\r
/*******************************************************************************************/\r
\r
+#if !defined(TRMMKERNEL)\r
+\r
\r
PROLOGUE\r
PROFCODE\r
movq OLD_B, B\r
movq OLD_C, C\r
movq OLD_LDC, LDC\r
-#ifdef TRMMKERNEL\r
- movsd OLD_OFFSET, %xmm12\r
-#endif\r
+\r
vmovaps %xmm3, %xmm0\r
\r
#else\r
movq STACKSIZE + 8(%rsp), LDC\r
-#ifdef TRMMKERNEL\r
- movsd STACKSIZE + 16(%rsp), %xmm12\r
-#endif\r
\r
#endif\r
\r
movq %rax, Ndiv6 // N / 6\r
movq %rdx, Nmod6 // N % 6\r
\r
- \r
-\r
-#ifdef TRMMKERNEL\r
- vmovsd %xmm12, OFFSET\r
- vmovsd %xmm12, KK\r
-#ifndef LEFT\r
- negq KK\r
-#endif \r
-#endif\r
\r
movq Ndiv6, J\r
cmpq $0, J\r
leaq (C, LDC, 2), C \r
leaq (C, LDC, 1), C // c += 3 * ldc\r
\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- movq OFFSET, %rax\r
- movq %rax, KK\r
-#endif\r
\r
movq A, AO // aoffset = a\r
addq $32 * SIZE, AO\r
ALIGN_4\r
\r
.L6_11:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
leaq BUFFER1, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $4, %rax // rax = rax * 16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $16, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
\r
andq $-8, %rax // K = K - ( K % 8 )\r
je .L6_16\r
ALIGN_4\r
\r
.L6_16:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L6_19\r
\r
vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
vfmaddps 8 * SIZE(CO1),%xmm0, %xmm10,%xmm10\r
vfmaddps 8 * SIZE(CO1, LDC, 2),%xmm0, %xmm12,%xmm12\r
vfmaddps 12 * SIZE(CO1, LDC, 2),%xmm0, %xmm15,%xmm15\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm7,%xmm7\r
- vmulps %xmm0, %xmm10,%xmm10\r
- vmulps %xmm0, %xmm13,%xmm13\r
-\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm8,%xmm8\r
- vmulps %xmm0, %xmm11,%xmm11\r
- vmulps %xmm0, %xmm14,%xmm14\r
-\r
- vmulps %xmm0, %xmm6,%xmm6\r
- vmulps %xmm0, %xmm9,%xmm9\r
- vmulps %xmm0, %xmm12,%xmm12\r
- vmulps %xmm0, %xmm15,%xmm15\r
-\r
-#endif\r
\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm7 , 4 * SIZE(CO1)\r
vmovups %xmm12, 8 * SIZE(CO1, LDC, 2)\r
vmovups %xmm15,12 * SIZE(CO1, LDC, 2)\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $4, %rax // rax = rax * 16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $16, KK \r
-#endif\r
\r
addq $16 * SIZE, CO1 # coffset += 16\r
decq I # i --\r
/**************************************************************************/\r
\r
.L6_20_1:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER1, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $8, %rax // number of values in A\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L6_20_6\r
ALIGN_4\r
\r
.L6_20_6:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L6_20_9\r
\r
vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
\r
vfmaddps (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
vfmaddps 4 * SIZE(CO1, LDC, 2),%xmm0, %xmm9,%xmm9\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm7,%xmm7\r
-\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm8,%xmm8\r
-\r
- vmulps %xmm0, %xmm6,%xmm6\r
- vmulps %xmm0, %xmm9,%xmm9\r
-\r
-#endif\r
\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm7 , 4 * SIZE(CO1)\r
vmovups %xmm9 , 4 * SIZE(CO1, LDC, 2)\r
\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $8, KK\r
-#endif\r
-\r
addq $8 * SIZE, CO1 # coffset += 8\r
ALIGN_4\r
\r
ALIGN_4\r
\r
.L6_21:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
leaq BUFFER1, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $4, %rax // number of values in A\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L6_26\r
ALIGN_4\r
\r
.L6_26:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L6_29\r
\r
vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps (CO1, LDC),%xmm0, %xmm5,%xmm5\r
vfmaddps (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm6,%xmm6\r
-\r
-#endif\r
\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm5 , (CO1, LDC)\r
vmovups %xmm6 , (CO1, LDC, 2)\r
\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $4, KK\r
-#endif\r
-\r
addq $4 * SIZE, CO1 # coffset += 4\r
ALIGN_4\r
\r
ALIGN_4\r
\r
.L6_31:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
leaq BUFFER1, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $2, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L6_36\r
ALIGN_4\r
\r
.L6_36:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L6_39\r
