refactor(ttm/ttmpy): rename ttm tags.

bassoy · Oct 27, 2024 · 90f3675 · 90f3675
1 parent 2a144f5
commit 90f3675
Show file tree

Hide file tree

Showing 9 changed files with 36 additions and 82 deletions.
diff --git a/example/interface1.cpp b/example/interface1.cpp
@@ -1,25 +1,3 @@
-/*
-# include MLK for fast execution
-
-MKL_ROOT_DIR="/opt/intel/oneapi"
-MKL_BLAS_DIR="${MKL_ROOT_DIR}/mkl/latest"
-MKL_COMP_DIR="${MKL_ROOT_DIR}/compiler/2023.2.0/linux/compiler"
-MKL_BLAS_INC="-I${MKL_BLAS_DIR}/include"
-MKL_BLAS_LIB="-Wl,--start-group ${MKL_BLAS_DIR}/lib/libmkl_intel_ilp64.a ${MKL_BLAS_DIR}/lib/libmkl_intel_thread.a ${MKL_BLAS_DIR}/lib/libmkl_core.a ${MKL_COMP_DIR}/lib/intel64_lin/libiomp5.a -Wl,--end-group -lpthread -lm -ldl -m64"
-MKL_BLAS_FLAGS="-DMKL_ILP64 -m64"
-
-
-TLIB_DIR=..
-TLIB_INC="-I${TLIB_DIR}/include"
-
-INCS="${TLIB_INC} ${MKL_BLAS_INC}"
-LIBS="${MKL_BLAS_LIB}"
-FLAGS="${MKL_BLAS_FLAGS} -DUSE_MKLBLAS"
-
-
-g++ ${INCS} -std=c++17 -Ofast -fopenmp interface1.cpp ${LIBS} ${FLAGS} -o interface1 && ./interface1
-*/
-
 #include <tlib/ttm.h>
 
 #include <vector>

diff --git a/example/interface2.cpp b/example/interface2.cpp
@@ -1,15 +1,3 @@
-/*
-TLIB_DIR=..
-TLIB_INC="-I${TLIB_DIR}/include"
-
-INCS="${TLIB_INC} ${MKL_BLAS_INC}"
-LIBS="${MKL_BLAS_LIB}"
-FLAGS="${MKL_BLAS_FLAGS} -DUSE_MKLBLAS"
-
-
-g++ ${INCS} -std=c++17 -Ofast -fopenmp interface1.cpp ${LIBS} ${FLAGS} -o interface1 && ./interface1
-*/
-
 #include <tlib/ttm.h>
 
 #include <vector>
@@ -67,10 +55,10 @@ int main()
 
 
   // correct shape, layout and strides of the output tensors C1,C2 are automatically computed and returned by the functions.  
-    auto C1 = tlib::ttm(q, A,B, tlib::parallel_policy::threaded_gemm_t{} , tlib::slicing_policy::slice_t{},     tlib::fusion_policy::none_t{} );
-    auto C2 = tlib::ttm(q, A,B, tlib::parallel_policy::omp_forloop_t{}   , tlib::slicing_policy::slice_t{},     tlib::fusion_policy::all_t{} );
-    auto C3 = tlib::ttm(q, A,B, tlib::parallel_policy::omp_forloop_t{}   , tlib::slicing_policy::subtensor_t{}, tlib::fusion_policy::outer_t{} );
-    auto C4 = tlib::ttm(q, A,B, tlib::parallel_policy::batched_gemm_t{}  , tlib::slicing_policy::subtensor_t{}, tlib::fusion_policy::outer_t{} );
+    auto C1 = tlib::ttm(q, A,B, tlib::parallel_policy::threaded_gemm , tlib::slicing_policy::slice,     tlib::fusion_policy::none );
+    auto C2 = tlib::ttm(q, A,B, tlib::parallel_policy::omp_forloop   , tlib::slicing_policy::slice,     tlib::fusion_policy::all );
+    auto C3 = tlib::ttm(q, A,B, tlib::parallel_policy::omp_forloop   , tlib::slicing_policy::subtensor, tlib::fusion_policy::all );
+    auto C4 = tlib::ttm(q, A,B, tlib::parallel_policy::batched_gemm  , tlib::slicing_policy::subtensor, tlib::fusion_policy::all );
 
 
     std::cout << "C1 = " << C1 << std::endl;

diff --git a/example/interface3.cpp b/example/interface3.cpp
@@ -1,8 +1,3 @@
-/*
-# include either -DUSE_OPENBLAS or -DUSE_INTELBLAS for fast execution
-g++ -I../include/ -std=c++17 -Ofast -fopenmp interface3.cpp -o interface3 && ./interface3
-*/
-
 #include <tlib/ttm.h>
 
