Sparse DMA (COO) Tests for Gemmini

.gitignore

@@ -6,3 +6,5 @@ build
 autom4te.cache
 configure
 .ipynb_checkpoints
+*.swp
+

Makefile.in

@@ -1,3 +1,4 @@
+<<<<<<< HEAD
 #=============================================================================
 # variables
 #=============================================================================
@@ -7,6 +8,7 @@ XLEN           := @XLEN@
 RISCVTOOLS     := @RISCVTOOLS@
 ROCC           := examples
 CATEGORIES     := bareMetalC mlps imagenet
+RUNNER         := "spike --extension=gemmini "
 
 #=============================================================================
 # targets
@@ -30,3 +32,36 @@ clean:
 	$(MAKE) -C imagenet -f $(abs_top_srcdir)/imagenet/Makefile abs_top_srcdir=$(abs_top_srcdir) PREFIX=$(ROCC)-imagenet clean
 	$(MAKE) -C mlps -f $(abs_top_srcdir)/mlps/Makefile abs_top_srcdir=$(abs_top_srcdir) PREFIX=$(ROCC)-mlps clean
 
+test-baremetal-bareMetalC:
+	make -C bareMetalC	\
+	        -f $(abs_top_srcdir)/bareMetalC/Makefile \
+                TARGET_MAKEFILE=$(abs_top_srcdir)/bareMetalC/Makefile \
+		abs_top_srcdir=$(abs_top_srcdir) \
+	 	src_dir=$(abs_top_srcdir)/bareMetalC \
+	 	XLEN=$(XLEN) \
+	 	PREFIX=$(ROCC)-bareMetalC \
+		RISCVTOOLS=$(RISCVTOOLS) \
+		RUNNER=$(RUNNER) \
+		run-baremetal
+
+test-baremetal: test-baremetal-bareMetalC
+	make -C mlps	\
+	        -f $(abs_top_srcdir)/mlps/Makefile \
+                TARGET_MAKEFILE=$(abs_top_srcdir)/mlps/Makefile \
+		abs_top_srcdir=$(abs_top_srcdir) \
+	 	src_dir=$(abs_top_srcdir)/mlps \
+	 	XLEN=$(XLEN) \
+	 	PREFIX=$(ROCC)-mlps \
+		RISCVTOOLS=$(RISCVTOOLS) \
+		RUNNER=$(RUNNER) \
+	        run-baremetal
+	make -C imagenet	\
+	        -f $(abs_top_srcdir)/imagenet/Makefile \
+                TARGET_MAKEFILE=$(abs_top_srcdir)/imagenet/Makefile \
+		abs_top_srcdir=$(abs_top_srcdir) \
+	 	src_dir=$(abs_top_srcdir)/imagenet \
+	 	XLEN=$(XLEN) \
+	 	PREFIX=$(ROCC)-imagenet \
+		RISCVTOOLS=$(RISCVTOOLS) \
+		RUNNER=$(RUNNER) \
+		run-baremetal

