New version of dgemm_kernel_4x4_bulldozer.S

The peak performance with 8 cores is now 90 GFlops

diff --git a/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S b/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
index f8a316b..ef98d9f 100644 (file)
--- a/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
+++ b/kernel/x86_64/dgemm_kernel_4x4_bulldozer.S
@@ -36,6 +36,17 @@
 /* or implied, of The University of Texas at Austin.                 */\r
 /*********************************************************************/\r
 \r
+/*********************************************************************\r
+* 2013/04/12   Saar\r
+* Performance:\r
+* 3584x3584    89 GFLOPS with 8 threads on 4 modules\r
+*              72 GFLOPS with 4 threads on 4 modules\r
+*              52 GFLOPS with 4 threads on 2 modules\r
+*              42 GFLOPS with 2 threads on 2 modules\r
+*              28 GFLOPS with 2 threads on 1 module\r
+*              22 GFLOPS with 1 thread  on 1 module\r
+*********************************************************************/\r
+\r
 #define ASSEMBLER\r
 #include "common.h"\r
  \r
@@ -88,133 +99,132 @@
 #define movupd movups\r
 \r
 #define KERNEL1(xx) \\r
-       vmovups         -16 * SIZE(AO, %rax, 4),%xmm0 ;\\r
-       vmovups         -14 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup        -16 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup        -15 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups         -14 * SIZE(AO, %rax, 4),%xmm2 ;\\r
+       vfmaddpd %xmm8,%xmm6,%xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2,%xmm7,%xmm12 ;\\r
        vmovddup        -14 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vfmaddpd %xmm8,%xmm0,%xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2,%xmm1,%xmm12 ;\\r
-       vmovddup        -13 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vmovups         -12 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups         -10 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0,%xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm9,%xmm6,%xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2,%xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0,%xmm5,%xmm10 ;\\r
+       vmovddup        -13 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vmovddup         -8 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6,%xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0,%xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm11,%xmm6,%xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups         -8 * SIZE(AO, %rax, 4),%xmm6 ;\\r
 \r
 #define KERNEL2(xx) \\r
        vmovddup        -12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup        -11 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups         -12 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups         -10 * SIZE(AO, %rax, 4),%xmm0 ;\\r
+       vfmaddpd %xmm8, %xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup        -10 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -9 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8, %xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9, %xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9, %xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         -9 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL3(xx) \\r
-       vmovups         -8 * SIZE(AO, %rax, 4),%xmm0 ;\\r
-       vmovups         -6 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup         -8 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup         -7 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups         -6 * SIZE(AO, %rax, 4),%xmm2 ;\\r