 #include <vector>
@@ -49,14 +44,14 @@ int main()
     std::cout << "B = [ "; std::copy(B.begin(), B.end(), iterator_t(std::cout, " ")); std::cout << " ];" << std::endl;
 
     tlib::ttm(
-        tlib::parallel_policy::threaded_gemm_t{} , tlib::slicing_policy::slice_t{},  tlib::fusion_policy::none_t{},
+        tlib::parallel_policy::threaded_gemm , tlib::slicing_policy::slice,  tlib::fusion_policy::none,
         q, p,
         A.data(), na.data(), wa.data(), pia.data(),
         B.data(), nb.data(),            pib.data(),
         C1.data(), nc.data(), wc.data());
 
     tlib::ttm(
-        tlib::parallel_policy::omp_forloop_t{} , tlib::slicing_policy::subtensor_t{}, tlib::fusion_policy::outer_t{},
+        tlib::parallel_policy::omp_forloop, tlib::slicing_policy::subtensor, tlib::fusion_policy::all,
         q, p,
         A.data(), na.data(), wa.data(), pia.data(),
         B.data(), nb.data(),            pib.data(),

diff --git a/example/measure.cpp b/example/measure.cpp
@@ -178,7 +178,7 @@ int main(int argc, char* argv[])
 
     if(method == 2 || method == 7){
       std::cout << "Algorithm: <par-loop | slice-qd, all>" << std::endl;
-      measure(q, A, B, C, tlib::parallel_policy::omp_forloop,   tlib::slicing_policy::subtensor, tlib::fusion_policy::outer);
+      measure(q, A, B, C, tlib::parallel_policy::omp_forloop,   tlib::slicing_policy::subtensor, tlib::fusion_policy::all );
       std::cout << "---------" << std::endl << std::endl;
     }
 
@@ -203,7 +203,7 @@ int main(int argc, char* argv[])
 
     if(method == 6 || method == 7){
       std::cout << "Algorithm: <par-gemm | slice-qd, all>" << std::endl;
-      measure(q, A, B, C, tlib::parallel_policy::threaded_gemm, tlib::slicing_policy::subtensor, tlib::fusion_policy::outer );
+      measure(q, A, B, C, tlib::parallel_policy::threaded_gemm, tlib::slicing_policy::subtensor, tlib::fusion_policy::all );
       std::cout << "---------" << std::endl << std::endl;  
     } 
 

diff --git a/include/tlib/detail/tags.h b/include/tlib/detail/tags.h
@@ -27,14 +27,14 @@ struct omp_taskloop_t   {}; // omp_taskloops with single threaded gemm
 struct omp_forloop_t    {}; // omp_for with single threaded gemm
 struct omp_forloop_and_threaded_gemm_t  {}; // omp_for with multi-threaded gemm
 struct batched_gemm_t   {}; // multithreaded batched gemm with collapsed loops
-struct depends_t        {};
+struct combined_t       {};
 
 inline constexpr sequential_t     sequential;
 inline constexpr threaded_gemm_t  threaded_gemm;
 inline constexpr omp_taskloop_t   omp_taskloop;
 inline constexpr omp_forloop_t    omp_forloop;
 inline constexpr batched_gemm_t   batched_gemm;
-inline constexpr depends_t        depends;
+inline constexpr combined_t       combined;
 
 }
 
@@ -43,27 +43,25 @@ namespace tlib::slicing_policy
 {
 struct slice_t     {};
 struct subtensor_t {};
-struct depends_t      {};
+struct combined_t  {};
 
 
-inline constexpr depends_t   depends;
-inline constexpr slice_t     slice;
-inline constexpr subtensor_t subtensor;
+inline constexpr combined_t   combined;
+inline constexpr slice_t      slice;
+inline constexpr subtensor_t  subtensor;
 
 }
 
 
 namespace tlib::fusion_policy
 {
-struct depends_t {};
-struct none_t    {};
-struct outer_t   {};
-struct all_t     {};
+struct none_t     {};
+struct outer_t    {};
+struct all_t      {};
 
-inline constexpr depends_t depends;
-inline constexpr none_t    none;
-inline constexpr outer_t   outer;
-inline constexpr all_t     all;
+inline constexpr none_t     none;
+inline constexpr outer_t    outer;
+inline constexpr all_t      all;
 }
 
 
diff --git a/include/tlib/detail/ttm.h b/include/tlib/detail/ttm.h
@@ -787,7 +787,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::outer_t,
+        parallel_policy::threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::all_t,
         unsigned const q, unsigned const p,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -842,7 +842,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::omp_forloop_t, slicing_policy::subtensor_t, fusion_policy::outer_t,
+        parallel_policy::omp_forloop_t, slicing_policy::subtensor_t, fusion_policy::all_t,
         unsigned const q, unsigned const p,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -906,7 +906,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::omp_forloop_and_threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::outer_t,
+        parallel_policy::omp_forloop_and_threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::all_t,
         unsigned const q, unsigned const p,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -965,7 +965,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::omp_forloop_and_threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::outer_t,
+        parallel_policy::omp_forloop_and_threaded_gemm_t, slicing_policy::subtensor_t, fusion_policy::all_t,
         unsigned const q, unsigned const p, double ratio,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -1028,7 +1028,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::batched_gemm_t, slicing_policy::subtensor_t, fusion_policy::outer_t,
+        parallel_policy::batched_gemm_t, slicing_policy::subtensor_t, fusion_policy::all_t,
         unsigned const q, unsigned const p,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -1115,7 +1115,7 @@ inline void ttm(
 