\r
vmovss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddss 1 * SIZE(CO1),%xmm0, %xmm8,%xmm8\r
vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
vfmaddss (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
vfmaddss 1 * SIZE(CO1, LDC, 2),%xmm0, %xmm12,%xmm12\r
\r
-#else\r
- vmulss %xmm0, %xmm4,%xmm4\r
- vmulss %xmm0, %xmm8,%xmm8\r
- vmulss %xmm0, %xmm5,%xmm5\r
- vmulss %xmm0, %xmm10,%xmm10\r
- vmulss %xmm0, %xmm6,%xmm6\r
- vmulss %xmm0, %xmm12,%xmm12\r
-\r
-#endif\r
\r
vmovss %xmm4 , (CO1)\r
vmovss %xmm8 , 1 * SIZE(CO1)\r
vmovss %xmm6 , (CO1, LDC, 2)\r
vmovss %xmm12, 1 * SIZE(CO1, LDC, 2)\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $2, KK\r
-#endif\r
-\r
addq $2 * SIZE, CO1 # coffset += 2\r
ALIGN_4\r
\r
ALIGN_4\r
\r
.L6_41:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER1, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $1, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
\r
andq $-8, %rax\r
je .L6_46\r
ALIGN_4\r
\r
.L6_46:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L6_49\r
\r
vmovss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
vfmaddss (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
\r
-#else\r
- vmulss %xmm0, %xmm4,%xmm4\r
- vmulss %xmm0, %xmm5,%xmm5\r
- vmulss %xmm0, %xmm6,%xmm6\r
-\r
-#endif\r
\r
vmovss %xmm4 , (CO1)\r
vmovss %xmm5 , (CO1, LDC)\r
vmovss %xmm6 , (CO1, LDC, 2)\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $1, KK\r
-#endif\r
-\r
addq $1 * SIZE, CO1 # coffset += 1\r
ALIGN_4\r
\r
leaq (C, LDC, 2), C \r
leaq (C, LDC, 1), C // c += 3 * ldc\r
\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- movq OFFSET, %rax\r
- movq %rax, KK\r
-#endif\r
\r
movq A, AO // aoffset = a\r
addq $32 * SIZE, AO\r
ALIGN_4\r
\r
.L7_11:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER2, BO // second buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER2, BO // second buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $4, %rax // rax = rax * 16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $16, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
\r
andq $-8, %rax // K = K - ( K % 8 )\r
je .L7_16\r
ALIGN_4\r
\r
.L7_16:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L7_19\r
\r
vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
vfmaddps 8 * SIZE(CO1),%xmm0, %xmm10,%xmm10\r
vfmaddps 8 * SIZE(CO1, LDC, 2),%xmm0, %xmm12,%xmm12\r
vfmaddps 12 * SIZE(CO1, LDC, 2),%xmm0, %xmm15,%xmm15\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm7,%xmm7\r
- vmulps %xmm0, %xmm10,%xmm10\r
- vmulps %xmm0, %xmm13,%xmm13\r
-\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm8,%xmm8\r
- vmulps %xmm0, %xmm11,%xmm11\r
- vmulps %xmm0, %xmm14,%xmm14\r
-\r
- vmulps %xmm0, %xmm6,%xmm6\r
- vmulps %xmm0, %xmm9,%xmm9\r
- vmulps %xmm0, %xmm12,%xmm12\r
- vmulps %xmm0, %xmm15,%xmm15\r
-\r
-#endif\r
\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm7 , 4 * SIZE(CO1)\r
vmovups %xmm12, 8 * SIZE(CO1, LDC, 2)\r
vmovups %xmm15,12 * SIZE(CO1, LDC, 2)\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $4, %rax // rax = rax * 16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $16, KK \r
-#endif\r
\r
addq $16 * SIZE, CO1 # coffset += 16\r
decq I # i --\r
/**************************************************************************/\r
\r
.L7_20_1:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER2, BO // first buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER2, BO // first buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $8, %rax // number of values in A\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L7_20_6\r
ALIGN_4\r
\r
.L7_20_6:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L7_20_9\r
jl .L7_20_7\r
ALIGN_4\r
\r
-\r
.L7_20_9:\r
\r
- vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
+ vbroadcastss ALPHA, %xmm0\r
\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
vfmaddps (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
vfmaddps 4 * SIZE(CO1, LDC, 2),%xmm0, %xmm9,%xmm9\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm7,%xmm7\r
-\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm8,%xmm8\r
-\r
- vmulps %xmm0, %xmm6,%xmm6\r
- vmulps %xmm0, %xmm9,%xmm9\r
-\r
-#endif\r
-\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm7 , 4 * SIZE(CO1)\r
\r
