Reduce duplication within debug code

CHANGELOG.md

@@ -6,6 +6,8 @@ Full documentation for rocSOLVER is available at [rocsolver.readthedocs.io](http
 ### Added
 ### Optimized
 ### Changed
+- Changed `ROCSOLVER_EMBED_FMT` default to `ON` for users building directly with CMake.
+  This matches the existing default when building with install.sh or rmake.py.
 ### Deprecated
 ### Removed
 ### Fixed

CMakeLists.txt

@@ -31,7 +31,7 @@ set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_EXTENSIONS OFF)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
-option(ROCSOLVER_EMBED_FMT "Hide libfmt symbols" OFF)
+option(ROCSOLVER_EMBED_FMT "Hide libfmt symbols" ON)
 option(OPTIMAL "Build specialized kernels for small matrix sizes" ON)
 option(ROCSOLVER_FIND_PACKAGE_LAPACK_CONFIG "Skip module mode search for LAPACK" ON)
 

clients/CMakeLists.txt

@@ -104,7 +104,7 @@ if(BUILD_CLIENTS_BENCHMARKS OR BUILD_CLIENTS_TESTS)
   )
   set(common_source_files
     common/lapack_host_reference.cpp
-    rocblascommon/utility.cpp
+    rocblascommon/clients_utility.cpp
     ${explicit_inst_files}
   )
 

clients/gtest/rocsolver_gtest_main.cpp

@@ -1,11 +1,12 @@
 /* ************************************************************************
- * Copyright (c) 2020-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2020-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #include <cstdio>
 #include <string>
 
 #include <fmt/core.h>
+#include <fmt/ostream.h>
 #include <gtest/gtest.h>
 #include <rocsolver/rocsolver.h>
 

clients/include/clientcommon.hpp

@@ -8,9 +8,9 @@
 #include "common_host_helpers.hpp"
 #include "rocsolver_datatype2string.hpp"
 // rocblas common
+#include "rocblascommon/clients_utility.hpp"
 #include "rocblascommon/rocblas_test.hpp"
 #include "rocblascommon/rocblas_vector.hpp"
-#include "rocblascommon/utility.hpp"
 //#include "rocblascommon/device_vector.hpp"
 //#include "rocblascommon/device_batch_vector.hpp"
 //#include "rocblascommon/device_strided_batch_vector.hpp"

clients/include/rocsolver_arguments.hpp

@@ -7,6 +7,7 @@
 #include <set>
 
 #include <fmt/format.h>
+#include <fmt/ostream.h>
 #include <fmt/ranges.h>
 #include <rocblas/rocblas.h>
 

clients/rocblascommon/utility.cpp → clients/rocblascommon/clients_utility.cpp

@@ -1,15 +1,16 @@
 /* ************************************************************************
- * Copyright (c) 2018-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2018-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #include <cstdlib>
 #include <cstring>
 #include <string>
 
 #include <fmt/core.h>
+#include <rocblas/rocblas.h>
 
+#include "clients_utility.hpp"
 #include "rocblas_random.hpp"
-#include "utility.hpp"
 
 // Random number generator
 // Note: We do not use random_device to initialize the RNG, because we want

clients/rocblascommon/utility.hpp → clients/rocblascommon/clients_utility.hpp

@@ -1,5 +1,5 @@
 /* ************************************************************************
- * Copyright (c) 2018-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2018-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #pragma once

clients/rocblascommon/d_vector.hpp

@@ -1,5 +1,5 @@
 /* ************************************************************************
- * Copyright (c) 2018-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2018-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #pragma once
@@ -8,6 +8,7 @@
 #include <cstdio>
 
 #include <fmt/core.h>
+#include <fmt/ostream.h>
 #include <rocblas/rocblas.h>
 
 #include "rocblas_init.hpp"

clients/rocblascommon/rocblas_init.hpp

@@ -1,5 +1,5 @@
 /* ************************************************************************
- * Copyright (c) 2018-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2018-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #pragma once
@@ -9,6 +9,7 @@
 #include <vector>
 
 #include <fmt/core.h>
+#include <fmt/ostream.h>
 #include <rocblas/rocblas.h>
 
 #include "rocblas_math.hpp"

clients/rocblascommon/rocblas_test.hpp

@@ -1,5 +1,5 @@
 /* ************************************************************************
- * Copyright (c) 2018-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2018-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #pragma once
@@ -9,14 +9,14 @@
 #include <cstring>
 #include <functional>
 #include <iostream>
-#include <rocblas/rocblas.h>
 #include <sstream>
 #include <string>
 #include <type_traits>
 #include <unordered_map>
 #include <utility>
 
