmpq $0, OLD_K + je .L999 + + movq OLD_M, M + movq OLD_N, N + movq OLD_K, K + + vmovss %xmm0, ALPHA + + salq $BASE_SHIFT, LDC + + movq N, %rax + xorq %rdx, %rdx + movq $12, %rdi + divq %rdi // N / 12 + movq %rax, Ndiv6 // N / 12 + movq %rdx, Nmod6 // N % 12 + + movq Ndiv6, J + cmpq $0, J + je .L4_00 + ALIGN_4 + + +/*******************************************************************************************/ + +.L6_01: + // copy to sub buffer + movq B, BO1 + leaq BUFFER1, BO // first buffer to BO + movq K, %rax + salq $2, %rax // 4 values of B + leaq (B, %rax,4), BO2 + movq BO2, B // next offset of B + movq K, %rax + + ALIGN_4 + + +.L6_02c: + + vmovups (BO1), %xmm0 + vmovsd (BO2), %xmm1 + vmovups %xmm0, (BO) + vmovsd %xmm1, 4*SIZE(BO) + addq $ 4*SIZE,BO1 + addq $ 4*SIZE,BO2 + addq $ 6*SIZE,BO + decq %rax + jnz .L6_02c + + +.L6_10: + movq C, CO1 + leaq (C, LDC, 2), CO2 + leaq (CO2, LDC, 1), CO2 // co2 = c + 3 * ldc + leaq (C, LDC, 4), C + leaq (C, LDC, 2), C // c = c + 6 * ldc + + movq A, AO // aoffset = a + addq $ 16 * SIZE, AO + + movq M, I + sarq $4, I // i = (m >> 4) + je .L6_20 + + ALIGN_4 + +.L6_11: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax // K = K - ( K % 8 ) + je .L6_16 + + ALIGN_4 + +.L6_12: + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + je .L6_16 + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + je .L6_16 + + jmp .L6_12 + ALIGN_4 + +.L6_16: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L6_19 + + ALIGN_4 + +.L6_17: + + KERNEL16x6_SUB + + jnz .L6_17 + ALIGN_4 + + +.L6_19: + + SAVE16x6 + + addq $16 * SIZE, CO1 # coffset += 16 + addq $16 * SIZE, CO2 # coffset += 16 + decq I # i -- + jg .L6_11 + ALIGN_4 + +/************************************************************************** +* Rest of M +***************************************************************************/ +.L6_20: + // Test rest of M + + testq $15, M + jz .L6_60 // to next 6 lines of N + + testq $8, M + jz .L6_21pre + ALIGN_4 + +/**************************************************************************/ + +.L6_20_1: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L6_20_6 + + ALIGN_4 + +.L6_20_2: + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + je .L6_20_6 + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + je .L6_20_6 + + jmp .L6_20_2 + ALIGN_4 + +.L6_20_6: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L6_20_9 + + ALIGN_4 + +.L6_20_7: + + KERNEL8x6_SUB + + jnz .L6_20_7 + ALIGN_4 + + +.L6_20_9: + + SAVE8x6 + + addq $8 * SIZE, CO1 # coffset += 8 + addq $8 * SIZE, CO2 # coffset += 8 + ALIGN_4 + + + +/**************************************************************************/ + +.L6_21pre: + + testq $4, M + jz .L6_30 + ALIGN_4 + +.L6_21: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L6_26 + + ALIGN_4 + +.L6_22: + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + je .L6_26 + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + je .L6_26 + + jmp .L6_22 + ALIGN_4 + +.L6_26: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L6_29 + + ALIGN_4 + +.L6_27: + + KERNEL4x6_SUB + + jnz .L6_27 + ALIGN_4 + + +.L6_29: + + SAVE4x6 + + addq $4 * SIZE, CO1 # coffset += 4 + addq $4 * SIZE, CO2 # coffset += 4 + ALIGN_4 + + +.L6_30: + testq $2, M + jz .L6_40 + + ALIGN_4 + +.L6_31: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L6_36 + + ALIGN_4 + +.L6_32: + + prefetcht0 A_PR1(AO) + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + je .L6_36 + + prefetcht0 A_PR1(AO) + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + je .L6_36 + + jmp .L6_32 + ALIGN_4 + +.L6_36: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L6_39 + + ALIGN_4 + +.L6_37: + + KERNEL2x6_SUB + + jnz .L6_37 + ALIGN_4 + + +.L6_39: + + SAVE2x6 + + addq $2 * SIZE, CO1 # coffset += 2 + addq $2 * SIZE, CO2 # coffset += 2 + ALIGN_4 + +.L6_40: + testq $1, M + jz .L6_60 // to next 4 lines of N + + ALIGN_4 + +.L6_41: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L6_46 + + ALIGN_4 + +.L6_42: + + prefetcht0 