 template<class value_t, class size_t>
 inline void ttm(
-        parallel_policy::depends_t, slicing_policy::depends_t, fusion_policy::depends_t,
+        parallel_policy::combined_t, slicing_policy::combined_t, fusion_policy::all_t,
         unsigned const q, unsigned const p,
         const value_t *a, size_t const*const na, size_t const*const wa, size_t const*const pia,
         const value_t *b, size_t const*const nb, size_t const*const pib,
@@ -1157,7 +1157,7 @@ inline void ttm(
     auto const outer = product(na, pia, qh+1,p+1);
 
     if( outer >= cores){
-        ttm(parallel_policy::omp_forloop, slicing_policy::subtensor, fusion_policy::outer,
+        ttm(parallel_policy::omp_forloop, slicing_policy::subtensor, fusion_policy::all,
             q, p,
             a, na, wa, pia,
             b, nb,     pib,

diff --git a/include/tlib/ttm.h b/include/tlib/ttm.h
@@ -134,5 +134,5 @@ template<class value_t>
 inline auto operator*(tlib::tensor_view<value_t> const& a,  tlib::tensor<value_t> const& b)
 {
     return ttm(a.contraction_mode(), a.get_tensor(),  b,
-               tlib::parallel_policy::depends, tlib::slicing_policy::depends, tlib::fusion_policy::depends) ;
+               tlib::parallel_policy::combined, tlib::slicing_policy::combined, tlib::fusion_policy::all) ;
 }
diff --git a/test/src/gtest_tlib_ttm.cpp b/test/src/gtest_tlib_ttm.cpp
@@ -366,7 +366,7 @@ TEST(TensorTimesMatrix, OmpForLoopSubtensorOuterFusion)
     using size_type        = std::size_t;
     using execution_policy = tlib::parallel_policy::omp_forloop_t;
     using slicing_policy   = tlib::slicing_policy::subtensor_t;
-    using fusion_policy    = tlib::fusion_policy::outer_t;
+    using fusion_policy    = tlib::fusion_policy::all_t;
 
     check_tensor_times_matrix<value_type,size_type,execution_policy,slicing_policy,fusion_policy,2u>(2u,3);
     check_tensor_times_matrix<value_type,size_type,execution_policy,slicing_policy,fusion_policy,3u>(2u,3);
@@ -380,7 +380,7 @@ TEST(TensorTimesMatrix, OmpForLoopThreadedGemmSubtensorOuterFusion)
     using size_type        = std::size_t;
     using execution_policy = tlib::parallel_policy::omp_forloop_and_threaded_gemm_t;
     using slicing_policy   = tlib::slicing_policy::subtensor_t;
-    using fusion_policy    = tlib::fusion_policy::outer_t;
+    using fusion_policy    = tlib::fusion_policy::all_t;
 
     check_tensor_times_matrix<value_type,size_type,execution_policy,slicing_policy,fusion_policy,2u>(2u,3);
     check_tensor_times_matrix<value_type,size_type,execution_policy,slicing_policy,fusion_policy,3u>(2u,3);
@@ -397,7 +397,7 @@ TEST(TensorTimesMatrix, BatchedGemmSubtensorOuterFusion)
     using size_type        = std::size_t;
     using execution_policy = tlib::parallel_policy::batched_gemm_t;
     using slicing_policy   = tlib::slicing_policy::subtensor_t;
-    using fusion_policy    = tlib::fusion_policy::outer_t;
+    using fusion_policy    = tlib::fusion_policy::all_t;
 
 
     check_tensor_times_matrix<value_type,size_type,execution_policy,slicing_policy,fusion_policy,2u>(2u,3);

diff --git a/ttmpy/src/wrapped_ttm.cpp b/ttmpy/src/wrapped_ttm.cpp
@@ -67,16 +67,11 @@ ttm(std::size_t const contraction_mode,
   auto* cptr        = static_cast<T*>(cinfo.ptr);    // extract data an shape of input array  
   // auto nnc          = std::size_t(cinfo.size);
 
-
-#if defined(USE_OPENBLAS) || defined(USE_MKL)
-  tlib::ttm<T>(tlib::parallel_policy::omp_forloop_t{}, tlib::slicing_policy::subtensor_t{}, tlib::fusion_policy::outer_t{}, 
-                               q, p, 
-                               aptr, na.data(), wa.data(), pia.data(),  
-                               bptr, nb.data(),            pib.data(), 
-                               cptr, nc.data(), wc.data());
-#else 
-
-#endif
+  tlib::ttm<T>(tlib::parallel_policy::combined, tlib::slicing_policy::combined, tlib::fusion_policy::all, 
+                q, p, 
+                aptr, na.data(), wa.data(), pia.data(),  
+                bptr, nb.data(),            pib.data(), 
+                cptr, nc.data(), wc.data());
 
   return c;  
 }