snippets: fp32 gemm.deit.cc

maekawatoshiki · Oct 5, 2024 · 493eaa3 · 493eaa3
1 parent 0a725ee
commit 493eaa3
Showing 1 changed file with 12 additions and 11 deletions.
diff --git a/snippets/sgemm/gemm.deit.cc b/snippets/sgemm/gemm.deit.cc
@@ -11,19 +11,19 @@
 #include <blis.h>
 
 static const size_t N = 4000;
-static const size_t ThreadBlockSizeI = 250;  // 5周（×16スレッド）
-static const size_t ThreadBlockSizeK = 4000; // 50周
-static const size_t ThreadBlockSizeJ = 4000; // 50周
+static const size_t ThreadBlockSizeI = N/16; // 250;  // 5周（×16スレッド）
+static const size_t ThreadBlockSizeK = N; // 50周
+static const size_t ThreadBlockSizeJ = N; // 50周
 static const size_t L3CacheBlockSizeI = 50;  // 1周
 static const size_t L3CacheBlockSizeK = 80;  // 2周
 static const size_t L3CacheBlockSizeJ = 80;  // 2周
 static const size_t L1DCacheBlockSizeI = 50; // 10周
 static const size_t L1DCacheBlockSizeK = 40; // 10周
-static const size_t L1DCacheBlockSizeJ = 40; // SIMD方向8要素×5周
+static const size_t L1DCacheBlockSizeJ = 80; // SIMD方向8要素×5周
 static const size_t RegisterBlockSizeI = 5;  // 5レジスタ並列に
 static const size_t RegisterBlockSizeK = 4;  // fma連鎖4回
 
-void mm( const double *__restrict__ a, const double *__restrict__ b, double *__restrict__ c ) {
+void mm( const float *__restrict__ a, const float *__restrict__ b, float *__restrict__ c ) {
     for( int i1 = 0; i1 < ThreadBlockSizeI; i1 += L3CacheBlockSizeI )
         for( int k1 = 0; k1 < ThreadBlockSizeK; k1 += L3CacheBlockSizeK )
             for( int j1 = 0; j1 < ThreadBlockSizeJ; j1 += L3CacheBlockSizeJ )
@@ -44,9 +44,9 @@ void mm( const double *__restrict__ a, const double *__restrict__ b, double *__r
                                             }
 }
 
-alignas(64) double ah[N*N];
-alignas(64) double bh[N*N];
-alignas(64) double ch[N*N], ch_cblas[N*N];
+alignas(64) float ah[N*N];
+alignas(64) float bh[N*N];
+alignas(64) float ch[N*N], ch_cblas[N*N];
 
 int main() {
     std::mt19937_64 mt;
@@ -73,15 +73,16 @@ int main() {
 
         const auto finish = std::chrono::system_clock::now();
 
-        cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, N, N, N, 1.0, ah, N, bh, N, 1.0, ch_cblas, N);
+        cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, N, N, N, 1.0, ah, N, bh, N, 1.0, ch_cblas, N);
 
         for( int i = 0; i < N*N; ++i )
-            if( std::abs( ch[i] - ch_cblas[i] ) > 1e-6 )
+            if( std::abs( ch[i] - ch_cblas[i] ) > 1e-2 ) {
+                std::cerr << "mismatch at " << i << ": " << ch[i] << " != " << ch_cblas[i] << std::endl;
                 exit(EXIT_FAILURE);
+            }
 
         const double s = std::chrono::duration_cast<std::chrono::nanoseconds>( finish - start ).count() * 1e-9;
         static constexpr double flop_per_fma = 2.0;
-        static constexpr double insn_per_fma = 1.0 / 8.0;
         std::cout << s << " seconds, " << N*N*N*flop_per_fma/s * 1e-9 << " GFLOPS" << std::endl;
     }
 }