vmovups %xmm9 , 4 * SIZE(CO1, LDC, 2)\r
\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $8, KK\r
-#endif\r
-\r
addq $8 * SIZE, CO1 # coffset += 8\r
ALIGN_4\r
\r
ALIGN_4\r
\r
.L7_21:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER2, BO // second buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER2, BO // second buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $4, %rax // number of values in A\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L7_26\r
ALIGN_4\r
\r
.L7_26:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L7_29\r
\r
vbroadcastss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddps (CO1, LDC),%xmm0, %xmm5,%xmm5\r
vfmaddps (CO1, LDC, 2),%xmm0, %xmm6 ,%xmm6\r
\r
-#else\r
- vmulps %xmm0, %xmm4,%xmm4\r
- vmulps %xmm0, %xmm5,%xmm5\r
- vmulps %xmm0, %xmm6,%xmm6\r
-\r
-#endif\r
-\r
vmovups %xmm4 , (CO1)\r
vmovups %xmm5 , (CO1, LDC)\r
vmovups %xmm6 , (CO1, LDC, 2)\r
\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $4, KK\r
-#endif\r
-\r
addq $4 * SIZE, CO1 # coffset += 4\r
ALIGN_4\r
\r
ALIGN_4\r
\r
.L7_31:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER2, BO // second buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
leaq BUFFER2, BO // second buffer to BO\r
addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $2, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
\r
andq $-8, %rax\r
je .L7_36\r
ALIGN_4\r
\r
.L7_36:\r
-#ifndef TRMMKERNEL\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
je .L7_39\r
\r
vmovss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
vfmaddss 1 * SIZE(CO1),%xmm0, %xmm8,%xmm8\r
vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
vfmaddss (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
vfmaddss 1 * SIZE(CO1, LDC, 2),%xmm0, %xmm12,%xmm12\r
\r
-#else\r
- vmulss %xmm0, %xmm4,%xmm4\r
- vmulss %xmm0, %xmm8,%xmm8\r
- vmulss %xmm0, %xmm5,%xmm5\r
- vmulss %xmm0, %xmm10,%xmm10\r
- vmulss %xmm0, %xmm6,%xmm6\r
- vmulss %xmm0, %xmm12,%xmm12\r
-\r
-#endif\r
-\r
vmovss %xmm4 , (CO1)\r
vmovss %xmm8 , 1 * SIZE(CO1)\r
vmovss %xmm5 , (CO1, LDC)\r
vmovss %xmm6 , (CO1, LDC, 2)\r
vmovss %xmm12, 1 * SIZE(CO1, LDC, 2)\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
+ addq $2 * SIZE, CO1 # coffset += 2\r
+ ALIGN_4\r
+\r
+.L7_40:\r
+ testq $1, M \r
+ jz .L7_60 // to next 3 lines of N\r
+\r
+ ALIGN_4\r
+\r
+.L7_41:\r
+ leaq BUFFER2, BO // second buffer to BO\r
+ addq $6 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L7_46\r
+ movq %rax, BI // Index for BO\r
leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- salq $1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L7_42:\r
+\r
+ KERNEL1x3_1(xxx)\r
+ KERNEL1x3_2(xxx)\r
+ KERNEL1x3_3(xxx)\r
+ KERNEL1x3_4(xxx)\r
+\r
+ KERNEL1x3_1(xxx)\r
+ KERNEL1x3_2(xxx)\r
+ KERNEL1x3_3(xxx)\r
+ KERNEL1x3_4(xxx)\r
+\r
+ je .L7_46\r
+\r
+ KERNEL1x3_1(xxx)\r
+ KERNEL1x3_2(xxx)\r
+ KERNEL1x3_3(xxx)\r
+ KERNEL1x3_4(xxx)\r
+\r
+ KERNEL1x3_1(xxx)\r
+ KERNEL1x3_2(xxx)\r
+ KERNEL1x3_3(xxx)\r
+ KERNEL1x3_4(xxx)\r
+\r
+ je .L7_46\r
+\r
+ jmp .L7_42\r
+ ALIGN_4\r
+\r
+.L7_46:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L7_49\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
+\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L7_47:\r
+\r
+ KERNEL1x3_SUB(xxx)\r
+ addq $3, BI\r
+ addq $1, %rax\r
+ jl .L7_47\r
+ ALIGN_4\r
+\r
+\r
+.L7_49:\r
+\r
+ vmovss ALPHA, %xmm0\r
+\r
+ vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+ vfmaddss (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
+\r
+ vmovss %xmm4 , (CO1)\r
+ vmovss %xmm5 , (CO1, LDC)\r
+ vmovss %xmm6 , (CO1, LDC, 2)\r
+\r
+ addq $1 * SIZE, CO1 # coffset += 1\r
+ ALIGN_4\r
+ \r
+\r
+\r
+.L7_60:\r
+\r
+ decq J // j --\r
+ jg .L6_01\r
+\r
+\r
+.L2_0:\r
+ cmpq $0, Nmod6 // N % 6 == 0\r
+ je .L999\r
+\r
+/************************************************************************************************\r
+* Loop for Nmod6 / 2 > 0\r
+*************************************************************************************************/\r
+\r
+ movq Nmod6, J \r
+ sarq $1, J // j = j / 2\r
+ je .L1_0\r
+ ALIGN_4\r
+\r
+.L2_01:\r
+ // copy to sub buffer\r
+ movq B, BO1\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ movq K, %rax\r
+ ALIGN_4\r
+\r
+.L2_02b:\r
+\r
+ vmovsd (BO1), %xmm0\r
