Use cmake to build OpenBLAS GENERIC Target on MSVC x86 64-bit.

Disable CBLAS and LAPACK.

6 files changed, 77 insertions, 30 deletions

diff --git a/interface/CMakeLists.txt b/interface/CMakeLists.txt
index ae949235b..91565d2f2 100644
--- a/interface/CMakeLists.txt
+++ b/interface/CMakeLists.txt
@@ -1,13 +1,16 @@
 
 include_directories(${CMAKE_SOURCE_DIR})
 
+
 set(BLAS1_SOURCES
   copy.c
-  asum.c nrm2.c
+  nrm2.c
 )
 
 set(BLAS1_REAL_ONLY_SOURCES
   rotm.c rotmg.c # N.B. these do not have complex counterparts
+  rot.c
+  asum.c
 )
 
 # these will have 'z' prepended for the complex version
@@ -15,7 +18,7 @@ set(BLAS1_MANGLED_SOURCES
   axpy.c swap.c
   scal.c
   dot.c
-  rot.c rotg.c
+  rotg.c
   axpby.c
 )
 
@@ -31,6 +34,13 @@ set(BLAS2_SOURCES
   tpsv.c tpmv.c
 )
 
+set(BLAS2_COMPLEX_ONLY_MANGLED_SOURCES
+  hemv.c hbmv.c
+  her.c her2.c
+  hpmv.c hpr.c
+  hpr2.c
+)
+
 # these do not have separate 'z' sources
 set(BLAS3_SOURCES
   gemm.c symm.c
@@ -39,6 +49,7 @@ set(BLAS3_SOURCES
 
 set(BLAS3_MANGLED_SOURCES
   omatcopy.c imatcopy.c
+  geadd.c
 )
 
 # generate the BLAS objs once with and once without cblas
@@ -65,9 +76,14 @@ foreach (CBLAS_FLAG ${CBLAS_FLAGS})
   GenerateNamedObjects("${BLAS1_REAL_ONLY_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 1)
   GenerateNamedObjects("${BLAS1_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
   GenerateNamedObjects("${BLAS2_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
+  GenerateNamedObjects("${BLAS2_COMPLEX_ONLY_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false 4)
   GenerateNamedObjects("${BLAS3_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${DISABLE_COMPLEX})
   GenerateNamedObjects("${BLAS3_MANGLED_SOURCES}" "" "" ${CBLAS_FLAG} "" "" false ${MANGLE_COMPLEX})
 
+  #sdsdot, dsdot
+  GenerateNamedObjects("sdsdot.c" "" "sdsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
+  GenerateNamedObjects("dsdot.c" "" "dsdot" ${CBLAS_FLAG} "" "" true "SINGLE")
+
   # trmm is trsm with a compiler flag set
   GenerateNamedObjects("trsm.c" "TRMM" "trmm" ${CBLAS_FLAG})
 
@@ -86,17 +102,36 @@ endforeach ()
 
 # complex-specific sources
 foreach (float_type ${FLOAT_TYPES})
+
   if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
     GenerateNamedObjects("zger.c" "" "geru" false "" "" false ${float_type})
     GenerateNamedObjects("zger.c" "CONJ" "gerc" false "" "" false ${float_type})
+    GenerateNamedObjects("zdot.c" "CONJ" "dotc" false "" "" false ${float_type})
+    GenerateNamedObjects("zdot.c" "" "dotu" false "" "" false ${float_type})
+
+    GenerateNamedObjects("symm.c" "HEMM" "hemm" false "" "" false ${float_type})
+    GenerateNamedObjects("syrk.c" "HEMM" "herk" false "" "" false ${float_type})
+    GenerateNamedObjects("syr2k.c" "HEMM" "her2k" false "" "" false ${float_type})
+
+    if (USE_GEMM3M)
+      GenerateNamedObjects("gemm.c" "GEMM3M" "gemm3m" false "" "" false ${float_type})
+    endif()
   endif ()
   if (${float_type} STREQUAL "COMPLEX")
     GenerateNamedObjects("zscal.c" "SSCAL" "sscal" false "" "" false "COMPLEX")
     GenerateNamedObjects("nrm2.c" "" "scnrm2" false "" "" true "COMPLEX")
+    GenerateNamedObjects("zrot.c" "" "csrot" false "" "" true "COMPLEX")
+    GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "scamin" false "" "" true "COMPLEX")
+    GenerateNamedObjects("max.c" "USE_ABS" "scamax" false "" "" true "COMPLEX")
+    GenerateNamedObjects("asum.c" "" "scasum" false "" "" true "COMPLEX")
   endif ()
   if (${float_type} STREQUAL "ZCOMPLEX")
     GenerateNamedObjects("zscal.c" "SSCAL" "dscal" false "" "" false "ZCOMPLEX")
     GenerateNamedObjects("nrm2.c" "" "dznrm2" false "" "" true "ZCOMPLEX")
+    GenerateNamedObjects("zrot.c" "" "zdrot" false "" "" true "ZCOMPLEX")
+    GenerateNamedObjects("max.c" "USE_ABS;USE_MIN" "dzamin" false "" "" true "ZCOMPLEX")
+    GenerateNamedObjects("max.c" "USE_ABS" "dzamax" false "" "" true "ZCOMPLEX")
+    GenerateNamedObjects("asum.c" "" "dzasum" false "" "" true "ZCOMPLEX")
   endif ()
 endforeach ()
 
