Make CI tests faster

Author: hngenc

bareMetalC/tiled_matmul_ws_At.c

@@ -21,6 +21,14 @@ typedef acc_t ACC_T;
 typedef elem_t ACC_T;
 #endif
 
+#ifdef FAST
+
+#define MAT_DIM_I 19
+#define MAT_DIM_K 18
+#define MAT_DIM_J 17
+
+#else
+
 #ifndef BAREMETAL
 #define MAT_DIM_I 500
 #define MAT_DIM_K 412
@@ -31,6 +39,8 @@ typedef elem_t ACC_T;
 #define MAT_DIM_J 30
 #endif
 
+#endif // ifdef FAST
+
 void print_tile(elem_t* in, int tile_dim) {
   for (size_t r = 0; r < tile_dim; r++) {
     printf("row starts at: %p\n", in +r*MAT_DIM_J);
@@ -104,30 +114,51 @@ int main() {
     // printf("Init A\n");
     for (size_t i = 0; i < MAT_DIM_K; ++i) {
       for (size_t j = 0; j < MAT_DIM_I; ++j) {
+#ifdef FAST
+        full_A[i][j] = 1;
+#else
         full_A[i][j] = rand() % 2;
+#endif
       }
     }
 
     // printf("Init B\n");
     for (size_t i = 0; i < MAT_DIM_K; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
+#ifdef FAST
+        full_B[i][j] = 1;
+#else
         full_B[i][j] = rand() % 2;
+#endif
       }
     }
 
     // printf("Init D\n");
     for (size_t i = 0; i < MAT_DIM_I; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
+#ifdef FAST
+        full_D[i][j] = NO_BIAS ? 0 : 1;
+#else
         full_D[i][j] = NO_BIAS ? 0 : rand() % 2;
+#endif
       }
     }
 
+#ifdef FAST
+    for (size_t i = 0; i < MAT_DIM_I; ++i) {
+      for (size_t j = 0; j < MAT_DIM_J; ++j) {
+        gold[i][j] = MAT_DIM_K + !NO_BIAS;
+      }
+    }
+#else
     printf("Starting slow CPU matmul\n");
     unsigned long cpu_start = read_cycles();
     full_matmul(full_A, full_B, full_D, gold_full);
     unsigned long cpu_end = read_cycles();
     printf("Cycles taken: %u\n", cpu_end-cpu_start);
     full_matscale(gold_full, gold, ACC_SCALE_IDENTITY);
+#endif // #ifdef FAST
+
 #endif
 
     printf("Starting gemmini matmul\n");
@@ -150,8 +181,12 @@ int main() {
       printf("C:\n");
       full_printMatrix(full_C);
       printf("Gold:\n");
+#ifdef FAST
+      printf("All elements must be %d\n", MAT_DIM_K + !NO_BIAS);
+#else
       full_printMatrix(gold);
       printf("\n");
+#endif // ifdef FAST
 
       exit(1);
     }

bareMetalC/tiled_matmul_ws_Bt.c

@@ -1,4 +1,3 @@
-// See LICENSE for license details.
 
 #include <stdint.h>
 #include <stddef.h>
@@ -42,12 +41,6 @@ typedef elem_t ACC_T;
 
 #endif // ifdef FAST
 
-#ifdef FAST
-#define RAND 1
-#else
-#define RAND rand()
-#endif
-
 void print_tile(elem_t* in, int tile_dim) {
   for (size_t r = 0; r < tile_dim; r++) {
     printf("row starts at: %p\n", in +r*MAT_DIM_J);
@@ -144,18 +137,17 @@ int main() {
     for (size_t i = 0; i < MAT_DIM_I; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
 #ifdef FAST
-        full_D[i][j] = 0;
+        full_D[i][j] = NO_BIAS ? 0 : 1;
 #else
         full_D[i][j] = NO_BIAS ? 0 : rand() % 2;
 #endif
       }
     }
 
 #ifdef FAST
-    // printf("Init D\n");
     for (size_t i = 0; i < MAT_DIM_I; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
-        gold[i][j] = MAT_DIM_K;
+        gold[i][j] = MAT_DIM_K + !NO_BIAS;
       }
     }
 #else
@@ -191,7 +183,7 @@ int main() {
 
       printf("Gold:\n");
 #ifdef FAST
-      printf("All elements must be %d\n", MAT_DIM_K);
+      printf("All elements must be %d\n", MAT_DIM_K + !NO_BIAS);
 #else
       full_printMatrix(gold);
       printf("\n");

bareMetalC/tiled_matmul_ws_full_C.c

@@ -74,30 +74,46 @@ int main() {
     static acc_t gold[MAT_DIM_I][MAT_DIM_J];
 
 #if CHECK_RESULT == 1
+
+#ifdef FAST
+#define RAND 1
+#else
+#define RAND rand()
+#endif
+
     // printf("Init A\n");
     for (size_t i = 0; i < MAT_DIM_I; ++i) {
       for (size_t j = 0; j < MAT_DIM_K; ++j) {
-        full_A[i][j] = rand() % 2;
+        full_A[i][j] = RAND % 2;
       }
     }
 
     // printf("Init B\n");
     for (size_t i = 0; i < MAT_DIM_K; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
-        full_B[i][j] = rand() % 2;
+        full_B[i][j] = RAND % 2;
       }
     }
 
     // printf("Init D\n");
     for (size_t i = 0; i < MAT_DIM_I; ++i) {
       for (size_t j = 0; j < MAT_DIM_J; ++j) {
-        full_D[i][j] = NO_BIAS ? 0 : rand() % 2;
+        full_D[i][j] = NO_BIAS ? 0 : RAND % 2;
       }
     }
 
     printf("Starting slow CPU matmul\n");
     unsigned long cpu_start = read_cycles();
+#ifdef FAST
+    for (size_t i = 0; i < MAT_DIM_I; ++i) {
+      for (size_t j = 0; j < MAT_DIM_J; ++j) {
+        gold_full[i][j] = MAT_DIM_K + (NO_BIAS ? 0 : (RAND % 2));
+      }
+    }
+
+#else
     full_matmul(full_A, full_B, full_D, gold);
+#endif
     unsigned long cpu_end = read_cycles();
     printf("Cycles taken: %u\n", cpu_end-cpu_start);
 #endif

bareMetalC/tiled_matmul_ws_perf.c

@@ -49,6 +49,8 @@
 #endif
 
 int main() {
+#ifndef FAST
+
 #ifndef BAREMETAL
     if (mlockall(MCL_CURRENT | MCL_FUTURE) != 0) {
       perror("mlockall failed");
@@ -99,6 +101,8 @@ int main() {
     const int utilization = 100 * ideal_cycles / (end-start);
     printf("Utilization: %d%%\n", utilization);
 
+#endif // ifndef FAST
+
   exit(0);
 }