diff --git a/control/context.c b/control/context.c
index 38d8920367593e569a6ed79ab4ca010447f4da1b..91b5434aa2035aac1cdc521aaaa4e86c17bfa084 100644
--- a/control/context.c
+++ b/control/context.c
@@ -83,6 +83,7 @@ MORSE_context_t *morse_context_create()
morse->parallel_enabled = MORSE_FALSE;
morse->profiling_enabled = MORSE_FALSE;
morse->progress_enabled = MORSE_FALSE;
+ morse->gemm3m_enabled = MORSE_FALSE;
morse->householder = MORSE_FLAT_HOUSEHOLDER;
morse->translation = MORSE_OUTOFPLACE;
@@ -132,6 +133,7 @@ int morse_context_destroy(){
* @arg MORSE_AUTOTUNING autotuning for tile size and inner block size.
* @arg MORSE_PROFILING_MODE activate profiling of kernels
* @arg MORSE_PROGRESS activate progress indicator
+ * @arg MORSE_GEMM3M Use z/cgemm3m for complexe matrix-matrix products
*
*******************************************************************************
*
@@ -166,6 +168,13 @@ int MORSE_Enable(MORSE_enum option)
case MORSE_PROGRESS:
morse->progress_enabled = MORSE_TRUE;
break;
+ case MORSE_GEMM3M:
+#ifdef CBLAS_HAS_ZGEMM3M
+ morse->gemm3m_enabled = MORSE_TRUE;
+#else
+ morse_error("MORSE_Enable", "cannot enable GEMM3M (not available in cblas)");
+#endif
+ break;
/* case MORSE_PARALLEL: */
/* morse->parallel_enabled = MORSE_TRUE; */
/* break; */
@@ -197,6 +206,7 @@ int MORSE_Enable(MORSE_enum option)
* @arg MORSE_AUTOTUNING autotuning for tile size and inner block size.
* @arg MORSE_PROFILING_MODE deactivate profiling of kernels
* @arg MORSE_PROGRESS deactivate progress indicator
+ * @arg MORSE_GEMM3M Use z/cgemm3m for complexe matrix-matrix products
*
*******************************************************************************
*
@@ -230,6 +240,9 @@ int MORSE_Disable(MORSE_enum option)
case MORSE_PROGRESS:
morse->progress_enabled = MORSE_FALSE;
break;
+ case MORSE_GEMM3M:
+ morse->gemm3m_enabled = MORSE_FALSE;
+ break;
case MORSE_PARALLEL_MODE:
morse->parallel_enabled = MORSE_FALSE;
break;
diff --git a/coreblas/compute/core_zgemm.c b/coreblas/compute/core_zgemm.c
index fe98b80a993a666c06b1e29a4d2156da6328df38..b5b7e4d4bc37ac0f8ac8d0203d3a91b2060f5182 100644
--- a/coreblas/compute/core_zgemm.c
+++ b/coreblas/compute/core_zgemm.c
@@ -42,6 +42,18 @@ void CORE_zgemm(MORSE_enum transA, int transB,
const MORSE_Complex64_t *B, int LDB,
MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
{
+#ifdef CBLAS_HAS_ZGEMM3M
+ MORSE_context_t *morse = morse_context_self();
+ if (morse->gemm3m_enabled)
+ cblas_zgemm3m(
+ CblasColMajor,
+ (CBLAS_TRANSPOSE)transA, (CBLAS_TRANSPOSE)transB,
+ M, N, K,
+ CBLAS_SADDR(alpha), A, LDA,
+ B, LDB,
+ CBLAS_SADDR(beta), C, LDC);
+ else
+#endif
cblas_zgemm(
CblasColMajor,
(CBLAS_TRANSPOSE)transA, (CBLAS_TRANSPOSE)transB,
diff --git a/include/morse_constants.h b/include/morse_constants.h
index 39e3c460535c3a6a8a5fab2b9020b181cc9e9ea8..8ffe90dc0868c8d574dd5fde01cc92fe63724fba 100644
--- a/include/morse_constants.h
+++ b/include/morse_constants.h