+ vmovsd %xmm0, (BO)\r
+ addq $2*SIZE,BO1\r
+ addq $2*SIZE,BO\r
+ decq %rax\r
+ jnz .L2_02b\r
+\r
+.L2_02c:\r
+\r
+ movq BO1, B // next offset of B\r
+\r
+.L2_10:\r
+ movq C, CO1\r
+ leaq (C, LDC, 2), C // c += 2 * ldc\r
+\r
+ \r
+ movq A, AO // aoffset = a\r
+ addq $32 * SIZE, AO\r
+\r
+ movq M, I\r
+ sarq $4, I // i = (m >> 4)\r
+ je .L2_20\r
+\r
+ ALIGN_4\r
+\r
+.L2_11:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $4 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax // K = K - ( K % 8 )\r
+ je .L2_16\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $4, %rax // rax = rax * 16 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_12:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL16x2_1(xxx)\r
+ KERNEL16x2_2(xxx)\r
+ KERNEL16x2_3(xxx)\r
+ KERNEL16x2_4(xxx)\r
+\r
+ KERNEL16x2_1(xxx)\r
+ KERNEL16x2_2(xxx)\r
+ KERNEL16x2_3(xxx)\r
+ KERNEL16x2_4(xxx)\r
+\r
+ je .L2_16\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL16x2_1(xxx)\r
+ KERNEL16x2_2(xxx)\r
+ KERNEL16x2_3(xxx)\r
+ KERNEL16x2_4(xxx)\r
+\r
+ KERNEL16x2_1(xxx)\r
+ KERNEL16x2_2(xxx)\r
+ KERNEL16x2_3(xxx)\r
+ KERNEL16x2_4(xxx)\r
+\r
+ je .L2_16\r
+\r
+ jmp .L2_12\r
+ ALIGN_4\r
+\r
+.L2_16:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L2_19\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $4, %rax // rax = rax * 16 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_17:\r
+\r
+ KERNEL16x2_SUB(xxx)\r
+ addq $2, BI\r
+ addq $16, %rax\r
+ jl .L2_17\r
+ ALIGN_4\r
+\r
+\r
+.L2_19:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
+ vfmaddps 8 * SIZE(CO1),%xmm0, %xmm10,%xmm10\r
+ vfmaddps 12 * SIZE(CO1),%xmm0, %xmm13,%xmm13\r
+\r
+ vfmaddps (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+ vfmaddps 4 * SIZE(CO1, LDC),%xmm0, %xmm8,%xmm8\r
+ vfmaddps 8 * SIZE(CO1, LDC),%xmm0, %xmm11,%xmm11\r
+ vfmaddps 12 * SIZE(CO1, LDC),%xmm0, %xmm14,%xmm14\r
+\r
+ vmovups %xmm4 , (CO1)\r
+ vmovups %xmm7 , 4 * SIZE(CO1)\r
+ vmovups %xmm10, 8 * SIZE(CO1)\r
+ vmovups %xmm13,12 * SIZE(CO1)\r
+\r
+ vmovups %xmm5 , (CO1, LDC)\r
+ vmovups %xmm8 , 4 * SIZE(CO1, LDC)\r
+ vmovups %xmm11, 8 * SIZE(CO1, LDC)\r
+ vmovups %xmm14,12 * SIZE(CO1, LDC)\r
+\r
+ addq $16 * SIZE, CO1 # coffset += 16\r
+ decq I # i --\r
+ jg .L2_11\r
+ ALIGN_4 \r
+\r
+/**************************************************************************\r
+* Rest of M \r
+***************************************************************************/\r
+.L2_20:\r
+ // Test rest of M\r
+\r
+ testq $15, M\r
+ jz .L2_60 // to next 3 lines of N\r
+\r
+ testq $8, M \r
+ jz .L2_21pre\r
+ ALIGN_4\r
+\r
+/**************************************************************************/\r
+\r
+.L2_20_1:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $4 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L2_20_6\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $3, %rax // rax = rax * 8 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_20_2:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL8x2_1(xxx)\r
+ KERNEL8x2_2(xxx)\r
+ KERNEL8x2_3(xxx)\r
+ KERNEL8x2_4(xxx)\r
+\r
+ KERNEL8x2_1(xxx)\r
+ KERNEL8x2_2(xxx)\r
+ KERNEL8x2_3(xxx)\r
+ KERNEL8x2_4(xxx)\r
+\r
+ je .L2_20_6\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL8x2_1(xxx)\r
+ KERNEL8x2_2(xxx)\r
+ KERNEL8x2_3(xxx)\r
+ KERNEL8x2_4(xxx)\r
+\r
+ KERNEL8x2_1(xxx)\r
+ KERNEL8x2_2(xxx)\r
+ KERNEL8x2_3(xxx)\r
+ KERNEL8x2_4(xxx)\r
+\r
+ je .L2_20_6\r
+\r
+ jmp .L2_20_2\r
+ ALIGN_4\r
+\r
+.L2_20_6:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L2_20_9\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $3, %rax // rax = rax * 8 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_20_7:\r
+\r
+ KERNEL8x2_SUB(xxx)\r
+ addq $2, BI\r
+ addq $8, %rax\r
+ jl .L2_20_7\r
+ ALIGN_4\r
+\r
+\r
+.L2_20_9:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
+\r
+ vfmaddps (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+ vfmaddps 4 * SIZE(CO1, LDC),%xmm0, %xmm8,%xmm8\r
+\r
+ vmovups %xmm4 , (CO1)\r
+ vmovups %xmm7 , 4 * SIZE(CO1)\r
+\r
+ vmovups %xmm5 , (CO1, LDC)\r
+ vmovups %xmm8 , 4 * SIZE(CO1, LDC)\r
+\r
+\r
+ addq $8 * SIZE, CO1 # coffset += 8\r
+ ALIGN_4\r
+ \r
+\r
+\r
+/**************************************************************************/\r
+\r
+.L2_21pre:\r
+\r
+ testq $4, M \r
+ jz .L2_30\r