 #include <fmt/core.h>
+#include <fmt/ostream.h>
 #include <rocblas/rocblas.h>
 
 // Suppress warnings about hipMalloc(), hipFree() except in rocblas-test and

common/include/common_host_helpers.hpp

@@ -14,7 +14,8 @@
 #include <fmt/ostream.h>
 #include <hip/hip_runtime_api.h>
 
-#include "rocblascommon/utility.hpp"
+#include "fmt_rocblas_types.hpp"
+#include "rocblas_utility.hpp"
 
 /*
  * ===========================================================================
@@ -81,8 +82,8 @@ void pairs_to_string(std::string& str, const char* sep, T1 arg1, T2 arg2, Ts...
 /** Set of helpers to print out data hosted in the CPU and/or the GPU **/
 /***********************************************************************/
 
-/*! \brief Print provided data into specified stream (real case)*/
-template <typename T, std::enable_if_t<!rocblas_is_complex<T>, int> = 0>
+/*! \brief Print provided data into specified stream */
+template <typename T>
 void print_to_stream(std::ostream& os,
                      const std::string name,
                      const rocblas_int m,
@@ -148,73 +149,6 @@ void print_to_stream(std::ostream& os,
     os.flush();
 }
 
-/*! \brief Print provided data into specified stream (complex cases)*/
-template <typename T, std::enable_if_t<rocblas_is_complex<T>, int> = 0>
-void print_to_stream(std::ostream& os,
-                     const std::string name,
-                     const rocblas_int m,
-                     const rocblas_int n,
-                     T* A,
-                     const rocblas_int lda,
-                     const rocblas_fill uplo)
-{
-    std::string s;
-    bool empty = name.empty();
-    if(!empty)
-        s += fmt::format("{}-by-{} matrix: {}\n", m, n, name);
-
-    if(uplo == rocblas_fill_full)
-    {
-        // normal case
-        for(int i = 0; i < m; i++)
-        {
-            if(!empty)
-                s += "    ";
-            for(int j = 0; j < n; j++)
-            {
-                s += fmt::format("{}+{}*i", A[j * lda + i].real(), A[j * lda + i].imag());
-                if(j < n - 1)
-                    s += ", ";
-            }
-            s += '\n';
-        }
-    }
-    else
-    {
-        // symmetric case
-        for(int i = 0; i < min(m, n); i++)
-        {
-            if(!empty)
-                s += "    ";
-            for(int j = 0; j < min(m, n); j++)
-            {
-                if(uplo == rocblas_fill_upper)
-                {
-                    if(i < j)
-                        s += fmt::format("{}+{}*i", A[j * lda + i].real(), A[j * lda + i].imag());
-                    else
-                        s += fmt::format("{}+{}*i", A[i * lda + j].real(), A[i * lda + j].imag());
-                }
-                else
-                {
-                    if(i > j)
-                        s += fmt::format("{}+{}*i", A[j * lda + i].real(), A[j * lda + i].imag());
-                    else
-                        s += fmt::format("{}+{}*i", A[i * lda + j].real(), A[i * lda + j].imag());
-                }
-
-                if(j < n - 1)
-                    s += ", ";
-            }
-            s += '\n';
-        }
-    }
-
-    s += '\n';
-    os << s;
-    os.flush();
-}
-
 /*! \brief Print data from a normal or strided_batched array on the GPU to screen*/
 template <typename T>
 void print_device_matrix(std::ostream& os,
@@ -380,37 +314,7 @@ void print_host_matrix(std::ostream& os,
     os.flush();
 }
 
-template <typename T, std::enable_if_t<!rocblas_is_complex<T>, int> = 0>
-void print_host_matrix(std::ostream& os,
-                       const std::string name,
-                       const rocblas_int m,
-                       const rocblas_int n,
-                       T* CPU_result,
-                       T* GPU_result,
-                       const rocblas_int lda,
-                       double error_tolerance)
-{
-    std::string s;
-    bool empty = name.empty();
-    if(!empty)
-        s += fmt::format("{}-by-{} matrix: {}\n", m, n, name);
-
-    for(size_t j = 0; j < n; j++)
-    {
-        for(size_t i = 0; i < m; i++)
-        {
-            T comp = (CPU_result[j * lda + i] - GPU_result[j * lda + i]) / CPU_result[j * lda + i];
-            if(abs(comp) > error_tolerance)
-                s += fmt::format("matrix  row {}, col {}, CPU result={}, GPU result={}\n", i, j,
-                                 CPU_result[j * lda + i], GPU_result[j * lda + i]);
-        }
-    }
-    s += '\n';
-    os << s;
-    os.flush();
-}
-
-template <typename T, std::enable_if_t<rocblas_is_complex<T>, int> = 0>
+template <typename T>
 void print_host_matrix(std::ostream& os,
                        const std::string name,
                        const rocblas_int m,
@@ -430,7 +334,7 @@ void print_host_matrix(std::ostream& os,
         for(size_t i = 0; i < m; i++)
         {
             T comp = (CPU_result[j * lda + i] - GPU_result[j * lda + i]) / CPU_result[j * lda + i];
-            if(sqrt(comp.real() * comp.real() + comp.imag() * comp.imag()) > error_tolerance)
+            if(std::abs(comp) > error_tolerance)
                 s += fmt::format("matrix  row {}, col {}, CPU result={}, GPU result={}\n", i, j,
                                  CPU_result[j * lda + i], GPU_result[j * lda + i]);
         }

