More matrix sizesmultiply-tiling-manymatrices

1 files changed, 45 insertions, 14 deletions

diff --git a/benchmarks/multiply_tiles_benchmark.cc b/benchmarks/multiply_tiles_benchmark.cc
index 33b7efe..7ae033b 100644
--- a/benchmarks/multiply_tiles_benchmark.cc
+++ b/benchmarks/multiply_tiles_benchmark.cc
@@ -8,6 +8,17 @@
 
 using namespace intgemm;
 
+struct matrix_size {
+   const int M;
+   const int K;
+   const int N;
+
+   friend std::ostream& operator<<(std::ostream& os, const matrix_size& m) {
+    os << "Matrix size: M: " << m.M << " K: " << m.K << " N: " << m.N;
+    return os;
+   }
+};
+
 static constexpr Index TestCases[][2] = {
   {1, 1}, {1, 2}, {1, 3}, {1, 4},
   {2, 1}, {2, 2}, {2, 3}, {2, 4},
@@ -33,6 +44,9 @@ struct BenchmarkLoop {
     AlignedVector<int32_t> C(A_rows * B_cols);
 
     std::chrono::duration<double> duration = std::chrono::nanoseconds::zero();
+    if (B_cols % TileColumnsMultiplier*8 != 0) {
+      return;
+    }
     for (Index i = 0; i < repeats; ++i) {
       std::generate(A_prepared.begin(), A_prepared.end(), [&]() {
         return dist(gen);
@@ -47,7 +61,7 @@ struct BenchmarkLoop {
       duration += std::chrono::system_clock::now() - start;
     }
 
-    std::cout << repeats << " iterations of " << Backend::kName << " with tile = " << TileRows << "x" << 8 * TileColumnsMultiplier << " took: " << duration.count() << " seconds." << std::endl;
+    std::cout << repeats << " " << A_rows << "x" << width << "x" << B_cols << " iterations of " << Backend::kName << " with tile = " << TileRows << "x" << 8 * TileColumnsMultiplier << " took: " << duration.count() << " seconds." << std::endl;
   }
 };
 
@@ -86,22 +100,39 @@ int main(int argc, char** argv) {
       std::cerr << "Warning: Unknown architecture '" << argv[i] << "'!" << std::endl;
   }
 
+std::vector<matrix_size> matrices = {
+    {1024, 1024, 1024},
+    {768, 768, 768},
+    {256, 10368, 256},
+    {256, 5312, 256},
+    {8, 2048, 256},
+    {320, 256, 256},
+    {472, 256, 256},
+    {248, 256, 256},
+    {200, 256, 256},
+    {1, 64, 8}};
+
 #define ARCH_BENCHMARK_IMPL(arch, A_rows, width, B_cols) \
   if (selected_archs[i].compare(#arch) == 0) { StaticLoop<BenchmarkLoop<arch>, MakeStaticLoopIterator<TestCasesN>>((A_rows), (width), (B_cols), repeats); }
 
-  for (int i  = 0; i < selected_archs.size(); ++i) {
-    ARCH_BENCHMARK_IMPL(SSSE3_8bit, 768, 768, 768)
-    ARCH_BENCHMARK_IMPL(SSE2_16bit, 768, 768, 768)
-    ARCH_BENCHMARK_IMPL(AVX2_8bit, 768, 768, 768)
-    ARCH_BENCHMARK_IMPL(AVX2_16bit, 768, 768, 768)
-    //  Tiling not supported yet
-    // #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
-      // ARCH_BENCHMARK_IMPL(AVX512BW_8bit, 768, 768, 768)
-      // ARCH_BENCHMARK_IMPL(AVX512BW_8bit, 768, 768, 768)
-    // #endif
-    #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
-      ARCH_BENCHMARK_IMPL(AVX512VNNI_8bit, 768, 768, 768)
-    #endif
+  for (auto&& matrix : matrices) {
+    Index m = matrix.M;
+    Index n = matrix.N;
+    Index k = matrix.K;
+    for (int i  = 0; i < selected_archs.size(); ++i) {
+      ARCH_BENCHMARK_IMPL(SSSE3_8bit, m, n, k)
+      ARCH_BENCHMARK_IMPL(SSE2_16bit, m, n, k)
+      ARCH_BENCHMARK_IMPL(AVX2_8bit, m, n, k)
+      ARCH_BENCHMARK_IMPL(AVX2_16bit, m, n, k)
+      //  Tiling not supported yet
+      // #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512BW
+        // ARCH_BENCHMARK_IMPL(AVX512BW_8bit, 768, 768, 768)
+        // ARCH_BENCHMARK_IMPL(AVX512BW_8bit, 768, 768, 768)
+      // #endif
+      #ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
+        ARCH_BENCHMARK_IMPL(AVX512VNNI_8bit, m, n, k)
+      #endif
+    }
   }
 
   return 0;