+ ALIGN_4\r
+\r
+.L2_21:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $4 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L2_26\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 1 ; number of values\r
+\r
+ salq $2, %rax // rax = rax * 4 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_22:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL4x2_1(xxx)\r
+ KERNEL4x2_2(xxx)\r
+ KERNEL4x2_3(xxx)\r
+ KERNEL4x2_4(xxx)\r
+\r
+ KERNEL4x2_1(xxx)\r
+ KERNEL4x2_2(xxx)\r
+ KERNEL4x2_3(xxx)\r
+ KERNEL4x2_4(xxx)\r
+\r
+ je .L2_26\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL4x2_1(xxx)\r
+ KERNEL4x2_2(xxx)\r
+ KERNEL4x2_3(xxx)\r
+ KERNEL4x2_4(xxx)\r
+\r
+ KERNEL4x2_1(xxx)\r
+ KERNEL4x2_2(xxx)\r
+ KERNEL4x2_3(xxx)\r
+ KERNEL4x2_4(xxx)\r
+\r
+ je .L2_26\r
+\r
+ jmp .L2_22\r
+ ALIGN_4\r
+\r
+.L2_26:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L2_29\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $2, %rax // rax = rax * 4 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_27:\r
+\r
+ KERNEL4x2_SUB(xxx)\r
+ addq $2, BI\r
+ addq $4, %rax\r
+ jl .L2_27\r
+ ALIGN_4\r
+\r
+\r
+.L2_29:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddps (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+\r
+ vmovups %xmm4 , (CO1)\r
+ vmovups %xmm5 , (CO1, LDC)\r
+\r
+\r
+ addq $4 * SIZE, CO1 # coffset += 4\r
+ ALIGN_4\r
+ \r
+\r
+.L2_30:\r
+ testq $2, M \r
+ jz .L2_40\r
+\r
+ ALIGN_4\r
+\r
+.L2_31:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $4 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L2_36\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ salq $1, %rax // rax = rax *2 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_32:\r
+\r
+ prefetcht0 B_PR1(BO,BI,SIZE)\r
+ KERNEL2x2_1(xxx)\r
+ KERNEL2x2_2(xxx)\r
+ KERNEL2x2_3(xxx)\r
+ KERNEL2x2_4(xxx)\r
+\r
+ KERNEL2x2_1(xxx)\r
+ KERNEL2x2_2(xxx)\r
+ KERNEL2x2_3(xxx)\r
+ KERNEL2x2_4(xxx)\r
+\r
+ je .L2_36\r
+\r
+ prefetcht0 B_PR1(BO,BI,SIZE)\r
+ KERNEL2x2_1(xxx)\r
+ KERNEL2x2_2(xxx)\r
+ KERNEL2x2_3(xxx)\r
+ KERNEL2x2_4(xxx)\r
+\r
+ KERNEL2x2_1(xxx)\r
+ KERNEL2x2_2(xxx)\r
+ KERNEL2x2_3(xxx)\r
+ KERNEL2x2_4(xxx)\r
+\r
+ je .L2_36\r
+\r
+ jmp .L2_32\r
+ ALIGN_4\r
+\r
+.L2_36:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L2_39\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+ \r
+ salq $1, %rax // rax = rax *2 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_37:\r
+\r
+ KERNEL2x2_SUB(xxx)\r
+ addq $2, BI\r
+ addq $2, %rax\r
+ jl .L2_37\r
+ ALIGN_4\r
+\r
+\r
+.L2_39:\r
+\r
+ vmovss ALPHA, %xmm0\r
+\r
+ vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddss 1 * SIZE(CO1),%xmm0, %xmm8,%xmm8\r
+ vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+ vfmaddss 1 * SIZE(CO1, LDC),%xmm0, %xmm10,%xmm10\r
+\r
+ vmovss %xmm4 , (CO1)\r
+ vmovss %xmm8 , 1 * SIZE(CO1)\r
+ vmovss %xmm5 , (CO1, LDC)\r
+ vmovss %xmm10, 1 * SIZE(CO1, LDC)\r
+\r
+ addq $2 * SIZE, CO1 # coffset += 2\r
+ ALIGN_4\r
+\r
+.L2_40:\r
+ testq $1, M \r
+ jz .L2_60 // to next 2 lines of N\r
+\r
+ ALIGN_4\r
+\r
+.L2_41:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $4 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L2_46\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_42:\r
+\r
+ KERNEL1x2_1(xxx)\r
+ KERNEL1x2_2(xxx)\r
+ KERNEL1x2_3(xxx)\r
+ KERNEL1x2_4(xxx)\r
+\r
+ KERNEL1x2_1(xxx)\r
+ KERNEL1x2_2(xxx)\r
+ KERNEL1x2_3(xxx)\r
+ KERNEL1x2_4(xxx)\r
+\r
+ je .L2_46\r
+\r
+ KERNEL1x2_1(xxx)\r
+ KERNEL1x2_2(xxx)\r
+ KERNEL1x2_3(xxx)\r
+ KERNEL1x2_4(xxx)\r
+\r
+ KERNEL1x2_1(xxx)\r
+ KERNEL1x2_2(xxx)\r
+ KERNEL1x2_3(xxx)\r
+ KERNEL1x2_4(xxx)\r
+\r
+ je .L2_46\r
+\r
+ jmp .L2_42\r
+ ALIGN_4\r
+\r
+.L2_46:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L2_49\r
+\r
+ movq %rax, BI // Index for BO\r
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values\r
+\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L2_47:\r
+\r
+ KERNEL1x2_SUB(xxx)\r
+ addq $2, BI\r
+ addq $1, %rax\r
+ jl .L2_47\r
+ ALIGN_4\r
+\r
+\r
+.L2_49:\r
+\r
+ vmovss ALPHA, %xmm0\r
+\r
+ vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
+\r
+ vmovss %xmm4 , (CO1)\r
+ vmovss %xmm5 , (CO1, LDC)\r
+\r
+ addq $1 * SIZE, CO1 # coffset += 1\r
+ ALIGN_4\r
+ \r
+.L2_60:\r
+\r
+ decq J // j --\r
+ jg .L2_01 // next 2 lines of N\r
+\r