diff --git a/interface/rotg.c b/interface/rotg.c
index 49088ab02..a0e6efdab 100644
--- a/interface/rotg.c
+++ b/interface/rotg.c
@@ -14,8 +14,7 @@ void CNAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
 
 #endif
 
-
-#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__) || defined(_M_X64) || defined(_M_IX86)
 
   long double da = *DA;
   long double db = *DB;
diff --git a/interface/zaxpby.c b/interface/zaxpby.c
index 9e8324432..1abb24de9 100644
--- a/interface/zaxpby.c
+++ b/interface/zaxpby.c
@@ -53,13 +53,13 @@ void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *BETA, FLOAT *
 
 #endif
 
-  if (n <= 0) return;
-
   FLOAT alpha_r = *(ALPHA + 0);
   FLOAT alpha_i = *(ALPHA + 1);
   FLOAT beta_r  = *(BETA  + 0);
   FLOAT beta_i  = *(BETA  + 1);
 
+  if (n <= 0) return;
+
   FUNCTION_PROFILE_START();
 
   if (incx < 0) x -= (n - 1) * incx * 2;
diff --git a/interface/zdot.c b/interface/zdot.c
index 1380ce292..34dfb731a 100644
--- a/interface/zdot.c
+++ b/interface/zdot.c
@@ -57,21 +57,25 @@
 #ifdef RETURN_BY_STRUCT
 MYTYPE         NAME(                        blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY) {
 #elif defined RETURN_BY_STACK
-void           NAME(FLOAT _Complex *result, blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY) {
+void           NAME(OPENBLAS_COMPLEX_FLOAT *result, blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY) {
 #else
-FLOAT _Complex NAME(                        blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY) {
+OPENBLAS_COMPLEX_FLOAT NAME(                        blasint *N, FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY) {
 #endif
 
   BLASLONG n    = *N;
   BLASLONG incx = *INCX;
   BLASLONG incy = *INCY;
 #ifndef RETURN_BY_STACK
-  FLOAT _Complex ret;
+  OPENBLAS_COMPLEX_FLOAT ret;
 #endif
 #ifdef RETURN_BY_STRUCT
   MYTYPE  myret;
 #endif
 
+#ifndef RETURN_BY_STRUCT
+  OPENBLAS_COMPLEX_FLOAT zero=OPENBLAS_MAKE_COMPLEX_FLOAT(0.0, 0.0);
+#endif
+
   PRINT_DEBUG_NAME;
 
   if (n <= 0) {
@@ -80,10 +84,10 @@ FLOAT _Complex NAME(                        blasint *N, FLOAT *x, blasint *INCX,
     myret.i = 0.;
     return myret;
 #elif defined RETURN_BY_STACK
-    *result = ZERO;
+    *result = zero;
     return;
 #else
-    return ZERO;
+    return zero;
 #endif
   }
 
@@ -144,21 +148,21 @@ FLOAT _Complex NAME(                        blasint *N, FLOAT *x, blasint *INCX,
 #else
 
