more optimizations

diff --git a/kernel/x86_64/cgemv_n_4.c b/kernel/x86_64/cgemv_n_4.c
index 2a5afd5..ff80585 100644 (file)
--- a/kernel/x86_64/cgemv_n_4.c
+++ b/kernel/x86_64/cgemv_n_4.c
@@ -131,6 +131,8 @@ static void cgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 #endif
 
 
+#ifndef HAVE_KERNEL_ADDY
+
 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)  __attribute__ ((noinline));
 
 static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
@@ -207,6 +209,8 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 
 }
 
+#endif
+
 int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
 {
        BLASLONG i;

diff --git a/kernel/x86_64/cgemv_n_microk_haswell-4.c b/kernel/x86_64/cgemv_n_microk_haswell-4.c
index 991c1dc..40a8432 100644 (file)
--- a/kernel/x86_64/cgemv_n_microk_haswell-4.c
+++ b/kernel/x86_64/cgemv_n_microk_haswell-4.c
@@ -180,3 +180,363 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 } 
 
 
+#define HAVE_KERNEL_4x2 1
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+       BLASLONG register i = 0;
+       BLASLONG register n1 = n & -8 ;
+       BLASLONG register n2 = n &  4 ;
+
+       __asm__  __volatile__
+       (
+       "vzeroupper                      \n\t"
+
+       "vbroadcastss     (%2), %%ymm0                  \n\t"  // real part x0
+       "vbroadcastss    4(%2), %%ymm1                  \n\t"  // imag part x0
+       "vbroadcastss    8(%2), %%ymm2                  \n\t"  // real part x1
+       "vbroadcastss   12(%2), %%ymm3                  \n\t"  // imag part x1
+
+       "cmpq           $0 , %1                         \n\t"
+       "je             .L01END%=                       \n\t"
+
+       ".align 16                                      \n\t"
+       ".L01LOOP%=:                                    \n\t"
+       "prefetcht0      320(%4,%0,4)                   \n\t"
+       "vmovups        (%4,%0,4), %%ymm8               \n\t" // 4 complex values form a0
+       "vmovups      32(%4,%0,4), %%ymm9               \n\t" // 4 complex values form a0
+
+       "prefetcht0      320(%5,%0,4)                   \n\t"
+       "vmovups        (%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+       "vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+       "vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+       "vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+       "vfmadd231ps      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vfmadd231ps      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+       "vfmadd231ps      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+       "vfmadd231ps      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+       "prefetcht0      320(%3,%0,4)                   \n\t"
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t"
+       "vmovups        32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+       "vmovups  %%ymm13, 32(%3,%0,4)                  \n\t"   
+
+        "addq          $16, %0                         \n\t"
+       "subq           $8 , %1                         \n\t"           
+       "jnz            .L01LOOP%=                      \n\t"
+
+       ".L01END%=:                                     \n\t"
+
+       "cmpq           $4, %6                          \n\t"
+       "jne            .L02END%=                       \n\t"
+
+       "vmovups        (%4,%0,4), %%ymm8               \n\t" // 4 complex values form a0
+       "vmovups        (%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+       "vfmadd231ps      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vfmadd231ps      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+
+       ".L02END%=:                                     \n\t"
+       "vzeroupper                      \n\t"
+
+       :
+        : 
+          "r" (i),     // 0    
+         "r" (n1),     // 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (n2)      // 6
+       : "cc", 
+         "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+         "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+         "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+         "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+         "memory"
+       );
+
+} 
+
+
+#define HAVE_KERNEL_4x1 1
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+
+       BLASLONG register i = 0;
+       BLASLONG register n1 = n & -8 ;
+       BLASLONG register n2 = n &  4 ;
+
+       __asm__  __volatile__
+       (
+       "vzeroupper                      \n\t"
+
+       "vbroadcastss     (%2), %%ymm0                  \n\t"  // real part x0
+       "vbroadcastss    4(%2), %%ymm1                  \n\t"  // imag part x0
+
+       "cmpq           $0 , %1                         \n\t"
+       "je             .L01END%=                       \n\t"
+
+       ".align 16                                      \n\t"
+       ".L01LOOP%=:                                    \n\t"
+       "prefetcht0      320(%4,%0,4)                   \n\t"
+       "vmovups        (%4,%0,4), %%ymm8               \n\t" // 4 complex values form a0
+       "vmovups      32(%4,%0,4), %%ymm9               \n\t" // 4 complex values form a0