+       vfmaddpd %xmm8, %xmm6, %xmm7, %xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup         -6 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -5 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8, %xmm0, %xmm1, %xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups          -4 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups          -2 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9, %xmm0, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm9, %xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup         -5 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vmovddup                  (BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7, %xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4, %xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups                   (AO, %rax, 4), %xmm6 ;\\r
 \r
 #define KERNEL4(xx) \\r
        vmovddup         -4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup         -3 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups          -4 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups          -2 * SIZE(AO, %rax, 4),%xmm0 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1 ,%xmm12;\\r
        vmovddup         -2 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         -1 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1 ,%xmm12;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         -1 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL5(xx) \\r
-       vmovups                   (AO, %rax, 4), %xmm0 ;\\r
-       vmovups           2 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup                  (BO, %rax, 4), %xmm1 ;\\r
        vmovddup          1 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups           2 * SIZE(AO, %rax, 4),%xmm2 ;\\r
+       vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup          2 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup          3 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups         4 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups         6 * SIZE(AO, %rax, 4),%xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup          3 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vmovddup          8 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups           8 * SIZE(AO, %rax, 4), %xmm6 ;\\r
 \r
 #define KERNEL6(xx) \\r
        vmovddup        4 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup        5 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups         4 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups         6 * SIZE(AO, %rax, 4),%xmm0 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup          6 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup          7 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13,%xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13,%xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup          7 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL7(xx) \\r
-       vmovups           8 * SIZE(AO, %rax, 4), %xmm0 ;\\r
-       vmovups          10 * SIZE(AO, %rax, 4),%xmm2 ;\\r
-       vmovddup          8 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup          9 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups          10 * SIZE(AO, %rax, 4),%xmm2 ;\\r
+       vfmaddpd %xmm8,%xmm6, %xmm7,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm2, %xmm7,%xmm12 ;\\r
        vmovddup         10 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         11 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm0, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm2, %xmm1,%xmm12 ;\\r
-       vmovups          12 * SIZE(AO, %rax, 4), %xmm4 ;\\r
-       vmovups          14 * SIZE(AO, %rax, 4), %xmm6 ;\\r
-       vfmaddpd %xmm9,%xmm0, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm9,%xmm6, %xmm3,%xmm9 ;\\r
        vfmaddpd %xmm13,%xmm2, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm0, %xmm5,%xmm10 ;\\r
+       vmovddup         11 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vmovddup         16 * SIZE(BO, %rax, 4), %xmm7 ;\\r
+       vfmaddpd %xmm10,%xmm6, %xmm5,%xmm10 ;\\r
        vfmaddpd %xmm14,%xmm2, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm0, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm2, %xmm7,%xmm15 ;\\r
