Merge pull request #48 from opencompl/christos/add-dgemm-8by8

Add 8x8 double precision matmul baseline

kernels/matmul/8x8xf64/Makefile

@@ -0,0 +1,8 @@
+.DEFAULT_GOAL := all
+
+include ../../../snitch/Makefile.rules
+
+TESTS =
+TESTS += baseline.x
+
+include ../../Makefile.kernels

kernels/matmul/8x8xf64/baseline.c

@@ -0,0 +1,16 @@
+#include "data.h"
+
+#include <snrt.h>
+
+#include <stdint.h>
+
+void matmul(const double* x, const double* y, double* g) {
+    for (uint32_t i = 0; i < M; ++i) {
+        for (uint32_t j = 0; j < N; ++j) {
+            for (uint32_t k = 0; k < K; ++k) {
+                // row-major accesses
+                g[i * N + j] += x[i * K + k] * y[k * N + j];
+            }
+        }
+    }
+}

kernels/matmul/8x8xf64/data.c

@@ -0,0 +1,207 @@
+#define M 8
+#define K 8
+#define N 8
+
+const double X[M * K] = {
+97.62700785,
+ 430.37873274,
+ 205.52675214,
+  89.76636599,
+-152.69040132,
+  291.78822613,
+-124.82557747,
+ 783.54600156,
+ 927.325521  ,
+-233.11696235,
+  583.45007617,
+  57.78983951,
+ 136.08912219,
+ 851.19327659,
+-857.9278836 ,
+ -825.7414006 ,
+-959.56320512,
+ 665.2396911 ,
+ 556.3135019 ,
+ 740.02429649,
+  957.23668447,
+ 598.31712843,
+ -77.04127549,
+ 561.05835257,
+-763.45114826,
+  279.84204266,
+-713.29342518,
+ 889.3378341 ,
+  43.6966435 ,
+-170.67612002,
+ -470.88877579,
+ 548.46737887,
+ -87.69933557,
+ 136.86789774,
+-962.42039913,
+  235.27099415,
+ 224.19144544,
+ 233.86799375,
+ 887.49615703,
+ 363.64059821,
+ -280.98419885,
+-125.9360924 ,
+ 395.26239185,
+-879.54905674,
+ 333.53343089,
+  341.27573924,
+-579.23487785,
+-742.14740469,
+-369.14329815,
+-272.57845811,
+  140.39354084,
+-122.79697308,
+ 976.74767612,
+-795.9103785 ,
+-582.24648781,
+ -677.38096423,
+ 306.21665093,
+-493.41679492,
+ -67.37845429,
+-511.148816  ,
+ -682.06083271,
+-779.24971767,
+ 312.65917893,
+-723.6340973
+};
+
+
+const double Y[K * N] = {
+-606.83527664,
+-262.54965868,
+ 641.9864597 ,
+-805.79744841,
+ 675.889815  ,
+ -807.80318421,
+ 952.91893003,
+ -62.6975967 ,
+ 953.52217638,
+ 209.69103949,
+  478.5271588 ,
+-921.62441549,
+-434.38607485,
+-759.60687757,
+-407.71960496,
+ -762.54456209,
+-364.03364121,
+-171.47401097,
+-871.7050073 ,
+ 384.94423874,
+  133.20290841,
+-469.22101812,
+  46.49610693,
+-812.11897848,
+ 151.89299111,
+  858.59239515,
+-362.8620951 ,
+ 334.82075993,
+-736.40427519,
+ 432.65440824,
+ -421.18781411,
+-633.61727599,
+ 173.02586962,
+-959.78490763,
+ 657.88005843,
+ -990.60904761,
+ 355.63307359,
+-459.98405362,
+ 470.38804425,
+ 924.37709023,
+ -502.49371296,
+ 152.31466884,
+ 184.08386254,
+ 144.50381158,
+-553.83673472,
+  905.49802303,
+-105.74924276,
+ 692.81734494,
+ 398.95855064,
+-405.12609829,
+  627.5956394 ,
+-206.98851831,
+ 762.20639422,
+ 162.54574527,
+ 763.47072371,
+  385.06318016,
+ 450.50855964,
+   2.64876385,
+ 912.16726945,
+ 287.98039846,
+ -152.28990288,
+ 212.78642826,
+-961.61360338,
+-396.85036665
+};
+
+
+const double G[M * N] = {
+4.20101194e+05,
+ 3.50083946e+05,
+ 6.46538181e+05,
+ 7.87611951e+04,
+ -5.90066796e+05,
+ 1.75051891e+04,
+-1.06214482e+06,
+-8.56097454e+05,
+ -2.10708718e+06,
+ 1.63457337e+03,
+-1.09121328e+06,
+-3.60471830e+05,
+ -1.87997154e+05,
+-4.27787358e+05,
+ 1.09454317e+06,
+ 3.22031941e+05,
+  1.21350670e+06,
+ 1.36502415e+05,
+ 1.52156393e+05,
+-6.22301317e+04,
+ -1.54349419e+06,
+ 5.37282592e+05,
+-1.68277900e+06,
+-3.20747360e+05,
+  1.27742014e+06,
+ 1.26930043e+06,
+ 1.44955858e+05,
+ 5.67935087e+05,
+ -1.71986709e+06,
+ 9.89135329e+05,
+-2.09766800e+06,
+-6.26582774e+05,
+  1.00898615e+06,
+-1.19381322e+05,
+ 1.84200115e+06,
+-6.14448878e+05,
+  1.51101302e+05,
+ 8.50535943e+05,
+ 1.25405040e+05,
+ 1.10035520e+06,
+ -9.06271345e+05,
+-8.11028383e+05,
+-1.02428859e+06,
+-1.74769700e+05,
+  1.66273819e+05,
+-3.39830959e+05,
+ 5.64655615e+05,
+ 9.66178163e+05,
+ -7.41759606e+04,
+-9.14352682e+05,
+-9.32476120e+05,
+-5.95542872e+05,
+  4.25567604e+05,
+-1.02251944e+06,
+ 5.68088238e+05,
+ 5.90867503e+05,
+ -6.37130208e+05,
+-2.03819589e+05,
+-8.51329047e+05,
+ 3.00857968e+05,
+  1.32626599e+06,
+-5.57125464e+05,
+ 1.40126649e+06,
+-2.71458462e+04
+};
+