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+       "vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+       "vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+       "prefetcht0      320(%3,%0,4)                   \n\t"
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t"
+       "vmovups        32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+       "vmovups  %%ymm13, 32(%3,%0,4)                  \n\t"   
+
+        "addq          $16, %0                         \n\t"
+       "subq           $8 , %1                         \n\t"           
+       "jnz            .L01LOOP%=                      \n\t"
+
+       ".L01END%=:                                     \n\t"
+
+       "cmpq           $4, %5                          \n\t"
+       "jne            .L02END%=                       \n\t"
+
+       "vmovups        (%4,%0,4), %%ymm8               \n\t" // 4 complex values form a0
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+
+       ".L02END%=:                                     \n\t"
+       "vzeroupper                      \n\t"
+
+       :
+        : 
+          "r" (i),     // 0    
+         "r" (n1),     // 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap),     // 4
+          "r" (n2)      // 5
+       : "cc", 
+         "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+         "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+         "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+         "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+         "memory"
+       );
+
+} 
+
+
+#define HAVE_KERNEL_ADDY 1
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)  __attribute__ ((noinline));
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+       BLASLONG i;
+
+       if ( inc_dest != 2 )
+       {
+
+               FLOAT temp_r;
+               FLOAT temp_i;
+               for ( i=0; i<n; i++ )
+               {
+#if !defined(XCONJ) 
+                       temp_r = alpha_r * src[0] - alpha_i * src[1];
+                       temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+                       temp_r =  alpha_r * src[0] + alpha_i * src[1];
+                       temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+                       *dest += temp_r;
+                       *(dest+1) += temp_i;
+
+                       src+=2;
+                       dest += inc_dest;
+               }
+               return;
+       }
+
+       i=0;
+       BLASLONG register n1 = n & -8 ;
+       BLASLONG register n2 = n &  4 ;
+
+       __asm__  __volatile__
+       (
+       "vzeroupper                      \n\t"
+
+       "vbroadcastss     (%4), %%ymm0                  \n\t"  // alpha_r
+       "vbroadcastss     (%5), %%ymm1                  \n\t"  // alpha_i
+
+       "cmpq           $0 , %1                         \n\t"
+       "je             .L01END%=                       \n\t"
+
+       ".align 16                                      \n\t"
+       ".L01LOOP%=:                                    \n\t"
+       "vmovups        (%2,%0,4), %%ymm8               \n\t" // 4 complex values from src
+       "vmovups      32(%2,%0,4), %%ymm9               \n\t" 
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+       "vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+       "vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t" // 4 complex values from dest
+       "vmovups        32(%3,%0,4),  %%ymm11           \n\t"
+
+#if  !defined(XCONJ)  
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+       "vmovups  %%ymm13, 32(%3,%0,4)                  \n\t"   
+
+        "addq          $16, %0                         \n\t"
+       "subq           $8 , %1                         \n\t"           
+       "jnz            .L01LOOP%=                      \n\t"
+
+       ".L01END%=:                                     \n\t"
+
+       "cmpq           $4, %6                          \n\t"
+       "jne            .L02END%=                       \n\t"
+
+       "vmovups        (%2,%0,4), %%ymm8               \n\t" // 4 complex values src
+
+       "vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+       "vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+
+       "vmovups          (%3,%0,4),  %%ymm10           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+
+       "vmovups  %%ymm12,   (%3,%0,4)                  \n\t" // 4 complex values to y  
+
+       ".L02END%=:                                     \n\t"
+       "vzeroupper                      \n\t"
+
+       :
+        : 
+          "r" (i),           // 0      
+         "r" (n1),           // 1
+          "r" (src),          // 2
+          "r" (dest),         // 3
+          "r" (&alpha_r),     // 4
+          "r" (&alpha_i),     // 5
+          "r" (n2)            // 6
+       : "cc", 
+         "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+         "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+         "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+         "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+         "memory"
+       );
+
+       return;
+
+}
+