Single precision kernel for AVX512 real block4

.gitlab-ci.yml

@@ -2167,6 +2167,19 @@ intel-double-precision-mpi-noopenmp-ftimings-redirect-real-avx512_block4-complex
     - export LD_LIBRARY_PATH=$MKL_HOME/lib/intel64:$LD_LIBRARY_PATH
     -  make check TEST_FLAGS='1000 500 128'
 
+intel-single-precision-mpi-noopenmp-ftimings-redirect-real-avx512_block4-complex-avx512_block2-kernel-jobs:
+  tags:
+    - KNL
+  script:
+    - ./autogen.sh
+    - ./configure FC=mpiifort CC=mpiicc CFLAGS="-O3 -mtune=knl -axMIC-AVX512" FCFLAGS="-O3 -mtune=knl -axMIC-AVX512" SCALAPACK_FCFLAGS="-L$MKLROOT/lib/intel64 -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_intelmpi_lp64 -lpthread -lm -I$MKLROOT/include/intel64/lp64" SCALAPACK_LDFLAGS="-L$MKLROOT/lib/intel64 -lmkl_scalapack_lp64 -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lmkl_blacs_intelmpi_lp64 -lpthread -lm -Wl,-rpath,$MKLROOT/lib/intel64" --with-real-avx512_block4-kernel-only  --with-complex-avx512_block2-kernel-only --enable-single-precision
+    - /home/elpa/wait_until_midnight.sh
+    - make -j 8
+    - export OMP_NUM_THREADS=1
+    - export LD_LIBRARY_PATH=$MKL_HOME/lib/intel64:$LD_LIBRARY_PATH
+    -  make check TEST_FLAGS='1000 500 128'
+
+
 intel-double-precision-mpi-noopenmp-ftimings-redirect-real-avx512_block6-complex-avx512_block1-kernel-jobs:
   tags:
     - KNL

Makefile.am

@@ -174,9 +174,9 @@ endif
 
 if WITH_REAL_AVX512_BLOCK4_KERNEL
   libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_4hv_double_precision.c
-#if WANT_SINGLE_PRECISION_REAL
-#  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_4hv_single_precision.c
-#endif
+if WANT_SINGLE_PRECISION_REAL
+  libelpa@SUFFIX@_private_la_SOURCES += src/elpa2_kernels/elpa2_kernels_real_avx512_4hv_single_precision.c
+endif
 endif
 
 

src/mod_compute_hh_trafo_real.F90

@@ -1241,52 +1241,52 @@ module compute_hh_trafo_real
 #endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
 #endif /* WITH_REAL_AVX_BLOCK4_KERNEL || WITH_REAL_AVX2_BLOCK4_KERNEL */
 