+\r
+\r
+.L1_0:\r
+\r
+/************************************************************************************************\r
+* Loop for Nmod6 % 2 > 0\r
+*************************************************************************************************/\r
+\r
+ movq Nmod6, J \r
+ andq $1, J // j % 2\r
+ je .L999\r
+ ALIGN_4\r
+\r
+.L1_01:\r
+ // copy to sub buffer\r
+ movq B, BO1\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ movq K, %rax\r
+ ALIGN_4\r
+\r
+.L1_02b:\r
+\r
+ vmovss (BO1), %xmm0\r
+ vmovss %xmm0, (BO)\r
+ addq $1*SIZE,BO1\r
+ addq $1*SIZE,BO\r
+ decq %rax\r
+ jnz .L1_02b\r
+\r
+.L1_02c:\r
+\r
+ movq BO1, B // next offset of B\r
+\r
+.L1_10:\r
+ movq C, CO1\r
+ leaq (C, LDC, 1), C // c += 1 * ldc\r
+\r
+ \r
+ movq A, AO // aoffset = a\r
+ addq $32 * SIZE, AO\r
+\r
+ movq M, I\r
+ sarq $4, I // i = (m >> 4)\r
+ je .L1_20\r
+\r
+ ALIGN_4\r
+\r
+.L1_11:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $2 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax // K = K - ( K % 8 )\r
+ je .L1_16\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $4, %rax // rax = rax * 16 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_12:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL16x1_1(xxx)\r
+ KERNEL16x1_2(xxx)\r
+ KERNEL16x1_3(xxx)\r
+ KERNEL16x1_4(xxx)\r
+\r
+ KERNEL16x1_1(xxx)\r
+ KERNEL16x1_2(xxx)\r
+ KERNEL16x1_3(xxx)\r
+ KERNEL16x1_4(xxx)\r
+\r
+ je .L1_16\r
+\r
+ KERNEL16x1_1(xxx)\r
+ KERNEL16x1_2(xxx)\r
+ KERNEL16x1_3(xxx)\r
+ KERNEL16x1_4(xxx)\r
+\r
+ KERNEL16x1_1(xxx)\r
+ KERNEL16x1_2(xxx)\r
+ KERNEL16x1_3(xxx)\r
+ KERNEL16x1_4(xxx)\r
+\r
+ je .L1_16\r
+\r
+ jmp .L1_12\r
+ ALIGN_4\r
+\r
+.L1_16:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L1_19\r
+\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $4, %rax // rax = rax * 16 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_17:\r
+\r
+ KERNEL16x1_SUB(xxx)\r
+ addq $1, BI\r
+ addq $16, %rax\r
+ jl .L1_17\r
+ ALIGN_4\r
+\r
+\r
+.L1_19:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
+ vfmaddps 8 * SIZE(CO1),%xmm0, %xmm10,%xmm10\r
+ vfmaddps 12 * SIZE(CO1),%xmm0, %xmm13,%xmm13\r
+\r
+ vmovups %xmm4 , (CO1)\r
+ vmovups %xmm7 , 4 * SIZE(CO1)\r
+ vmovups %xmm10, 8 * SIZE(CO1)\r
+ vmovups %xmm13,12 * SIZE(CO1)\r
+\r
+ addq $16 * SIZE, CO1 # coffset += 16\r
+ decq I # i --\r
+ jg .L1_11\r
+ ALIGN_4 \r
+\r
+/**************************************************************************\r
+* Rest of M \r
+***************************************************************************/\r
+.L1_20:\r
+ // Test rest of M\r
+\r
+ testq $15, M\r
+ jz .L999\r
+\r
+ testq $8, M \r
+ jz .L1_21pre\r
+ ALIGN_4\r
+\r
+/**************************************************************************/\r
+\r
+.L1_20_1:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $2 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L1_20_6\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $3, %rax // rax = rax * 8 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_20_2:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL8x1_1(xxx)\r
+ KERNEL8x1_2(xxx)\r
+ KERNEL8x1_3(xxx)\r
+ KERNEL8x1_4(xxx)\r
+\r
+ KERNEL8x1_1(xxx)\r
+ KERNEL8x1_2(xxx)\r
+ KERNEL8x1_3(xxx)\r
+ KERNEL8x1_4(xxx)\r
+\r
+ je .L1_20_6\r
+\r
+ KERNEL8x1_1(xxx)\r
+ KERNEL8x1_2(xxx)\r
+ KERNEL8x1_3(xxx)\r
+ KERNEL8x1_4(xxx)\r
+\r
+ KERNEL8x1_1(xxx)\r
+ KERNEL8x1_2(xxx)\r
+ KERNEL8x1_3(xxx)\r
+ KERNEL8x1_4(xxx)\r
+\r
+ je .L1_20_6\r
+\r
+ jmp .L1_20_2\r
+ ALIGN_4\r
+\r
+.L1_20_6:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L1_20_9\r
+\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $3, %rax // rax = rax * 8 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_20_7:\r
+\r
+ KERNEL8x1_SUB(xxx)\r
+ addq $1, BI\r
+ addq $8, %rax\r
+ jl .L1_20_7\r
+ ALIGN_4\r
+\r
+\r
+.L1_20_9:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddps 4 * SIZE(CO1),%xmm0, %xmm7,%xmm7\r
+\r
+ vmovups %xmm4 , (CO1)\r
+ vmovups %xmm7 , 4 * SIZE(CO1)\r
+\r
+ addq $8 * SIZE, CO1 # coffset += 8\r
+ ALIGN_4\r
+ \r
+\r
+\r
+/**************************************************************************/\r
+\r
+.L1_21pre:\r
+\r
+ testq $4, M \r
+ jz .L1_30\r
+ ALIGN_4\r
+\r
