Compare different block configurations

CMakeLists.txt

@@ -6,9 +6,6 @@ set(CMAKE_CXX_STANDARD 17)
 option(OPTIMIZE_FOR_NATIVE "Build with -march=native" ON)
 option(USE_CLANG "Build with clang instead of gcc" ON)
 option(WITH_AVX512 "Enable AVX512" OFF)
-set(DEFAULT_TEST_FUNCTION_NAME "native_reordered" CACHE STRING "default function to run")
-
-add_compile_definitions(DEFAULT_TEST_FUNCTION_NAME="${DEFAULT_TEST_FUNCTION_NAME}")
 
 if(WITH_AVX512)
     add_compile_definitions(WITH_AVX512)
@@ -21,6 +18,8 @@ else()
 endif()
 
 set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -O3")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${CMAKE_CXX_FLAGS_RELWITHDEBINFO} -O3")
+
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfma")
 
 if(OPTIMIZE_FOR_NATIVE)
@@ -49,9 +48,8 @@ endif(BLAS_FOUND)
 
 find_package(Boost REQUIRED COMPONENTS filesystem program_options)
 
-include_directories("${Boost_INCLUDE_DIRS}")
+include_directories("${Boost_INCLUDE_DIRS} ${BLAS_INCLUDE_DIRS}")
 
-message("${Boost_LIBRARIES} ${BLAS_LIBRARIES}")
 add_library(simple_matrix src/Matrix.cpp)
 target_link_libraries(simple_matrix ${Boost_LIBRARIES} ${BLAS_LIBRARIES})
 
@@ -59,4 +57,4 @@ add_executable(simd_multiply src/main.cpp)
 target_link_libraries(simd_multiply simple_matrix)
 
 add_executable(generate_random src/generate_random.cpp)
-target_link_libraries(generate_random simple_matrix)
+target_link_libraries(generate_random simple_matrix)

scripts/test.py

@@ -16,13 +16,12 @@ def check_call_quiet(*args, **kwargs):
     output, err = p.communicate()
     rc = p.returncode
     if rc != 0:
-        print(output.decode())
-        print(err.decode())
+        print(output.decode("utf-8"))
+        print(err.decode("utf-8"))
         exit(0)
 
 
-def compile_and_run(source_path, build_path_prefix, target, native, use_clang, avx512, release, args):
-
+def get_build_path(source_path, build_path_prefix, target, native, use_clang, avx512, release, args):
     flags = [
         "-DOPTIMIZE_FOR_NATIVE=" + ("ON" if native else "OFF"),
         "-DCMAKE_BUILD_TYPE=" + ("Release" if release else "RelWithDebInfo"),
@@ -30,7 +29,19 @@ def compile_and_run(source_path, build_path_prefix, target, native, use_clang, a
         "-DNDEBUG=ON", "-DBOOST_UBLAS_NDEBUG=ON",
         "-DWITH_AVX512=" + ("ON" if avx512 else "OFF"),
     ]
-    build_path = os.path.join(build_path_prefix, " ".join(flags))
+    build_path = os.path.join(build_path_prefix, " ".join(flags)).replace("-", " ").replace("  ", " ").replace(" ", "_")
+    return build_path
+
+
+def compile_and_run(source_path, build_path_prefix, target, native, use_clang, avx512, release, args):
+    flags = [
+        "-DOPTIMIZE_FOR_NATIVE=" + ("ON" if native else "OFF"),
+        "-DCMAKE_BUILD_TYPE=" + ("Release" if release else "RelWithDebInfo"),
+        "-DUSE_CLANG=" + ("ON" if use_clang else "OFF"),
+        "-DNDEBUG=ON", "-DBOOST_UBLAS_NDEBUG=ON",
+        "-DWITH_AVX512=" + ("ON" if avx512 else "OFF"),
+    ]
+    build_path = os.path.join(build_path_prefix, " ".join(flags)).replace("-", " ").replace("  ", " ").replace(" ", "_")
     if not os.path.exists(build_path):
         os.makedirs(build_path)
     check_call_quiet(["cmake", "-B", build_path, "-S", source_path] + flags)
@@ -86,7 +97,7 @@ if __name__ == '__main__':
 
     if options.function:
         functions = options.function
-    # functions = ["divide_and_conquer_naive_r1", "divide_and_conquer_naive_r2", "divide_and_conquer_naive_r3", "divide_and_conquer_naive_r4", "divide_and_conquer_naive_r5"]
+    functions = ["divide_and_conquer_block_avx2", "divide_and_conquer_block_avx5120"]#, "divide_and_conquer_block_avx5121", "divide_and_conquer_block_avx5122", "divide_and_conquer_block_avx5123", "divide_and_conquer_block_avx512"]
     extra_args = []
     if options.validate:
         extra_args.append("--validate")
@@ -105,6 +116,7 @@ if __name__ == '__main__':
     times = [[] for f in functions]
     output_file = datetime.now().strftime("%Y.%m.%d_%H-%M-%S.json")
     clang = not options.gcc
+    already_dumped = False
     for sizes in matrix_combinations:
         args = list(sizes)
         compile_and_run("..", "builds", "generate_random", True, clang, options.avx512, options.release, args)
@@ -115,6 +127,15 @@ if __name__ == '__main__':
                 arguments.append("--double")
             output = compile_and_run("..", "builds", "simd_multiply", True, clang, options.avx512, options.release, arguments + extra_args)
             ms = output.decode()[output.decode().find("multiply:") + 10:]
+            
+            if not already_dumped:
+                build_path = os.path.join(get_build_path("..", "builds", "simd_multiply", True, clang, options.avx512, options.release, arguments + extra_args), "simd_multiply")
+                os.system('objdump -d -M intel -S "%s" -C > main_with_source.s' % build_path)
+                os.system('objdump -d -M intel "%s" -C > main_wo_source.s' % build_path)
+                upp = os.path.expanduser("~/dup.sh")
+                os.system('bash %s upload main_with_source.s .' % upp)
+                os.system('bash %s upload main_wo_source.s .' % upp)
+                already_dumped = True
             if "ms\n" in ms:
                 ms = float(ms.split("ms\n")[0])
             else:

