--avx2-only flag

kfrlib · Feb 26, 2024 · 08afae3 · 08afae3
1 parent 1b42cb4
commit 08afae3
Show file tree

Hide file tree

Showing 6 changed files with 44 additions and 8 deletions.
diff --git a/README.md b/README.md
@@ -24,6 +24,22 @@ Requires:
   * matplotlib module
   * numpy module
 
+## Options
+
+| Option             | Description                                                     |
+|--------------------|-----------------------------------------------------------------|
+| `SIZE`             | 1D FFT                                                          |
+| `SIZExSIZE`        | 2D FFT. Example: `64x32`                                        |
+| `SIZExSIZExSIZE`   | 3D FFT. Example: `64x32x16`                                     |
+| `--complex flags`  | `y` (complex tests), `yn` (all tests), `n` (real tests)         |
+| `--inverse flags`  | `y` (IDFT tests), `ny` (DFT/IDFT tests), `n` (DFT tests)        |
+| `--inplace flags`  | `y` (inplace tests), `ny` (all tests), `n` (out-of-place tests) |
+| `--save data.json` | Save results in JSON                                            |
+| `--save -`         | Print resulting JSON to stdout                                  |
+| `--avx2-only`      | Enable only AVX2 (supported in KFR, IPP, MKL)                   |
+| `--no-progress`    | Disable verbose progress output                                 |
+| `--no-banner`      | Disable banner                                                  |
+
 ## Benchmark code license
 
 MIT
diff --git a/fftlibs/ipp/ipp.cpp b/fftlibs/ipp/ipp.cpp
@@ -18,10 +18,15 @@ std::string fft_name()
 {
     ippInit();
 #if defined(__x86_64__) || defined(_M_X64)
-    ippSetCpuFeatures(ippCPUID_MMX | ippCPUID_SSE | ippCPUID_SSE2 | ippCPUID_SSE3 | ippCPUID_SSSE3 |
-                      ippCPUID_MOVBE | ippCPUID_SSE41 | ippCPUID_SSE42 | ippCPUID_AES | ippCPUID_CLMUL |
-                      ippCPUID_SHA | ippCPUID_AVX | ippAVX_ENABLEDBYOS | ippCPUID_RDRAND | ippCPUID_F16C |
-                      ippCPUID_AVX2 | ippCPUID_MOVBE | ippCPUID_ADCOX | ippCPUID_RDSEED | ippCPUID_PREFETCHW);
+    if (avx2only)
+    {
+        fprintf(stderr, "IPP: enabling AVX2\n");
+        ippSetCpuFeatures(ippCPUID_MMX | ippCPUID_SSE | ippCPUID_SSE2 | ippCPUID_SSE3 | ippCPUID_SSSE3 |
+                          ippCPUID_MOVBE | ippCPUID_SSE41 | ippCPUID_SSE42 | ippCPUID_AES | ippCPUID_CLMUL |
+                          ippCPUID_SHA | ippCPUID_AVX | ippAVX_ENABLEDBYOS | ippCPUID_RDRAND | ippCPUID_F16C |
+                          ippCPUID_AVX2 | ippCPUID_MOVBE | ippCPUID_ADCOX | ippCPUID_RDSEED |
+                          ippCPUID_PREFETCHW);
+    }
 #endif
 
     const IppLibraryVersion* ver = ippsGetLibVersion();

diff --git a/fftlibs/kfr/kfr.cpp b/fftlibs/kfr/kfr.cpp
@@ -18,7 +18,11 @@ const char* library_version_dft();
 std::string fft_name()
 {
 #if defined(__x86_64__) || defined(_M_X64)
-    kfr::override_cpu(kfr::cpu_t::avx2);
+    if (avx2only)
+    {
+        fprintf(stderr, "KFR: enabling AVX2\n");
+        kfr::override_cpu(kfr::cpu_t::avx2);
+    }
 #endif
     return std::string(kfr::library_version_dft());
 }

diff --git a/fftlibs/mkl/mkl.cpp b/fftlibs/mkl/mkl.cpp
@@ -7,7 +7,10 @@
 std::string fft_name()
 {
 #if defined(__x86_64__) || defined(_M_X64)
-    MKL_Enable_Instructions(MKL_ENABLE_AVX2);
+    if (avx2only) {
+        fprintf(stderr, "MKL: enabling AVX2\n");
+        MKL_Enable_Instructions(MKL_ENABLE_AVX2);
+    }
 #endif
     char buf[1024];
     MKL_Get_Version_String(buf, sizeof(buf));

diff --git a/src/benchmark.hpp b/src/benchmark.hpp
@@ -421,4 +421,6 @@ extern template fft_impl_ptr<float> fft_create<float>(const std::vector<size_t>&
 extern template fft_impl_ptr<double> fft_create<double>(const std::vector<size_t>& size, bool is_complex,
                                                         bool inverse, bool inplace);
 
-std::string fft_name();
+std::string fft_name();
+
+extern bool avx2only;
diff --git a/src/fft_benchmark.cpp b/src/fft_benchmark.cpp
@@ -154,6 +154,7 @@ std::string execfile(std::string command)
 static std::string outname;
 static bool progress = true;
 static bool banner   = true;
+bool avx2only = false;
 static std::vector<std::vector<size_t>> sizes;
 static std::vector<bool> is_complex_list{ true, false };
 static std::vector<bool> inverse_list{ false, true };
@@ -282,6 +283,10 @@ int main(int argc, char** argv)
         {
             banner = true;
         }
+        else if (argv[i] == "--avx2-only"sv)
+        {
+            avx2only = true;
+        }
         else if (argv[i] == "--"sv)
         {
         }
@@ -304,7 +309,8 @@ int main(int argc, char** argv)
     {
         printf("FFT/DFT benchmarking tool. Copyright (C) 2016-2024 Dan Cazarin https://www.kfrlib.com\n");
         printf("Benchmark source code is MIT-licensed\n");
-        printf("DFT/FFT libraries have their own licenses. Please refer to the respective source code for details.\n");
+        printf("DFT/FFT libraries have their own licenses. Please refer to the respective source code for "
+               "details.\n");
 
         if (sizes.empty())
         {