New dgemm kernel for BULLDOZER: dgemm_kernel_8x2_bulldozer.S

driver/level3/level3.c

@@ -332,7 +332,20 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 #else
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+
+#if defined(BULLDOZER) && defined(ARCH_X86_64) && defined(DOUBLE) && !defined(COMPLEX)
+        if (min_jj >= 12*GEMM_UNROLL_N) min_jj = 12*GEMM_UNROLL_N;
+        else
+                if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
+                else
+                        if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
+                        else
+                                if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+#else
+
+        if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+#endif
+
 	
 	START_RPCC();
 	

driver/level3/level3_thread.c

@@ -360,7 +360,19 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
       for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){
 	min_jj = MIN(n_to, xxx + div_n) - jjs;
+
+#if defined(BULLDOZER) && defined(ARCH_X86_64) && defined(DOUBLE) && !defined(COMPLEX)
+	if (min_jj >= 12*GEMM_UNROLL_N) min_jj = 12*GEMM_UNROLL_N;
+	else
+		if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
+		else
+			if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
+			else
+				if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+#else
+
 	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+#endif	
 
 	START_RPCC();
 	

kernel/x86_64/KERNEL.BULLDOZER

@@ -10,13 +10,13 @@ SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
 SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX) 
 SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
 SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
-DGEMMKERNEL    =  dgemm_kernel_4x4_bulldozer.S
+DGEMMKERNEL    =  dgemm_kernel_8x2_bulldozer.S
-DGEMMINCOPY    =
+DGEMMINCOPY    =  ../generic/gemm_ncopy_8.c
-DGEMMITCOPY    =
+DGEMMITCOPY    =  ../generic/gemm_tcopy_8.c
-DGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
+DGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
-DGEMMOTCOPY    =  ../generic/gemm_tcopy_4.c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
-DGEMMINCOPYOBJ =
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
-DGEMMITCOPYOBJ =
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
 DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
 DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMMKERNEL    =  zgemm_kernel_4x2_barcelona.S
@@ -38,25 +38,27 @@ ZGEMMITCOPYOBJ =
 ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
 ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 
-STRSMKERNEL_LN	=  trsm_kernel_LN_8x4_sse.S
-STRSMKERNEL_LT	=  trsm_kernel_LT_8x4_sse.S
-STRSMKERNEL_RN	=  trsm_kernel_LT_8x4_sse.S
-STRSMKERNEL_RT	=  trsm_kernel_RT_8x4_sse.S
-
-DTRSMKERNEL_LN	=  trsm_kernel_LN_4x4_barcelona.S
-DTRSMKERNEL_LT	=  trsm_kernel_LT_4x4_barcelona.S
-DTRSMKERNEL_RN	=  trsm_kernel_LT_4x4_barcelona.S
-DTRSMKERNEL_RT	=  trsm_kernel_RT_4x4_barcelona.S
-
-CTRSMKERNEL_LN	=  ztrsm_kernel_LN_4x2_sse.S
-CTRSMKERNEL_LT	=  ztrsm_kernel_LT_4x2_sse.S
-CTRSMKERNEL_RN	=  ztrsm_kernel_LT_4x2_sse.S
-CTRSMKERNEL_RT	=  ztrsm_kernel_RT_4x2_sse.S
-
-ZTRSMKERNEL_LN	=  ztrsm_kernel_LN_2x2_sse2.S
-ZTRSMKERNEL_LT	=  ztrsm_kernel_LT_2x2_sse2.S
-ZTRSMKERNEL_RN	=  ztrsm_kernel_LT_2x2_sse2.S
-ZTRSMKERNEL_RT	=  ztrsm_kernel_RT_2x2_sse2.S
-
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
+
+STRSMKERNEL_LN  =  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT  =  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN  =  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT  =  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+
+

param.h

@@ -153,7 +153,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define GEMM_DEFAULT_ALIGN 0x0fffUL
 
 #define SGEMM_DEFAULT_UNROLL_N 4
+#if defined(BULLDOZER) && defined(ARCH_X86_64) && !defined(COMPLEX)
+#define DGEMM_DEFAULT_UNROLL_N 2
+#else
 #define DGEMM_DEFAULT_UNROLL_N 4
+#endif
 #define QGEMM_DEFAULT_UNROLL_N 2
 #define CGEMM_DEFAULT_UNROLL_N 2
 #define ZGEMM_DEFAULT_UNROLL_N 2
@@ -161,14 +165,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifdef ARCH_X86
 #define SGEMM_DEFAULT_UNROLL_M 4
-#define DGEMM_DEFAULT_UNROLL_M 2
+#define DGEMM_DEFAULT_UNROLL_M 4
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 2
 #define ZGEMM_DEFAULT_UNROLL_M 1
 #define XGEMM_DEFAULT_UNROLL_M 1
 #else
 #define SGEMM_DEFAULT_UNROLL_M 8
+#if defined(BULLDOZER) && !defined(COMPLEX)
+#define DGEMM_DEFAULT_UNROLL_M 8
+#else
 #define DGEMM_DEFAULT_UNROLL_M 4
+#endif
 #define QGEMM_DEFAULT_UNROLL_M 2
 #define CGEMM_DEFAULT_UNROLL_M 4
 #define ZGEMM_DEFAULT_UNROLL_M 2
@@ -193,26 +201,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 
 #define SGEMM_DEFAULT_P 448
-
+#if defined(BULLDOZER) && defined(ARCH_X86_64) && !defined(COMPLEX)
-#if defined(BULLDOZER) && defined(ARCH_X86_64)
+#define DGEMM_DEFAULT_P 384
-#define DGEMM_DEFAULT_P 248
 #else
 #define DGEMM_DEFAULT_P 224
 #endif
-
 #define QGEMM_DEFAULT_P 112
 #define CGEMM_DEFAULT_P 224
 #define ZGEMM_DEFAULT_P 112 
 #define XGEMM_DEFAULT_P  56
 
 #define SGEMM_DEFAULT_Q 224
-
+#if defined(BULLDOZER) && defined(ARCH_X86_64) && !defined(COMPLEX)
-#if defined(BULLDOZER) && defined(ARCH_X86_64)
+#define DGEMM_DEFAULT_Q 168
-#define DGEMM_DEFAULT_Q 248
 #else
 #define DGEMM_DEFAULT_Q 224
 #endif
-
 #define QGEMM_DEFAULT_Q 224
 #define CGEMM_DEFAULT_Q 224
 #define ZGEMM_DEFAULT_Q 224
@@ -230,7 +234,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SYMV_P	16
 #define HAVE_EXCLUSIVE_CACHE
 
-#define GEMM_THREAD gemm_thread_mn
+#define GEMM_THREAD gemm_thread_m
 
 #endif