A_PR1(AO) + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + je .L6_46 + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + je .L6_46 + + jmp .L6_42 + ALIGN_4 + +.L6_46: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L6_49 + + ALIGN_4 + +.L6_47: + + KERNEL1x6_SUB + + jnz .L6_47 + ALIGN_4 + + +.L6_49: + + SAVE1x6 + + addq $1 * SIZE, CO1 # coffset += 1 + addq $1 * SIZE, CO2 # coffset += 1 + ALIGN_4 + + + + + +.L6_60: + + +/*******************************************************************************************/ + + +.L7_01: + // copy to sub buffer + movq B, BO1 + leaq BUFFER1, BO // first buffer to BO + movq K, %rax + salq $2, %rax // 4 values of B + leaq (B, %rax,4), BO2 + movq K, %rax + + ALIGN_4 + + +.L7_02c: + + vmovsd 2*SIZE(BO1), %xmm0 + vmovups (BO2), %xmm1 + vmovsd %xmm0, (BO) + vmovups %xmm1, 2*SIZE(BO) + addq $ 4*SIZE,BO1 + addq $ 4*SIZE,BO2 + addq $ 6*SIZE,BO + decq %rax + jnz .L7_02c + + movq BO2, B // next offset of B + +.L7_10: + movq C, CO1 + leaq (C, LDC, 2), CO2 + leaq (CO2, LDC, 1), CO2 // co2 = c + 3 * ldc + leaq (C, LDC, 4), C + leaq (C, LDC, 2), C // c = c + 6 * ldc + + movq A, AO // aoffset = a + addq $ 16 * SIZE, AO + + movq M, I + sarq $4, I // i = (m >> 4) + je .L7_20 + + ALIGN_4 + +.L7_11: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax // K = K - ( K % 8 ) + je .L7_16 + + ALIGN_4 + +.L7_12: + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + je .L7_16 + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + KERNEL16x6_SUB + + je .L7_16 + + jmp .L7_12 + ALIGN_4 + +.L7_16: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L7_19 + + ALIGN_4 + +.L7_17: + + KERNEL16x6_SUB + + jnz .L7_17 + ALIGN_4 + + +.L7_19: + + SAVE16x6 + + addq $16 * SIZE, CO1 # coffset += 16 + addq $16 * SIZE, CO2 # coffset += 16 + decq I # i -- + jg .L7_11 + ALIGN_4 + +/************************************************************************** +* Rest of M +***************************************************************************/ +.L7_20: + // Test rest of M + + testq $15, M + jz .L7_60 // to next 6 lines of N + + testq $8, M + jz .L7_21pre + ALIGN_4 + +/**************************************************************************/ + +.L7_20_1: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L7_20_6 + + ALIGN_4 + +.L7_20_2: + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + je .L7_20_6 + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + prefetcht0 A_PR1(AO) + KERNEL8x6_SUB + KERNEL8x6_SUB + + je .L7_20_6 + + jmp .L7_20_2 + ALIGN_4 + +.L7_20_6: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L7_20_9 + + ALIGN_4 + +.L7_20_7: + + KERNEL8x6_SUB + + jnz .L7_20_7 + ALIGN_4 + + +.L7_20_9: + + SAVE8x6 + + addq $8 * SIZE, CO1 # coffset += 8 + addq $8 * SIZE, CO2 # coffset += 8 + ALIGN_4 + + + +/**************************************************************************/ + +.L7_21pre: + + testq $4, M + jz .L7_30 + ALIGN_4 + +.L7_21: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L7_26 + + ALIGN_4 + +.L7_22: + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + je .L7_26 + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + prefetcht0 A_PR1(AO) + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + KERNEL4x6_SUB + + je .L7_26 + + jmp .L7_22 + ALIGN_4 + +.L7_26: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L7_29 + + ALIGN_4 + +.L7_27: + + KERNEL4x6_SUB + + jnz .L7_27 + ALIGN_4 + + +.L7_29: + + SAVE4x6 + + addq $4 * SIZE, CO1 # coffset += 4 + addq $4 * SIZE, CO2 # coffset += 4 + ALIGN_4 + + +.L7_30: + testq $2, M + jz .L7_40 + + ALIGN_4 + +.L7_31: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L7_36 + + ALIGN_4 + +.L7_32: + + prefetcht0 A_PR1(AO) + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + je .L7_36 + + prefetcht0 A_PR1(AO) + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + KERNEL2x6_SUB + + je .L7_36 + + jmp .L7_32 + ALIGN_4 + +.L7_36: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L7_39 + + ALIGN_4 + +.L7_37: + + KERNEL2x6_SUB + + jnz .L7_37 + ALIGN_4 + + +.L7_39: + + SAVE2x6 + + addq $2 * SIZE, CO1 # coffset += 2 + addq $2 * SIZE, CO2 # coffset += 2 + ALIGN_4 + +.L7_40: + testq $1, M + jz .L7_60 // to next 4 lines of N + + ALIGN_4 + +.L7_41: + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + + vzeroall + + movq K, %rax + + andq $-8, %rax + je .L7_46 + + ALIGN_4 + +.L7_42: + + prefetcht0 A_PR1(AO) + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + je .L7_46 + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + KERNEL1x6_SUB + + je .L7_46 + + jmp .L7_42 + ALIGN_4 + +.L7_46: + movq K, %rax + + andq $7, %rax # if (k & 1) + je .L7_49 + + ALIGN_4 + +.L7_47: + + KERNEL1x6_SUB + + jnz .L7_47 + ALIGN_4 + + +.L7_49: + + SAVE1x6 + + addq $1 * SIZE, CO1 # coffset += 1 + addq $1 * SIZE, CO2 # coffset += 1 + ALIGN_4 + + + + + +.L7_60: + + decq J // j -- + jg .L6_01 // next 12 lines of N + + + + +/*******************************************************************************************/ +.L4_00: + + movq Nmod6, J + sarq $2, J // j = j / 4 + cmpq $ 0, J + je .L2_00 + ALIGN_4 + + +.L4_01: + // copy to sub buffer + movq B, BO1 + leaq BUFFER1, BO // first buffer to BO + movq K, %rax + sarq $2, %rax // K / 4 + jz .L4_01b + ALIGN_4 + + +.L4_01a: + prefetcht0 512(BO1) + prefetchw 512(BO) + + vmovups (BO1), %xmm0 + vmovups 4*SIZE(BO1), %xmm1 + vmovups 8*SIZE(BO1), %xmm2 + vmovups 12*SIZE(BO1), %xmm3 + + vmovups %xmm0, (BO) + vmovups %xmm1, 4*SIZE(BO) + vmovups %xmm2, 8*SIZE(BO) + vmovups %xmm3,12*SIZE(BO) + + addq $ 16*SIZE,BO1 + addq $ 16*SIZE,BO + decq %rax + jnz .L4_01a + + +.L4_01b: + + movq K, %rax + andq $3, %rax // K % 4 + jz .L4_02d + ALIGN_4 + +.L4_02c: + + vmovups (BO1), %xmm0 + vmovups %xmm0, (BO) + addq $ 4*SIZE,BO1 + addq $ 4*SIZE,BO + decq %rax + jnz .L4_02c + +.L4_02d: + + movq BO1, B // next offset of B + +.L4_10: + movq C, CO1 + leaq (C, LDC, 2), CO2 + leaq (C, LDC, 4), C // c += 4 * ldc + +#if defined(TRMMKERNEL) && defined(LEFT) + movq OFFSET, %rax + movq %rax, KK +#endif + + movq A, AO // aoffset = a + addq $ 16 * SIZE, AO + + movq M, I + sarq $4, I // i = (m >> 4) + je .L4_20 + + ALIGN_4 + +.L4_11: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $16, %rax // number of values in AO +#else + addq $4, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + andq $-8, %rax // K = K - ( K % 8 ) + je .L4_16 + movq %rax, BI // Index for BO + leaq (,BI,4) , BI // BI = BI * 4 ; number of values + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_12: + + prefetcht0 A_PR1(AO, %rax, SIZE) + prefetcht0 B_PR1(BO, BI , SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + + prefetcht0 A_PR1(AO, %rax, SIZE) + prefetcht0 B_PR1(BO, BI , SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + + je .L4_16 + + prefetcht0 A_PR1(AO, %rax, SIZE) + prefetcht0 B_PR1(BO, BI , SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + + prefetcht0 A_PR1(AO, %rax, SIZE) + prefetcht0 B_PR1(BO, BI , SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + prefetcht0 A_PR1(AO, %rax, SIZE) + KERNEL16x4_SUB + + je .L4_16 + + jmp .L4_12 + ALIGN_4 + +.L4_16: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L4_19 + + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_17: + + KERNEL16x4_SUB + + jl .L4_17 + ALIGN_4 + + +.L4_19: + + SAVE16x4 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $16, KK +#endif + + addq $16 * SIZE, CO1 # coffset += 16 + addq $16 * SIZE, CO2 # coffset += 16 + decq I # i -- + jg .L4_11 + ALIGN_4 + +/************************************************************************** +* Rest of M +***************************************************************************/ +.L4_20: + // Test rest of M + + testq $15, M + jz .L4_60 // to next 3 lines of N + + testq $8, M + jz .L4_21pre + ALIGN_4 + +/**************************************************************************/ + +.L4_20_1: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $8, %rax // number of values in A +#else + addq $4, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L4_20_6 + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_20_2: + + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + + je .L4_20_6 + + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + KERNEL8x4_SUB + + je .L4_20_6 + + jmp .L4_20_2 + ALIGN_4 + +.L4_20_6: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L4_20_9 + + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_20_7: + + KERNEL8x4_SUB + + jl .L4_20_7 + ALIGN_4 + + +.L4_20_9: + + SAVE8x4 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $8, KK +#endif + + addq $8 * SIZE, CO1 # coffset += 8 + addq $8 * SIZE, CO2 # coffset += 8 + ALIGN_4 + + + +/**************************************************************************/ + +.L4_21pre: + + testq $4, M + jz .L4_30 + ALIGN_4 + +.L4_21: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $4, %rax // number of values in A +#else + addq $4, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L4_26 + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_22: + + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + + je .L4_26 + + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + KERNEL4x4_SUB + + je .L4_26 + + jmp .L4_22 + ALIGN_4 + +.L4_26: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L4_29 + + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_27: + + KERNEL4x4_SUB + + jl .L4_27 + ALIGN_4 + + +.L4_29: + + SAVE4x4 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $4, KK +#endif + + addq $4 * SIZE, CO1 # coffset += 4 + addq $4 * SIZE, CO2 # coffset += 4 + ALIGN_4 + + +.L4_30: + testq $2, M + jz .L4_40 + + ALIGN_4 + +.L4_31: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $2, %rax // number of values in AO +#else + addq $4, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L4_36 + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_32: + + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + + je .L4_36 + + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + KERNEL2x4_SUB + + je .L4_36 + + jmp .L4_32 + ALIGN_4 + +.L4_36: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L4_39 + + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_37: + + KERNEL2x4_SUB + + jl .L4_37 + ALIGN_4 + + +.L4_39: + + SAVE2x4 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $2, KK +#endif + + addq $2 * SIZE, CO1 # coffset += 2 + addq $2 * SIZE, CO2 # coffset += 2 + ALIGN_4 + +.L4_40: + testq $1, M + jz .L4_60 // to next 4 lines of N + + ALIGN_4 + +.L4_41: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $1, %rax // number of values in AO +#else + addq $4, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + andq $-8, %rax + je .L4_46 + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_42: + + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + + je .L4_46 + + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + KERNEL1x4_SUB + + je .L4_46 + + jmp .L4_42 + ALIGN_4 + +.L4_46: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L4_49 + + movq %rax, BI // Index for BO + leaq (,BI,4), BI // BI = BI * 4 ; number of values + + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L4_47: + + KERNEL1x4_SUB + + jl .L4_47 + ALIGN_4 + + +.L4_49: + + SAVE1x4 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (,BI, 4), BI // BI = BI * 4 ; number of values + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $1, KK +#endif + + addq $1 * SIZE, CO1 # coffset += 1 + addq $1 * SIZE, CO2 # coffset += 1 + ALIGN_4 + + + + + +.L4_60: +#if defined(TRMMKERNEL) && !defined(LEFT) + addq $4, KK +#endif + + decq J // j -- + jg .L4_01 // next 4 lines of N + + + +/*******************************************************************************************/ +.L2_00: + + movq Nmod6, J + andq $3, J // j % 4 + je .L999 + + movq Nmod6, J + andq $2, J // j % 4 + je .L1_0 + +.L2_01: + + // copy to sub buffer + movq B, BO1 + leaq BUFFER1, BO // first buffer to BO + movq K, %rax + sarq $2, %rax // K / 4 + jz .L2_01b + ALIGN_4 + +.L2_01a: + + vmovsd (BO1), %xmm0 + vmovsd 2*SIZE(BO1), %xmm1 + vmovsd 4*SIZE(BO1), %xmm2 + vmovsd 6*SIZE(BO1), %xmm3 + + vmovsd %xmm0, (BO) + vmovsd %xmm1, 2*SIZE(BO) + vmovsd %xmm2, 4*SIZE(BO) + vmovsd %xmm3, 6*SIZE(BO) + + addq $8*SIZE,BO1 + addq $8*SIZE,BO + decq %rax + jnz .L2_01a + + +.L2_01b: + + movq K, %rax + andq $3, %rax // K % 4 + jz .L2_02d + ALIGN_4 + +.L2_02c: + + vmovsd (BO1), %xmm0 + vmovsd %xmm0, (BO) + addq $2*SIZE,BO1 + addq $2*SIZE,BO + decq %rax + jnz .L2_02c + +.L2_02d: + + movq BO1, B // next offset of B + +.L2_10: + movq C, CO1 + leaq (C, LDC, 2), C // c += 2 * ldc + +#if defined(TRMMKERNEL) && defined(LEFT) + movq OFFSET, %rax + movq %rax, KK +#endif + + movq A, AO // aoffset = a + addq $16 * SIZE, AO + + movq M, I + sarq $4, I // i = (m >> 4) + je .L2_20 + + ALIGN_4 + +.L2_11: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $16, %rax // number of values in AO +#else + addq $2, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + andq $-8, %rax // K = K - ( K % 8 ) + je .L2_16 + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_12: + + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + + je .L2_16 + + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + KERNEL16x2_SUB + + je .L2_16 + + jmp .L2_12 + ALIGN_4 + +.L2_16: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L2_19 + + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_17: + + KERNEL16x2_SUB + + jl .L2_17 + ALIGN_4 + + +.L2_19: + + SAVE16x2 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $16, KK +#endif + + addq $16 * SIZE, CO1 # coffset += 16 + decq I # i -- + jg .L2_11 + ALIGN_4 + +/************************************************************************** +* Rest of M +***************************************************************************/ +.L2_20: + // Test rest of M + + testq $15, M + jz .L2_60 // to next 2 lines of N + + testq $8, M + jz .L2_21pre + ALIGN_4 + +/**************************************************************************/ + +.L2_20_1: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $8, %rax // number of values in A +#else + addq $2, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L2_20_6 + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_20_2: + + + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + + je .L2_20_6 + + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + KERNEL8x2_SUB + + je .L2_20_6 + + jmp .L2_20_2 + ALIGN_4 + +.L2_20_6: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L2_20_9 + + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_20_7: + + KERNEL8x2_SUB + + jl .L2_20_7 + ALIGN_4 + + +.L2_20_9: + + SAVE8x2 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $8, KK +#endif + + addq $8 * SIZE, CO1 # coffset += 8 + ALIGN_4 + + + +/**************************************************************************/ + +.L2_21pre: + + testq $4, M + jz .L2_30 + ALIGN_4 + +.L2_21: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $4, %rax // number of values in A +#else + addq $2, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L2_26 + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 1 ; number of values + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_22: + + + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + + je .L2_26 + + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + KERNEL4x2_SUB + + je .L2_26 + + jmp .L2_22 + ALIGN_4 + +.L2_26: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L2_29 + + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_27: + + KERNEL4x2_SUB + + jl .L2_27 + ALIGN_4 + + +.L2_29: + + SAVE4x2 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $4, KK +#endif + + addq $4 * SIZE, CO1 # coffset += 4 + ALIGN_4 + + +.L2_30: + testq $2, M + jz .L2_40 + + ALIGN_4 + +.L2_31: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $2, %rax // number of values in AO +#else + addq $2, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L2_36 + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_32: + + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + + je .L2_36 + + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + KERNEL2x2_SUB + + je .L2_36 + + jmp .L2_32 + ALIGN_4 + +.L2_36: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L2_39 + + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_37: + + KERNEL2x2_SUB + + jl .L2_37 + ALIGN_4 + + +.L2_39: + + SAVE2x2 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $2, KK +#endif + + addq $2 * SIZE, CO1 # coffset += 2 + ALIGN_4 + +.L2_40: + testq $1, M + jz .L2_60 // to next 2 lines of N + + ALIGN_4 + +.L2_41: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $1, %rax // number of values in AO +#else + addq $2, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + andq $-8, %rax + je .L2_46 + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_42: + + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + + je .L2_46 + + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + KERNEL1x2_SUB + + je .L2_46 + + jmp .L2_42 + ALIGN_4 + +.L2_46: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L2_49 + + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L2_47: + + KERNEL1x2_SUB + + jl .L2_47 + ALIGN_4 + + +.L2_49: + + SAVE1x2 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BI,BI,1), BI // BI = BI * 2 ; number of values + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $1, KK +#endif + + addq $1 * SIZE, CO1 # coffset += 1 + ALIGN_4 + + + + + +.L2_60: +#if defined(TRMMKERNEL) && !defined(LEFT) + addq $2, KK +#endif + + + + +.L1_0: + +/************************************************************************************************ +* Loop for Nmod6 % 2 > 0 +*************************************************************************************************/ + + movq Nmod6, J + andq $1, J // j % 2 + je .L999 + ALIGN_4 + +.L1_01: + // copy to sub buffer + movq B, BO1 + leaq BUFFER1, BO // first buffer to BO + movq K, %rax + ALIGN_4 + +.L1_02b: + + vmovss (BO1), %xmm0 + vmovss %xmm0, (BO) + addq $1*SIZE,BO1 + addq $1*SIZE,BO + decq %rax + jnz .L1_02b + +.L1_02c: + + movq BO1, B // next offset of B + +.L1_10: + movq C, CO1 + leaq (C, LDC, 1), C // c += 1 * ldc + +#if defined(TRMMKERNEL) && defined(LEFT) + movq OFFSET, %rax + movq %rax, KK +#endif + + movq A, AO // aoffset = a + addq $16 * SIZE, AO + + movq M, I + sarq $4, I // i = (m >> 4) + je .L1_20 + + ALIGN_4 + +.L1_11: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $16, %rax // number of values in AO +#else + addq $1, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + andq $-8, %rax // K = K - ( K % 8 ) + je .L1_16 + movq %rax, BI // Index for BO + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_12: + + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + + je .L1_16 + + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + KERNEL16x1_SUB + + je .L1_16 + + jmp .L1_12 + ALIGN_4 + +.L1_16: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L1_19 + + movq %rax, BI // Index for BO + + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_17: + + KERNEL16x1_SUB + + jl .L1_17 + ALIGN_4 + + +.L1_19: + + SAVE16x1 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $4, %rax // rax = rax * 16 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $16, KK +#endif + + addq $16 * SIZE, CO1 # coffset += 16 + decq I # i -- + jg .L1_11 + ALIGN_4 + +/************************************************************************** +* Rest of M +***************************************************************************/ +.L1_20: + // Test rest of M + + testq $15, M + jz .L999 + + testq $8, M + jz .L1_21pre + ALIGN_4 + +/**************************************************************************/ + +.L1_20_1: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $8, %rax // number of values in A +#else + addq $1, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L1_20_6 + movq %rax, BI // Index for BO + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_20_2: + + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + + je .L1_20_6 + + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + KERNEL8x1_SUB + + je .L1_20_6 + + jmp .L1_20_2 + ALIGN_4 + +.L1_20_6: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L1_20_9 + + movq %rax, BI // Index for BO + + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_20_7: + + KERNEL8x1_SUB + + jl .L1_20_7 + ALIGN_4 + + +.L1_20_9: + + SAVE8x1 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $3, %rax // rax = rax * 8 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $8, KK +#endif + + addq $8 * SIZE, CO1 # coffset += 8 + ALIGN_4 + + + +/**************************************************************************/ + +.L1_21pre: + + testq $4, M + jz .L1_30 + ALIGN_4 + +.L1_21: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $4, %rax // number of values in A +#else + addq $1, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L1_26 + movq %rax, BI // Index for BO + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_22: + + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + + je .L1_26 + + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + KERNEL4x1_SUB + + je .L1_26 + + jmp .L1_22 + ALIGN_4 + +.L1_26: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif + + andq $7, %rax # if (k & 1) + je .L1_29 + + movq %rax, BI // Index for BO + + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_27: + + KERNEL4x1_SUB + + jl .L1_27 + ALIGN_4 + + +.L1_29: + + SAVE4x1 + +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $2, %rax // rax = rax * 4 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + +#if defined(TRMMKERNEL) && defined(LEFT) + addq $4, KK +#endif + + addq $4 * SIZE, CO1 # coffset += 4 + ALIGN_4 + + +.L1_30: + testq $2, M + jz .L1_40 + + ALIGN_4 + +.L1_31: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif + + + vzeroall + +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $2, %rax // number of values in AO +#else + addq $1, %rax // number of values in BO +#endif + movq %rax, KKK +#endif + + + andq $-8, %rax + je .L1_36 + movq %rax, BI // Index for BO + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 + +.L1_32: + + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB + + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB - vbroadcastss ALPHA, %ymm0 + je .L1_36 - vmulps %ymm0 , %ymm4 , %ymm4 - vmulps %ymm0 , %ymm5 , %ymm5 + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB -#if !defined(TRMMKERNEL) + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB + KERNEL2x1_SUB - vaddps (CO1), %ymm4,%ymm4 - vaddps 8 * SIZE(CO1), %ymm5,%ymm5 + je .L1_36 + + jmp .L1_32 + ALIGN_4 +.L1_36: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax #endif - vmovups %ymm4 , (CO1) - vmovups %ymm5 , 8 * SIZE(CO1) + andq $7, %rax # if (k & 1) + je .L1_39 -.endm + movq %rax, BI // Index for BO + + salq $1, %rax // rax = rax *2 ; number of values + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 +.L1_37: -/*******************************************************************************************/ + KERNEL2x1_SUB -.macro KERNEL8x1_SUB - vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0 - vbroadcastss -4 * SIZE(BO, BI, SIZE), %ymm2 - VFMADD231PS_( %ymm4,%ymm2,%ymm0 ) - addq $ 1 , BI - addq $ 8 , %rax -.endm + jl .L1_37 + ALIGN_4 -.macro SAVE8x1 - vbroadcastss ALPHA, %ymm0 +.L1_39: - vmulps %ymm0 , %ymm4 , %ymm4 + SAVE2x1 -#if !defined(TRMMKERNEL) +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + salq $1, %rax // rax = rax * 2 ; number of values + leaq (AO, %rax, SIZE), AO +#endif - vaddps (CO1), %ymm4,%ymm4 +#if defined(TRMMKERNEL) && defined(LEFT) + addq $2, KK #endif - vmovups %ymm4 , (CO1) - -.endm + addq $2 * SIZE, CO1 # coffset += 2 + ALIGN_4 +.L1_40: + testq $1, M + jz .L999 + ALIGN_4 -/*******************************************************************************************/ +.L1_41: +#if !defined(TRMMKERNEL) || \ + (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO +#else + movq KK, %rax + leaq BUFFER1, BO // first buffer to BO + addq $4 * SIZE, BO + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif -.macro KERNEL4x1_SUB - vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0 - vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm2 - VFMADD231PS_( %xmm4,%xmm2,%xmm0 ) - addq $ 1 , BI - addq $ 4 , %rax -.endm -.macro SAVE4x1 + vzeroall - vbroadcastss ALPHA, %xmm0 +#ifndef TRMMKERNEL + movq K, %rax +#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) + movq K, %rax + subq KK, %rax + movq %rax, KKK +#else + movq KK, %rax +#ifdef LEFT + addq $1, %rax // number of values in AO +#else + addq $1, %rax // number of values in BO +#endif + movq %rax, KKK +#endif - vmulps %xmm0 , %xmm4 , %xmm4 + andq $-8, %rax + je .L1_46 + movq %rax, BI // Index for BO -#if !defined(TRMMKERNEL) + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 - vaddps (CO1), %xmm4,%xmm4 +.L1_42: -#endif + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB - vmovups %xmm4 , (CO1) + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB -.endm + je .L1_46 + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB -/*******************************************************************************************/ + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB + KERNEL1x1_SUB -.macro KERNEL2x1_SUB - vmovss -16 * SIZE(AO, %rax, SIZE), %xmm0 - vmovss -15 * SIZE(AO, %rax, SIZE), %xmm1 - vmovss -4 * SIZE(BO, BI, SIZE), %xmm2 - VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) - VFMADD231SS_( %xmm5,%xmm2,%xmm1 ) - addq $ 1 , BI - addq $ 2 , %rax -.endm + je .L1_46 -.macro SAVE2x1 + jmp .L1_42 + ALIGN_4 - vmovss ALPHA, %xmm0 +.L1_46: +#ifndef TRMMKERNEL + movq K, %rax +#else + movq KKK, %rax +#endif - vmulss %xmm0 , %xmm4 , %xmm4 - vmulss %xmm0 , %xmm5 , %xmm5 + andq $7, %rax # if (k & 1) + je .L1_49 -#if !defined(TRMMKERNEL) + movq %rax, BI // Index for BO - vaddps (CO1), %xmm4,%xmm4 - vaddps 1 * SIZE(CO1), %xmm5,%xmm5 + leaq (AO, %rax, SIZE), AO + leaq (BO, BI, SIZE), BO + negq BI + negq %rax + ALIGN_4 -#endif +.L1_47: - vmovss %xmm4 , (CO1) - vmovss %xmm5 , 1 * SIZE(CO1) + KERNEL1x1_SUB -.endm + jl .L1_47 + ALIGN_4 -/*******************************************************************************************/ +.L1_49: -.macro KERNEL1x1_SUB - vmovss -16 * SIZE(AO, %rax, SIZE), %xmm0 - vmovss -4 * SIZE(BO, BI, SIZE), %xmm2 - VFMADD231SS_( %xmm4,%xmm2,%xmm0 ) - addq $ 1 , BI - addq $ 1 , %rax -.endm + SAVE1x1 -.macro SAVE1x1 +#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \ + (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA)) + movq K, %rax + subq KKK, %rax + movq %rax, BI // Index for BO + leaq (BO, BI, SIZE), BO + leaq (AO, %rax, SIZE), AO +#endif - vmovss ALPHA, %xmm0 - vmulss %xmm0 , %xmm4 , %xmm4 +#if defined(TRMMKERNEL) && defined(LEFT) + addq $1, KK +#endif -#if !defined(TRMMKERNEL) + addq $1 * SIZE, CO1 # coffset += 1 + ALIGN_4 + - vaddps (CO1), %xmm4,%xmm4 +.L999: + movq SP, %rsp + movq (%rsp), %rbx + movq 8(%rsp), %rbp + movq 16(%rsp), %r12 + movq 24(%rsp), %r13 + movq 32(%rsp), %r14 + movq 40(%rsp), %r15 +#ifdef WINDOWS_ABI + movq 48(%rsp), %rdi + movq 56(%rsp), %rsi + movups 64(%rsp), %xmm6 + movups 80(%rsp), %xmm7 + movups 96(%rsp), %xmm8 + movups 112(%rsp), %xmm9 + movups 128(%rsp), %xmm10 + movups 144(%rsp), %xmm11 + movups 160(%rsp), %xmm12 + movups 176(%rsp), %xmm13 + movups 192(%rsp), %xmm14 + movups 208(%rsp), %xmm15 #endif - vmovss %xmm4 , (CO1) + addq $STACKSIZE, %rsp + ret -.endm + EPILOGUE -/*******************************************************************************************/ + +#else /************************************************************************************* * TRMM Kernel @@ -3152,6 +6802,5 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. EPILOGUE - - +#endif diff --git a/param.h b/param.h index 863e83c..c545d21 100644 --- a/param.h +++ b/param.h @@ -1237,10 +1237,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define CGEMM_DEFAULT_P 384 #define ZGEMM_DEFAULT_P 256 -#define SGEMM_DEFAULT_Q 384 #ifdef WINDOWS_ABI +#define SGEMM_DEFAULT_Q 320 #define DGEMM_DEFAULT_Q 128 #else +#define SGEMM_DEFAULT_Q 384 #define DGEMM_DEFAULT_Q 256 #endif #define CGEMM_DEFAULT_Q 192 -- 2.7.4