@@ -131,6 +131,7 @@
#define MORSE_PARALLEL_MODE 6
#define MORSE_BOUND 7
#define MORSE_PROGRESS 8
+#define MORSE_GEMM3M 9
/** ****************************************************************************
* MORSE constants - configuration parameters
diff --git a/include/morse_struct.h b/include/morse_struct.h
index 2958ce3d021bc68e2b5a91195b1b8dd7580a43b0..397d96e1e332171a8e2eef5eb7eff4073b43ca52 100644
--- a/include/morse_struct.h
+++ b/include/morse_struct.h
@@ -132,6 +132,7 @@ typedef struct morse_context_s {
MORSE_bool parallel_enabled;
MORSE_bool profiling_enabled;
MORSE_bool progress_enabled;
+ MORSE_bool gemm3m_enabled;
MORSE_enum householder; // "domino" (flat) or tree-based (reduction) Householder
MORSE_enum translation; // In place or Out of place layout conversion
diff --git a/timing/timing.c b/timing/timing.c
index 796441d6bf9adfcedb0f5500a7c891f7df186c5b..1c83dbd161498577226a1ea55e3cc15d544ed800 100644
--- a/timing/timing.c
+++ b/timing/timing.c
@@ -349,6 +349,7 @@ show_help(char *prog_name) {
" --[a]sync Enable/Disable synchronous calls in wrapper function such as POTRI. (default: async)\n"
" --[no]check Check result (default: nocheck)\n"
" --[no]progress Display progress indicator (default: noprogress)\n"
+ " --[no]gemm3m Use gemm3m complex method (default: nogemm3m)\n"
" --[no]inv Check on inverse (default: noinv)\n"
" --[no]warmup Perform a warmup run to pre-load libraries (default: warmup)\n"
" --[no]trace Enable/Disable trace generation (default: notrace)\n"
@@ -487,6 +488,7 @@ main(int argc, char *argv[]) {
iparam[IPARAM_NMPI ] = 1;
iparam[IPARAM_P ] = 1;
iparam[IPARAM_Q ] = 1;
+ iparam[IPARAM_GEMM3M ] = 0;
iparam[IPARAM_PROGRESS ] = 0;
iparam[IPARAM_PROFILE ] = 0;
iparam[IPARAM_PRINT_ERRORS ] = 0;
@@ -526,6 +528,10 @@ main(int argc, char *argv[]) {
iparam[IPARAM_TRACE] = 1;
} else if (startswith( argv[i], "--notrace" )) {
iparam[IPARAM_TRACE] = 0;
+ } else if (startswith( argv[i], "--gemm3m" )) {
+ iparam[IPARAM_GEMM3M] = 1;
+ } else if (startswith( argv[i], "--nogemm3m" )) {
+ iparam[IPARAM_GEMM3M] = 0;
} else if (startswith( argv[i], "--progress" )) {
iparam[IPARAM_PROGRESS] = 1;
} else if (startswith( argv[i], "--noprogress" )) {
@@ -637,6 +643,9 @@ main(int argc, char *argv[]) {
if (iparam[IPARAM_PROGRESS] == 1)
MORSE_Enable(MORSE_PROGRESS);
+ if (iparam[IPARAM_GEMM3M] == 1)
+ MORSE_Enable(MORSE_GEMM3M);
+
#if defined(CHAMELEON_USE_MPI)
MORSE_Comm_size( &nbnode );
iparam[IPARAM_NMPI] = nbnode;
diff --git a/timing/timing.h b/timing/timing.h
index 4f8771bb6be45316752cf6649eeb989fb508c380..a1636491c6664b49868265fb622b623b449649a9 100644
--- a/timing/timing.h
+++ b/timing/timing.h
@@ -48,6 +48,7 @@ enum iparam_timing {
IPARAM_Q, /* Parameter for 2D cyclic distribution */
IPARAM_PROGRESS, /* Use a progress indicator during computations */
+ IPARAM_GEMM3M, /* Use GEMM3M for complex matrix vector products */
/* Added for StarPU version */
IPARAM_PROFILE,
IPARAM_PRINT_ERRORS,