optimized zgemv_t_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index 2d54920..6d0792f 100644 (file)
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -4,7 +4,7 @@ DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n.c
-#ZGEMVTKERNEL = zgemv_t.c
+ZGEMVTKERNEL = zgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c

diff --git a/kernel/x86_64/zgemv_t.c b/kernel/x86_64/zgemv_t.c
index df75afe..bb3f904 100644 (file)
--- a/kernel/x86_64/zgemv_t.c
+++ b/kernel/x86_64/zgemv_t.c
@@ -141,6 +141,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
        BLASLONG n2;
        FLOAT ybuffer[8],*xbuffer;
 
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
         inc_x *= 2;
         inc_y *= 2;
         lda   *= 2;

diff --git a/kernel/x86_64/zgemv_t_microk_haswell-2.c b/kernel/x86_64/zgemv_t_microk_haswell-2.c
index 99a620e..8325db5 100644 (file)
--- a/kernel/x86_64/zgemv_t_microk_haswell-2.c
+++ b/kernel/x86_64/zgemv_t_microk_haswell-2.c
@@ -49,23 +49,22 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
        ".align 16                                      \n\t"
        ".L01LOOP%=:                                    \n\t"
 
-        "prefetcht0      384(%2,%0,8)                   \n\t"
+        "prefetcht0      192(%2,%0,8)                   \n\t"
        "vmovddup          (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+        "prefetcht0      192(%4,%0,8)                   \n\t"
+       "vmovups        (%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
        "vmovddup         8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+       "vmovups        (%4,%0,8), %%ymm4               \n\t" // 2 complex values from a0
+        "prefetcht0      192(%5,%0,8)                   \n\t"
        "vmovddup        16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+        "prefetcht0      192(%6,%0,8)                   \n\t"
+       "vmovups        (%6,%0,8), %%ymm6               \n\t" // 2 complex values from a2
        "vmovddup        24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+        "prefetcht0      192(%7,%0,8)                   \n\t"
+       "vmovups        (%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
        "vinsertf128     $1, %%xmm2, %%ymm0 , %%ymm0    \n\t"  // real values from x0 and x1
        "vinsertf128     $1, %%xmm3, %%ymm1 , %%ymm1    \n\t"  // imag values from x0 and x1
 
-        "prefetcht0      384(%4,%0,8)                   \n\t"
-       "vmovups        (%4,%0,8), %%ymm4               \n\t" // 2 complex values from a0
-        "prefetcht0      384(%5,%0,8)                   \n\t"
-       "vmovups        (%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
-        "prefetcht0      384(%6,%0,8)                   \n\t"
-       "vmovups        (%6,%0,8), %%ymm6               \n\t" // 2 complex values from a2
-        "prefetcht0      384(%7,%0,8)                   \n\t"
-       "vmovups        (%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
-
        "vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
        "vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
        "vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
@@ -75,17 +74,16 @@ static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
        "vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
        "vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
 
+       "vmovups       32(%4,%0,8), %%ymm4              \n\t" // 2 complex values from a0
+       "vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
        "vmovddup        32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
        "vmovddup        40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
        "vmovddup        48(%2,%0,8), %%xmm2            \n\t"  // real value from x1
        "vmovddup        56(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
-       "vinsertf128     $1, %%xmm2, %%ymm0 , %%ymm0    \n\t"  // real values from x0 and x1
-       "vinsertf128     $1, %%xmm3, %%ymm1 , %%ymm1    \n\t"  // imag values from x0 and x1
-
-       "vmovups       32(%4,%0,8), %%ymm4              \n\t" // 2 complex values from a0
-       "vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
        "vmovups       32(%6,%0,8), %%ymm6              \n\t" // 2 complex values from a2
        "vmovups       32(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+       "vinsertf128     $1, %%xmm2, %%ymm0 , %%ymm0    \n\t"  // real values from x0 and x1
+       "vinsertf128     $1, %%xmm3, %%ymm1 , %%ymm1    \n\t"  // imag values from x0 and x1
 
        "vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
        "vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1