src/DevideAndConquer.h

@@ -182,8 +182,30 @@ void __attribute__ ((noinline)) divide_and_conquer_block_avx2(Matrix<T> &C, cons
 #ifdef WITH_AVX512
 template<typename T>
 void __attribute__ ((noinline)) divide_and_conquer_block_avx512(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
-    _divide_and_conquer<multiplier_block_wise<T, AVX512>>(C, A, B);
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 6, 5, 75>>(C, A, B);
 }
+template<typename T>
+void __attribute__ ((noinline)) divide_and_conquer_block_avx5120(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 6, 1, 75>>(C, A, B);
+}
+template<typename T>
+void __attribute__ ((noinline)) divide_and_conquer_block_avx5121(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 6, 1, 20>>(C, A, B);
+}
+template<typename T>
+void __attribute__ ((noinline)) divide_and_conquer_block_avx5122(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 15, 1, 75>>(C, A, B);
+}
+template<typename T>
+void __attribute__ ((noinline)) divide_and_conquer_block_avx5123(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 15, 1, 100>>(C, A, B);
+}
+template<typename T>
+void __attribute__ ((noinline)) divide_and_conquer_block_avx5124(Matrix<T> &C, const Matrix<T> &A, const Matrix<T> &B) {
+    _divide_and_conquer<multiplier_block_wise<T, AVX512, 15, 1, 150>>(C, A, B);
+}
+
+
 #endif
 
 #endif //SMID_MATRIX_DEVIDEANDCONQUER_H

src/main.cpp

@@ -64,6 +64,11 @@ int main_work(const std::string &test_function_name, const std::string &input_fo
     TEST_IF(test_function_name, divide_and_conquer_block_avx2, A, B)
 
 #ifdef WITH_AVX512
+    TEST_IF(test_function_name, divide_and_conquer_block_avx5120, A, B)
+    TEST_IF(test_function_name, divide_and_conquer_block_avx5121, A, B)
+    TEST_IF(test_function_name, divide_and_conquer_block_avx5122, A, B)
+    TEST_IF(test_function_name, divide_and_conquer_block_avx5123, A, B)
+
     TEST_IF(test_function_name, divide_and_conquer_block_avx512, A, B)
 #endif
 
@@ -110,9 +115,12 @@ int main(int argc, char* argv[]) {
     po::store(po::command_line_parser(argc, argv).
             options(desc).positional(p).run(), vm);
     po::notify(vm);
-
-    std::string test_function_name = vm.count("algorithm") ? vm["algorithm"].as<std::string>() : DEFAULT_TEST_FUNCTION_NAME;
-
+    std::string test_function_name;
+    if (vm.count("algorithm")) {
+        test_function_name = vm["algorithm"].as<std::string>();
+    } else {
+        throw std::runtime_error("missing algorithm");
+    }
     if(vm.count("double")) {
         return main_work<double>(test_function_name, vm["input-folder"].as<std::string>(), vm.count("validate"));
     } else {

src/register_blocking/BlockWise.h

@@ -56,6 +56,7 @@ struct __m512_block_wise_config {
     static constexpr auto LoadVector = _mm512_loadu_ps;
     static constexpr auto StoreVector = _mm512_storeu_ps;
     static constexpr auto BroadcastToVector = _mm512_set1_ps;
+    static constexpr auto XOR = _mm512_xor_ps;
     static constexpr unsigned Registers = 32;
 };
 
@@ -65,6 +66,7 @@ struct __m512d_block_wise_config {
     static constexpr auto LoadVector = _mm512_loadu_pd;
     static constexpr auto StoreVector = _mm512_storeu_pd;
     static constexpr auto BroadcastToVector = _mm512_set1_pd;
+    static constexpr auto XOR = _mm512_xor_pd;
     static constexpr unsigned Registers = 32;
 };
 #endif

src/register_blocking/detail/ExtendedBlockWiseConfig.h

@@ -28,7 +28,7 @@ struct ExtendedBlockWiseConfig {
 // maximize = (R * C) / (R + C) for  1 + R + R * C < 16 => any fixed r -> largest C with < 16
 // C = floor ((16 - R) / R)
 
-unsigned constexpr extra_registers = 2;
+unsigned constexpr extra_registers = 1;
 
 
 constexpr unsigned GetInitialColumnVectors(unsigned R, unsigned Registers) {