+       vfmaddpd %xmm11,%xmm6, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm2, %xmm4,%xmm15 ;\\r
+       vmovups          16 * SIZE(AO, %rax, 4),%xmm6 ;\\r
 \r
 #define KERNEL8(xx) \\r
        vmovddup         12 * SIZE(BO, %rax, 4), %xmm1 ;\\r
        vmovddup         13 * SIZE(BO, %rax, 4), %xmm3 ;\\r
+       vmovups          12 * SIZE(AO, %rax, 4), %xmm2 ;\\r
+       vmovups          14 * SIZE(AO, %rax, 4), %xmm0 ;\\r
+       vfmaddpd %xmm8,%xmm2, %xmm1,%xmm8 ;\\r
+       vfmaddpd %xmm12,%xmm0, %xmm1,%xmm12 ;\\r
        vmovddup         14 * SIZE(BO, %rax, 4), %xmm5 ;\\r
-       vmovddup         15 * SIZE(BO, %rax, 4), %xmm7 ;\\r
-       vfmaddpd %xmm8,%xmm4, %xmm1,%xmm8 ;\\r
-       vfmaddpd %xmm12,%xmm6, %xmm1,%xmm12 ;\\r
-       vfmaddpd %xmm9,%xmm4, %xmm3,%xmm9 ;\\r
-       vfmaddpd %xmm13, %xmm6, %xmm3,%xmm13 ;\\r
-       vfmaddpd %xmm10,%xmm4, %xmm5,%xmm10 ;\\r
-       vfmaddpd %xmm14,%xmm6, %xmm5,%xmm14 ;\\r
-       vfmaddpd %xmm11,%xmm4, %xmm7,%xmm11 ;\\r
-       vfmaddpd %xmm15,%xmm6, %xmm7,%xmm15 ;\\r
-       addq    $8 * SIZE, %rax ;\\r
+       vfmaddpd %xmm9,%xmm2, %xmm3,%xmm9 ;\\r
+       vfmaddpd %xmm13, %xmm0, %xmm3,%xmm13 ;\\r
+       vmovddup         15 * SIZE(BO, %rax, 4), %xmm4 ;\\r
+       vfmaddpd %xmm10,%xmm2, %xmm5,%xmm10 ;\\r
+       vfmaddpd %xmm14,%xmm0, %xmm5,%xmm14 ;\\r
+       vfmaddpd %xmm11,%xmm2, %xmm4,%xmm11 ;\\r
+       vfmaddpd %xmm15,%xmm0, %xmm4,%xmm15 ;\\r
 \r
 #define KERNEL_SUB1(xx) \\r
        vmovups         -16 * SIZE(AO),%xmm0 ;\\r
@@ -368,7 +378,7 @@
        movq    A, AO           # aoffset = a\r
 \r
        movq    K, %rax\r
-       salq    $BASE_SHIFT + 2, %rax\r
+       salq    $BASE_SHIFT + 2, %rax    # k << 5 # K * 32\r
        leaq     (B, %rax), BB\r
 \r
        movq    M,  I\r
@@ -389,6 +399,8 @@
        leaq    (B,  %rax, 4), BO\r
 #endif \r
 \r
+       \r
+\r
        vxorpd  %xmm8, %xmm8,%xmm8\r
        vxorpd  %xmm9, %xmm9,%xmm9\r
        vxorpd  %xmm10, %xmm10,%xmm10\r
@@ -397,12 +409,12 @@
        vxorpd  %xmm13, %xmm13,%xmm13\r
        vxorpd  %xmm14, %xmm14,%xmm14\r
        vxorpd  %xmm15, %xmm15,%xmm15\r
-\r
+/*\r
        prefetcht0       (CO1)\r
        prefetcht0       (CO1,LDC)\r
        prefetcht0       (CO2)\r
        prefetcht0       (CO2,LDC)\r
-\r
+*/\r
 #ifndef TRMMKERNEL\r
        movq    K, %rax\r
 #elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
@@ -428,25 +440,36 @@
        je      .L15\r
        // ALIGN_4\r
 \r
+       vmovups         -16 * SIZE(AO, %rax, 4),%xmm6 \r
+       vmovddup        -16 * SIZE(BO, %rax, 4), %xmm7 \r
        .align 16\r
 \r
-#define PR1 16\r
-#define PR2 24\r
+#define A_PR1 512\r
+#define A_PR2 576\r
+#define B_PR1 256\r
+#define B_PR2 576\r
 \r
 .L12:\r
-       prefetcht0 PR1*SIZE(AO,%rax,4)\r
-       prefetcht0 PR2*SIZE(AO,%rax,4)\r
-       prefetcht0 PR1*SIZE(BO,%rax,4)\r
-       prefetcht0 PR2*SIZE(BO,%rax,4)\r
+\r
+       //prefetcht0 A_PR1(AO,%rax,4)\r
+       // prefetcht0 B_PR1(BO,%rax,4)\r
        KERNEL1(16 *  0)\r
        KERNEL2(16 *  0)\r
+       //prefetcht0 A_PR1+64(AO,%rax,4)\r
+       // prefetcht0 B_PR1+64(BO,%rax,4)\r
        KERNEL3(16 *  0)\r
        KERNEL4(16 *  0)\r
+       //prefetcht0 A_PR1+128(AO,%rax,4)\r
+       // prefetcht0 B_PR1+128(BO,%rax,4)\r
        KERNEL5(16 *  0)\r
        KERNEL6(16 *  0)\r
+       //prefetcht0 A_PR1+192(AO,%rax,4)\r
+       // prefetcht0 B_PR1+192(BO,%rax,4)\r
        KERNEL7(16 *  0)\r
        KERNEL8(16 *  0)\r
-       jl      .L12\r
+       addq    $8 * SIZE, %rax \r
+       jnz     .L12\r
+\r
        ALIGN_4\r
 \r
 .L15:\r
@@ -518,13 +541,10 @@
 \r
        vfmaddpd (CO1),%xmm7, %xmm8,%xmm8\r
        vfmaddpd 2 * SIZE(CO1),%xmm7, %xmm12,%xmm12\r
-       .align 2\r
        vfmaddpd (CO1, LDC),%xmm7, %xmm9,%xmm9\r
        vfmaddpd 2 * SIZE(CO1, LDC),%xmm7, %xmm13,%xmm13\r
-       .align 2\r
        vfmaddpd (CO2),%xmm7, %xmm10,%xmm10\r
        vfmaddpd 2 * SIZE(CO2),%xmm7, %xmm14,%xmm14\r
-       .align 2\r
        vfmaddpd (CO2, LDC),%xmm7, %xmm11,%xmm11\r
        vfmaddpd 2 * SIZE(CO2, LDC),%xmm7, %xmm15,%xmm15\r
 \r