common/include/fmt_rocblas_types.hpp

@@ -0,0 +1,34 @@
+/* ************************************************************************
+ * Copyright (c) 2021-2022 Advanced Micro Devices, Inc.
+ * ************************************************************************ */
+
+#pragma once
+
+#include <fmt/format.h>
+#include <fmt/ostream.h>
+#include <rocblas/rocblas.h>
+
+/* The format function for user-defined types cannot be const before fmt v8.0
+   but must be const in fmt v8.1 if the type is used in a tuple. */
+#if FMT_VERSION < 80000
+#define ROCSOLVER_FMT_CONST
+#else
+#define ROCSOLVER_FMT_CONST const
+#endif
+
+namespace fmt
+{
+template <typename T>
+struct formatter<rocblas_complex_num<T>> : formatter<T>
+{
+    template <typename FormatCtx>
+    auto format(const rocblas_complex_num<T>& value, FormatCtx& ctx) ROCSOLVER_FMT_CONST
+    {
+        formatter<T>::format(value.real(), ctx);
+        format_to(ctx.out(), "+");
+        formatter<T>::format(value.imag(), ctx);
+        format_to(ctx.out(), "*i");
+        return ctx.out();
+    }
+};
+}

library/src/rocblascommon/utility.hpp → common/include/rocblas_utility.hpp

@@ -4,16 +4,16 @@
 
 #pragma once
 
-#include "definitions.hpp"
-#include "rocblas/rocblas.h"
-#include "rocsolver/rocsolver.h"
 #include <cmath>
 #include <complex>
 #include <exception>
-#include <hip/hip_runtime.h>
 #include <new>
 #include <type_traits>
 
+#include <hip/hip_runtime.h>
+#include <rocblas/rocblas.h>
+#include <rocsolver/rocsolver.h>
+
 #pragma STDC CX_LIMITED_RANGE ON
 
 // half vectors

install.sh

@@ -540,7 +540,6 @@ cmake_common_options+=(
   '-DCPACK_SET_DESTDIR=OFF'
   "-DCMAKE_INSTALL_PREFIX=${lib_dir}"
   "-DCPACK_PACKAGING_INSTALL_PREFIX=${install_dir}"
-  '-DROCSOLVER_EMBED_FMT=ON'
   "-DCMAKE_BUILD_TYPE=${build_type}"
 )
 

library/src/common/rocsolver_logger.cpp

@@ -8,7 +8,7 @@
 #include <iostream>
 #include <string>
 
-#include "rocblascommon/utility.hpp"
+#include "rocblas_utility.hpp"
 #include "rocsolver_logger.hpp"
 
 #define STRINGIFY(s) STRINGIFY_HELPER(s)

library/src/include/libcommon.hpp

@@ -1,9 +1,10 @@
 /* ************************************************************************
- * Copyright (c) 2019-2021 Advanced Micro Devices, Inc.
+ * Copyright (c) 2019-2022 Advanced Micro Devices, Inc.
  * ************************************************************************ */
 
 #pragma once
 
-#include "rocblascommon/utility.hpp"
 #include <string>
 #include <vector>
+
+#include "rocblas_utility.hpp"

library/src/include/rocsolver_logvalue.hpp

@@ -5,17 +5,11 @@
 #pragma once
 
 #include <fmt/format.h>
+#include <fmt/ostream.h>
 
+#include "fmt_rocblas_types.hpp"
 #include "rocsolver_datatype2string.hpp"
 
-/* The format function for user-defined types cannot be const before fmt v8.0
-   but must be const in fmt v8.1 if the type is used in a tuple. */
-#if FMT_VERSION < 80000
-#define ROCSOLVER_FMT_CONST
-#else
-#define ROCSOLVER_FMT_CONST const
-#endif
-
 /***************************************************************************
  * Wrapper for types passed to logger, so we can more easily adjust the
  * default way of printing built-in types without doing it globally. (e.g.