+.L1_21:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $2 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L1_26\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $2, %rax // rax = rax * 4 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_22:\r
+\r
+ prefetcht0 B_PR1(BO,BI, SIZE)\r
+ KERNEL4x1_1(xxx)\r
+ KERNEL4x1_2(xxx)\r
+ KERNEL4x1_3(xxx)\r
+ KERNEL4x1_4(xxx)\r
+\r
+ KERNEL4x1_1(xxx)\r
+ KERNEL4x1_2(xxx)\r
+ KERNEL4x1_3(xxx)\r
+ KERNEL4x1_4(xxx)\r
+\r
+ je .L1_26\r
+\r
+ KERNEL4x1_1(xxx)\r
+ KERNEL4x1_2(xxx)\r
+ KERNEL4x1_3(xxx)\r
+ KERNEL4x1_4(xxx)\r
+\r
+ KERNEL4x1_1(xxx)\r
+ KERNEL4x1_2(xxx)\r
+ KERNEL4x1_3(xxx)\r
+ KERNEL4x1_4(xxx)\r
+\r
+ je .L1_26\r
+\r
+ jmp .L1_22\r
+ ALIGN_4\r
+\r
+.L1_26:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L1_29\r
+\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $2, %rax // rax = rax * 4 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_27:\r
+\r
+ KERNEL4x1_SUB(xxx)\r
+ addq $1, BI\r
+ addq $4, %rax\r
+ jl .L1_27\r
+ ALIGN_4\r
+\r
+\r
+.L1_29:\r
+\r
+ vbroadcastss ALPHA, %xmm0\r
+\r
+ vfmaddps (CO1),%xmm0, %xmm4,%xmm4\r
+\r
+ vmovups %xmm4 , (CO1)\r
+\r
+ addq $4 * SIZE, CO1 # coffset += 4\r
+ ALIGN_4\r
+ \r
+\r
+.L1_30:\r
+ testq $2, M \r
+ jz .L1_40\r
+\r
+ ALIGN_4\r
+\r
+.L1_31:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $2 * SIZE, BO\r
+\r
+ vzeroall\r
+\r
+ movq K, %rax\r
+\r
+ andq $-8, %rax\r
+ je .L1_36\r
+ movq %rax, BI // Index for BO\r
+\r
+ salq $1, %rax // rax = rax *2 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_32:\r
+\r
+ prefetcht0 B_PR1(BO,BI,SIZE)\r
+ KERNEL2x1_1(xxx)\r
+ KERNEL2x1_2(xxx)\r
+ KERNEL2x1_3(xxx)\r
+ KERNEL2x1_4(xxx)\r
+\r
+ KERNEL2x1_1(xxx)\r
+ KERNEL2x1_2(xxx)\r
+ KERNEL2x1_3(xxx)\r
+ KERNEL2x1_4(xxx)\r
+\r
+ je .L1_36\r
+\r
+ KERNEL2x1_1(xxx)\r
+ KERNEL2x1_2(xxx)\r
+ KERNEL2x1_3(xxx)\r
+ KERNEL2x1_4(xxx)\r
+\r
+ KERNEL2x1_1(xxx)\r
+ KERNEL2x1_2(xxx)\r
+ KERNEL2x1_3(xxx)\r
+ KERNEL2x1_4(xxx)\r
+\r
+ je .L1_36\r
+\r
+ jmp .L1_32\r
+ ALIGN_4\r
+\r
+.L1_36:\r
+ movq K, %rax\r
+\r
+ andq $7, %rax # if (k & 1)\r
+ je .L1_39\r
+\r
+ movq %rax, BI // Index for BO\r
+ \r
+ salq $1, %rax // rax = rax *2 ; number of values\r
+ leaq (AO, %rax, SIZE), AO\r
+ leaq (BO, BI, SIZE), BO\r
+ negq BI\r
+ negq %rax\r
+ ALIGN_4\r
+\r
+.L1_37:\r
+\r
+ KERNEL2x1_SUB(xxx)\r
+ addq $1, BI\r
+ addq $2, %rax\r
+ jl .L1_37\r
+ ALIGN_4\r
+\r
+\r
+.L1_39:\r
+\r
+ vmovss ALPHA, %xmm0\r
+\r
+ vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
+ vfmaddss 1 * SIZE(CO1),%xmm0, %xmm8,%xmm8\r
\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $2, KK\r
-#endif\r
+ vmovss %xmm4 , (CO1)\r
+ vmovss %xmm8 , 1 * SIZE(CO1)\r
\r
addq $2 * SIZE, CO1 # coffset += 2\r
ALIGN_4\r
\r
-.L7_40:\r
+.L1_40:\r
testq $1, M \r
- jz .L7_60 // to next 3 lines of N\r
+ jz .L999\r
\r
ALIGN_4\r
\r
-.L7_41:\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER2, BO // second buffer to BO\r
- addq $6 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER2, BO // second buffer to BO\r
- addq $6 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
+.L1_41:\r
+ leaq BUFFER1, BO // first buffer to BO\r
+ addq $2 * SIZE, BO\r
\r
vzeroall\r
\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $1, %rax // number of values in AO\r
-#else\r
- addq $3, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
\r
andq $-8, %rax\r
- je .L7_46\r
+ je .L1_46\r
movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
\r
leaq (AO, %rax, SIZE), AO\r
leaq (BO, BI, SIZE), BO\r
negq %rax\r
ALIGN_4\r
\r
-.L7_42:\r
+.L1_42:\r
\r
- KERNEL1x3_1(xxx)\r
- KERNEL1x3_2(xxx)\r
- KERNEL1x3_3(xxx)\r
- KERNEL1x3_4(xxx)\r
+ KERNEL1x1_1(xxx)\r
+ KERNEL1x1_2(xxx)\r
+ KERNEL1x1_3(xxx)\r
+ KERNEL1x1_4(xxx)\r
\r
- KERNEL1x3_1(xxx)\r
- KERNEL1x3_2(xxx)\r
- KERNEL1x3_3(xxx)\r
- KERNEL1x3_4(xxx)\r
+ KERNEL1x1_1(xxx)\r
+ KERNEL1x1_2(xxx)\r
+ KERNEL1x1_3(xxx)\r
+ KERNEL1x1_4(xxx)\r
\r
- je .L7_46\r
+ je .L1_46\r
\r
- KERNEL1x3_1(xxx)\r
- KERNEL1x3_2(xxx)\r
- KERNEL1x3_3(xxx)\r
- KERNEL1x3_4(xxx)\r
+ KERNEL1x1_1(xxx)\r
+ KERNEL1x1_2(xxx)\r
+ KERNEL1x1_3(xxx)\r
+ KERNEL1x1_4(xxx)\r
\r
- KERNEL1x3_1(xxx)\r
- KERNEL1x3_2(xxx)\r
- KERNEL1x3_3(xxx)\r
- KERNEL1x3_4(xxx)\r
+ KERNEL1x1_1(xxx)\r
+ KERNEL1x1_2(xxx)\r
+ KERNEL1x1_3(xxx)\r
+ KERNEL1x1_4(xxx)\r
\r
- je .L7_46\r
+ je .L1_46\r
\r
- jmp .L7_42\r
+ jmp .L1_42\r
ALIGN_4\r
\r
-.L7_46:\r