-!#if defined(WITH_REAL_AVX512_BLOCK4_KERNEL)
-!#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
-!           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4) then
-!#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
-!
-!#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX512_BLOCK6_KERNEL))
-!             ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
-!             do j = ncols, 4, -4
-!               w(:,1) = bcast_buffer(1:nbw,j+off)
-!               w(:,2) = bcast_buffer(1:nbw,j+off-1)
-!               w(:,3) = bcast_buffer(1:nbw,j+off-2)
-!               w(:,4) = bcast_buffer(1:nbw,j+off-3)
-!#ifdef WITH_OPENMP
-!               call quad_hh_trafo_real_avx512_4hv_single(c_loc(a(1,j+off+a_off-3,istripe,my_thread)), w, &
-!                                                  nbw, nl, stripe_width, nbw)
-!#else
-!               call quad_hh_trafo_real_avx512_4hv_single(c_loc(a(1,j+off+a_off-3,istripe)), w, &
-!                                                  nbw, nl, stripe_width, nbw)
-!#endif
-!             enddo
-!             do jj = j, 2, -2
-!               w(:,1) = bcast_buffer(1:nbw,jj+off)
-!               w(:,2) = bcast_buffer(1:nbw,jj+off-1)
-!#ifdef WITH_OPENMP
-!               call double_hh_trafo_real_avx512_2hv_single(c_loc(a(1,jj+off+a_off-1,istripe,my_thread)), &
-!                                                    w, nbw, nl, stripe_width, nbw)
-!#else
-!               call double_hh_trafo_real_avx512_2hv_single(c_loc(a(1,jj+off+a_off-1,istripe)), &
-!                                                    w, nbw, nl, stripe_width, nbw)
-!#endif
-!             enddo
-!#ifdef WITH_OPENMP
-!             if (jj==1) call single_hh_trafo_real_cpu_openmp_single(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1, &
-!                                                                      istripe,my_thread), &
-!                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
-!#else
-!             if (jj==1) call single_hh_trafo_real_cpu_single(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
-!                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
-!#endif
-!
-!#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX_BLOCK6_KERNEL) ) */
-!
-!#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
-!           endif
-!#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
-!#endif /* WITH_REAL_AVX512_BLOCK4_KERNEL */
+#if defined(WITH_REAL_AVX512_BLOCK4_KERNEL)
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
+           if (THIS_REAL_ELPA_KERNEL .eq. REAL_ELPA_KERNEL_AVX512_BLOCK4) then
+#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX512_BLOCK6_KERNEL))
+             ! X86 INTRINSIC CODE, USING 4 HOUSEHOLDER VECTORS
+             do j = ncols, 4, -4
+               w(:,1) = bcast_buffer(1:nbw,j+off)
+               w(:,2) = bcast_buffer(1:nbw,j+off-1)
+               w(:,3) = bcast_buffer(1:nbw,j+off-2)
+               w(:,4) = bcast_buffer(1:nbw,j+off-3)
+#ifdef WITH_OPENMP
+               call quad_hh_trafo_real_avx512_4hv_single(c_loc(a(1,j+off+a_off-3,istripe,my_thread)), w, &
+                                                  nbw, nl, stripe_width, nbw)
+#else
+               call quad_hh_trafo_real_avx512_4hv_single(c_loc(a(1,j+off+a_off-3,istripe)), w, &
+                                                  nbw, nl, stripe_width, nbw)
+#endif
+             enddo
+             do jj = j, 2, -2
+               w(:,1) = bcast_buffer(1:nbw,jj+off)
+               w(:,2) = bcast_buffer(1:nbw,jj+off-1)
+#ifdef WITH_OPENMP
+               call double_hh_trafo_real_avx512_2hv_single(c_loc(a(1,jj+off+a_off-1,istripe,my_thread)), &
+                                                    w, nbw, nl, stripe_width, nbw)
+#else
+               call double_hh_trafo_real_avx512_2hv_single(c_loc(a(1,jj+off+a_off-1,istripe)), &
+                                                    w, nbw, nl, stripe_width, nbw)
+#endif
+             enddo
+#ifdef WITH_OPENMP
+             if (jj==1) call single_hh_trafo_real_cpu_openmp_single(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1, &
+                                                                      istripe,my_thread), &
+                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
+#else
+             if (jj==1) call single_hh_trafo_real_cpu_single(a(1:stripe_width,1+off+a_off:1+off+a_off+nbw-1,istripe), &
+                                          bcast_buffer(1:nbw,off+1), nbw, nl, stripe_width)
+#endif
+
+#endif /* defined(WITH_NO_SPECIFIC_REAL_KERNEL) || (defined(WITH_ONE_SPECIFIC_REAL_KERNEL) && !defined(WITH_REAL_AVX_BLOCK6_KERNEL) ) */
+
+#if defined(WITH_NO_SPECIFIC_REAL_KERNEL)
+           endif
+#endif /* WITH_NO_SPECIFIC_REAL_KERNEL */
+#endif /* WITH_REAL_AVX512_BLOCK4_KERNEL */
 
 #if defined(WITH_REAL_SSE_BLOCK6_KERNEL)
 #if defined(WITH_NO_SPECIFIC_REAL_KERNEL)