 #ifdef FORCE_USE_STACK
-void           CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy, FLOAT _Complex *result){
+void           CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy, OPENBLAS_COMPLEX_FLOAT *result){
 #else
-FLOAT _Complex CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
+OPENBLAS_COMPLEX_FLOAT CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
 
-  FLOAT _Complex ret;
+  OPENBLAS_COMPLEX_FLOAT ret;
 #endif
 
   PRINT_DEBUG_CNAME;
 
   if (n <= 0) {
 #ifdef FORCE_USE_STACK
-    *result = ZERO;
+    *result = OPENBLAS_MAKE_COMPLEX_FLOAT(0.0, 0.0);
     return;
 #else
-    return ZERO;
+    return OPENBLAS_MAKE_COMPLEX_FLOAT(0.0, 0.0);
 #endif
   }
 
diff --git a/interface/zgemv.c b/interface/zgemv.c
index 704034aaf..792f799e5 100644
--- a/interface/zgemv.c
+++ b/interface/zgemv.c
@@ -79,6 +79,9 @@ void NAME(char *TRANS, blasint *M, blasint *N,
   FLOAT *buffer;
 #ifdef SMP
   int nthreads;
+  int nthreads_max;
+  int nthreads_avail;
+  double MNK;
 #endif
 
   int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
@@ -91,14 +94,14 @@ void NAME(char *TRANS, blasint *M, blasint *N,
   blasint    lenx, leny;
   blasint    i;
 
-  PRINT_DEBUG_NAME;
-
   FLOAT alpha_r = *(ALPHA + 0);
   FLOAT alpha_i = *(ALPHA + 1);
 
   FLOAT beta_r  = *(BETA + 0);
   FLOAT beta_i  = *(BETA + 1);
 
+  PRINT_DEBUG_NAME;
+
   TOUPPER(trans);
 
   info = 0;
@@ -153,14 +156,14 @@ void CNAME(enum CBLAS_ORDER order,
 	      GEMV_O, GEMV_U, GEMV_S, GEMV_D,
 	    };
 
-  PRINT_DEBUG_CNAME;
-
   FLOAT alpha_r = *(ALPHA + 0);
   FLOAT alpha_i = *(ALPHA + 1);
 
   FLOAT beta_r  = *(BETA + 0);
   FLOAT beta_i  = *(BETA + 1);
 
+  PRINT_DEBUG_CNAME;
+
   trans = -1;
   info  =  0;
 
@@ -234,10 +237,10 @@ void CNAME(enum CBLAS_ORDER order,
 
 #ifdef SMP
 
-  int  nthreads_max = num_cpu_avail(2);
-  int  nthreads_avail = nthreads_max;
+  nthreads_max = num_cpu_avail(2);
+  nthreads_avail = nthreads_max;
 
-  double MNK = (double) m * (double) n;
+  MNK = (double) m * (double) n;
   if ( MNK <= ( 256.0  * (double) (GEMM_MULTITHREAD_THRESHOLD * GEMM_MULTITHREAD_THRESHOLD)  ))
         nthreads_max = 1;
 
diff --git a/interface/zrotg.c b/interface/zrotg.c
index e9e8a11df..187343d41 100644
--- a/interface/zrotg.c
+++ b/interface/zrotg.c
@@ -6,13 +6,7 @@
 
 void NAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
 
-  PRINT_DEBUG_NAME;
-
-  IDEBUG_START;
-
-  FUNCTION_PROFILE_START();
-
-#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__)
+#if defined(__i386__) || defined(__x86_64__) || defined(__ia64__) || defined(_M_X64) || defined(_M_IX86)
 
   long double da_r = *(DA + 0);
   long double da_i = *(DA + 1);
@@ -22,6 +16,12 @@ void NAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
 
   long double ada = fabs(da_r) + fabs(da_i);
 
+  PRINT_DEBUG_NAME;
+
+  IDEBUG_START;
+
+  FUNCTION_PROFILE_START();
+
   if (ada == ZERO) {
     *C        = ZERO;
     *(S  + 0) = ONE;
@@ -54,6 +54,12 @@ void NAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
   FLOAT ada = fabs(da_r) + fabs(da_i);
   FLOAT adb;
 
+  PRINT_DEBUG_NAME;
+
+  IDEBUG_START;
+
+  FUNCTION_PROFILE_START();
+
   if (ada == ZERO) {
     *C        = ZERO;
     *(S  + 0) = ONE;