kernels/matmul/8x8xf64/data.h

@@ -0,0 +1,9 @@
+#pragma once
+
+#define M 8
+#define K 8
+#define N 8
+
+extern const double X[M * K];
+extern const double Y[K * N];
+extern const double G[M * N];

kernels/matmul/8x8xf64/main.c

@@ -0,0 +1,43 @@
+#include "data.h"
+
+#include <snrt.h>
+
+#include <math.h>
+
+// Kernel provided via external definition
+void matmul(double *x, double *y, double *g);
+
+int main() {
+    // Allocate shared local memory
+    // By avoiding allocators and bumping by a known offset a base pointer
+    // (snrt_l1_next()) that is the same for all the cores in the cluster, we are
+    // essentially providing the same memory regions to all the cores in this cluster.
+    double *local_x = (double *)snrt_l1_next();
+    double *local_y = local_x + K * N;
+    double *local_z = local_y + M * N;
+
+    // Copy data in shared local memory
+    if (snrt_is_dm_core()) {
+        snrt_dma_start_1d(local_x, X, M * N * sizeof(double));
+        snrt_dma_start_1d(local_y, Y, M * N * sizeof(double));
+    }
+
+    snrt_cluster_hw_barrier();
+
+    // Launch kernel: from this point on only core 0 is required to be alive.
+    int thiscore = snrt_cluster_core_idx();
+    if (thiscore != 0) return 0;
+
+    (void)snrt_mcycle();
+    matmul(local_x, local_y, local_z);
+    (void)snrt_mcycle();
+
+    // Correctness check
+    int nerr = 0;
+    for (int i = 0; i < M * N; i++) {
+        double d = fabs(local_z[i] - G[i]);
+        nerr += !(d <= 1E-2f);  // Make sure to take into account NaNs (e.g.: happy path
+                                // on the taken branch)
+    }
+    return nerr;
+}

scripts/run.sh

@@ -84,6 +84,7 @@ KERNEL_DIRS=(
   # "ssum/8x16xf32/"
   # "ssum/14x26xf32/"
   "dsum/8x16xf32/"
+  "matmul/8x8xf64/"
   #"matmul/16x16xf64/"
   "relu/16x16xf64/"
 )