-#ifndef TRMMKERNEL\r
+.L1_46:\r
movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
\r
andq $7, %rax # if (k & 1)\r
- je .L7_49\r
+ je .L1_49\r
\r
movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
\r
leaq (AO, %rax, SIZE), AO\r
leaq (BO, BI, SIZE), BO\r
negq %rax\r
ALIGN_4\r
\r
-.L7_47:\r
+.L1_47:\r
\r
- KERNEL1x3_SUB(xxx)\r
- addq $3, BI\r
+ KERNEL1x1_SUB(xxx)\r
+ addq $1, BI\r
addq $1, %rax\r
- jl .L7_47\r
+ jl .L1_47\r
ALIGN_4\r
\r
\r
-.L7_49:\r
+.L1_49:\r
\r
vmovss ALPHA, %xmm0\r
\r
-#ifndef TRMMKERNEL\r
-\r
vfmaddss (CO1),%xmm0, %xmm4,%xmm4\r
- vfmaddss (CO1, LDC),%xmm0, %xmm5,%xmm5\r
- vfmaddss (CO1, LDC, 2),%xmm0, %xmm6,%xmm6\r
\r
-#else\r
- vmulss %xmm0, %xmm4,%xmm4\r
- vmulss %xmm0, %xmm5,%xmm5\r
- vmulss %xmm0, %xmm6,%xmm6\r
+ vmovss %xmm4 , (CO1)\r
+\r
+ addq $1 * SIZE, CO1 # coffset += 1\r
+ ALIGN_4\r
+ \r
+\r
+.L999:\r
+ movq SP, %rsp\r
+ movq (%rsp), %rbx\r
+ movq 8(%rsp), %rbp\r
+ movq 16(%rsp), %r12\r
+ movq 24(%rsp), %r13\r
+ movq 32(%rsp), %r14\r
+ movq 40(%rsp), %r15\r
\r
+#ifdef WINDOWS_ABI\r
+ movq 48(%rsp), %rdi\r
+ movq 56(%rsp), %rsi\r
+ movups 64(%rsp), %xmm6\r
+ movups 80(%rsp), %xmm7\r
+ movups 96(%rsp), %xmm8\r
+ movups 112(%rsp), %xmm9\r
+ movups 128(%rsp), %xmm10\r
+ movups 144(%rsp), %xmm11\r
+ movups 160(%rsp), %xmm12\r
+ movups 176(%rsp), %xmm13\r
+ movups 192(%rsp), %xmm14\r
+ movups 208(%rsp), %xmm15\r
#endif\r
\r
- vmovss %xmm4 , (CO1)\r
- vmovss %xmm5 , (CO1, LDC)\r
- vmovss %xmm6 , (CO1, LDC, 2)\r
+ addq $STACKSIZE, %rsp\r
+ ret\r
\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax \r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq (BI,BI,2), BI // BI = BI * 3 ; number of values\r
- leaq (BO, BI, SIZE), BO \r
- leaq (AO, %rax, SIZE), AO\r
-#endif \r
+ EPILOGUE\r
\r
\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $1, KK\r
-#endif\r
+#else\r
+/*************************************************************************************\r
+* TRMM Kernel\r
+*************************************************************************************/\r
\r
- addq $1 * SIZE, CO1 # coffset += 1\r
- ALIGN_4\r
+\r
+ PROLOGUE\r
+ PROFCODE\r
\r
+ subq $STACKSIZE, %rsp\r
+ movq %rbx, (%rsp)\r
+ movq %rbp, 8(%rsp)\r
+ movq %r12, 16(%rsp)\r
+ movq %r13, 24(%rsp)\r
+ movq %r14, 32(%rsp)\r
+ movq %r15, 40(%rsp)\r
\r
+ vzeroupper\r
\r
-.L7_60:\r
+#ifdef WINDOWS_ABI\r
+ movq %rdi, 48(%rsp)\r
+ movq %rsi, 56(%rsp)\r
+ movups %xmm6, 64(%rsp)\r
+ movups %xmm7, 80(%rsp)\r
+ movups %xmm8, 96(%rsp)\r
+ movups %xmm9, 112(%rsp)\r
+ movups %xmm10, 128(%rsp)\r
+ movups %xmm11, 144(%rsp)\r
+ movups %xmm12, 160(%rsp)\r
+ movups %xmm13, 176(%rsp)\r
+ movups %xmm14, 192(%rsp)\r
+ movups %xmm15, 208(%rsp)\r
\r
- decq J // j --\r
- jg .L6_01\r
+ movq ARG1, OLD_M\r
+ movq ARG2, OLD_N\r
+ movq ARG3, OLD_K\r
+ movq OLD_A, A\r
+ movq OLD_B, B\r
+ movq OLD_C, C\r
+ movq OLD_LDC, LDC\r
+#ifdef TRMMKERNEL\r
+ movsd OLD_OFFSET, %xmm12\r
+#endif\r
+ vmovaps %xmm3, %xmm0\r
+\r
+#else\r
+ movq STACKSIZE + 8(%rsp), LDC\r
+#ifdef TRMMKERNEL\r
+ movsd STACKSIZE + 16(%rsp), %xmm12\r
+#endif\r
\r
+#endif\r
\r
-.L2_0:\r
- cmpq $0, Nmod6 // N % 6 == 0\r
+ movq %rsp, SP # save old stack\r
+ subq $128 + L_BUFFER_SIZE, %rsp\r
+ andq $-4096, %rsp # align stack\r
+\r
+ STACK_TOUCH\r
+\r
+ cmpq $0, OLD_M\r
je .L999\r
\r
-/************************************************************************************************\r
-* Loop for Nmod6 / 2 > 0\r
-*************************************************************************************************/\r
+ cmpq $0, OLD_N\r
+ je .L999\r
\r
- movq Nmod6, J \r
- sarq $1, J // j = j / 2\r
+ cmpq $0, OLD_K\r
+ je .L999\r
+\r
+ movq OLD_M, M\r
+ movq OLD_N, N\r
+ movq OLD_K, K\r
+\r
+ vmovsd %xmm0, ALPHA\r
+\r
+ salq $BASE_SHIFT, LDC\r
+\r
+ movq N, %rax\r
+ xorq %rdx, %rdx\r
+ movq $2, %rdi\r
+ divq %rdi // N / 6\r
+ movq %rax, Ndiv6 // N / 6\r
+ movq %rdx, Nmod6 // N % 6\r
+\r
+ \r
+\r
+#ifdef TRMMKERNEL\r
+ vmovsd %xmm12, OFFSET\r
+ vmovsd %xmm12, KK\r
+#ifndef LEFT\r
+ negq KK\r
+#endif \r
+#endif\r
+\r
+ movq Ndiv6, J\r
+ cmpq $0, J\r
je .L1_0\r
ALIGN_4\r
\r
\r
\r
.L2_60:\r
+#if defined(TRMMKERNEL) && !defined(LEFT)\r
+ addq $2, KK\r
+#endif\r
\r
decq J // j --\r
jg .L2_01 // next 2 lines of N\r
ret\r
\r
EPILOGUE\r
+\r
+\r
+\r
+\r
+\r
+#endif\r