bareMetalC/Makefile

@@ -1,9 +1,13 @@
 
 tests := \
 	mvin_mvout \
+	mvin_mvout_sparse \
 	mvin_mvout_stride \
 	mvin_mvout_acc \
+	mvin_mvout_acc_zero_stride \
 	mvin_mvout_acc_stride \
+	mvin_mvout_acc_full \
+	mvin_mvout_acc_full_stride \
 	matmul_os \
 	matmul_ws \
 	matmul \
@@ -15,9 +19,86 @@ tests := \
 	conv_with_pool \
 	tiled_matmul_os \
 	tiled_matmul_ws \
+	tiled_matmul_gcn_1 \
+	tiled_matmul_gcn_1b \
+	tiled_matmul_gcn_2 \
+	tiled_matmul_gcn_3 \
+	tiled_matmul_gcn_4 \
+	tiled_matmul_gcn_4b \
+	tiled_matmul_ws_At \
+	tiled_matmul_ws_Bt \
+	tiled_matmul_ws_full_C \
+	tiled_matmul_ws_low_D \
 	tiled_matmul_cpu \
 	tiled_matmul_option \
+	tiled_matmul_ws_perf \
 	transpose \
 	template \
 	gemm
+	matrix_add \
+	resadd \
+	template
+
+tests_baremetal = $(tests:=-baremetal)
+
+ifeq ($(findstring spike,$(RUNNER)),spike)
+# Currently don't support conv or conv-with-pool on spike
+runs_baremetal = $(addsuffix .run,$(filter-out conv-baremetal conv_with_pool-baremetal,$(tests_baremetal)))
+else
+# Don't run very long benchmarks for RTL sim
+runs_baremetal = $(addsuffix .run,$(filter-out tiled_matmul_cpu-baremetal tiled_matmul_option-baremetal,$(tests_baremetal)))
+endif
+
+ifdef BAREMETAL_ONLY
+	tests_linux =
+else
+	tests_linux = $(tests:=-linux)
+endif
+
+BENCH_COMMON = $(abs_top_srcdir)/riscv-tests/benchmarks/common
+GEMMINI_HEADERS = $(abs_top_srcdir)/include/gemmini.h $(abs_top_srcdir)/include/gemmini_params.h $(abs_top_srcdir)/include/gemmini_testutils.h
+
+CFLAGS := $(CFLAGS) \
+	-DPREALLOCATE=1 \
+	-DMULTITHREAD=1 \
+	-mcmodel=medany \
+	-std=gnu99 \
+	-O2 \
+	-ffast-math \
+	-fno-common \
+	-fno-builtin-printf \
+	-march=rv64gc -Wa,-march=rv64gcxhwacha \
+	-lm \
+	-lgcc \
+	-I$(abs_top_srcdir)/riscv-tests \
+	-I$(abs_top_srcdir)/riscv-tests/env \
+	-I$(abs_top_srcdir) \
+	-I$(BENCH_COMMON) \
+	-DID_STRING=$(ID_STRING) \
+
+CFLAGS_BAREMETAL := \
+	$(CFLAGS) \
+	-nostdlib \
+	-nostartfiles \
+	-static \
+	-T $(BENCH_COMMON)/test.ld \
+	-DBAREMETAL=1 \
+
+all: $(tests_baremetal) $(tests_linux)
+
+vpath %.c $(src_dir)
+
+%-baremetal: %.c $(GEMMINI_HEADERS)
+	$(CC_BAREMETAL) $(CFLAGS_BAREMETAL) $< $(LFLAGS) -o $@ \
+		$(wildcard $(BENCH_COMMON)/*.c) $(wildcard $(BENCH_COMMON)/*.S) $(LIBS)
+
+%-linux: %.c $(GEMMINI_HEADERS)
+	$(CC_LINUX) $(CFLAGS) $< $(LFLAGS) -o $@
+
+run-baremetal: $(runs_baremetal)
+
+%-baremetal.run: %-baremetal
+	$(RUNNER)$(abs_top_srcdir)/build/bareMetalC/$^
+
+junk += $(tests_baremetal) $(tests_linux)
 

bareMetalC/aligned.c

@@ -41,6 +41,7 @@ int main() {
     }
 
   gemmini_config_ld(DIM * sizeof(elem_t));
+  gemmini_config_st(DIM * sizeof(elem_t));
 
   // printf("Mvin\n");
   gemmini_mvin(In.data, 0);

bareMetalC/conv.c

@@ -15,13 +15,21 @@
 #define PADDING 1
 #define STRIDE 2
 #else
-#define BATCH_SIZE 3
+#ifdef FAST
+#define IN_DIM 9
+#define IN_CHANNELS 5
+#define OUT_CHANNELS 7
+#else
 #define IN_DIM 23
 #define IN_CHANNELS 17
 #define OUT_CHANNELS 31
+#endif
+
+#define BATCH_SIZE 3
 #define KERNEL_DIM 3
 #define PADDING 1
 #define STRIDE 2
+
 #endif
 
 #define NO_BIAS false
@@ -69,9 +77,6 @@ void conv(int batch_size, int in_channels, int in_dim,
                         }
                     }
 
-                    // Shift while rounding to nearest integer (ties round to negative infinity)
-                    result = ROUNDING_RIGHT_SHIFT(result, 0);
-
                     // Clip result
                     result = result > elem_t_max ? elem_t_max : (result < elem_t_min ? elem_t_min : result);
 
@@ -113,16 +118,26 @@ bool vec_is_equal(elem_t * a, elem_t * b, int len) {
 }
 
 void init_random(elem_t * buf, int len) {
+    elem_t i = 0;
     for (elem_t * ptr = buf; ptr < buf + len; ptr++) {
         // *ptr = (rand() % 32) - 16;
-        *ptr = (rand() % 5) - 2;
+#ifdef FAST
+      *ptr = 1;
+#else
+      *ptr = (rand() % 5) - 2;
+#endif
     }
 }
 
 void init_random_acc(acc_t * buf, int len) {
+    elem_t i = 0;
     for (acc_t * ptr = buf; ptr < buf + len; ptr++) {
         // *ptr = (rand() % 32) - 16;
-        *ptr = (rand() % 5) - 2;
+#ifdef FAST
+      *ptr = 1;
+#else
+      *ptr = (rand() % 5) - 2;
+#endif
     }
 }
 
@@ -159,6 +174,7 @@ int main() {
 
     printf("CPU conv...\n");
     uint64_t start_cpu = read_cycles();
+#ifndef FAST
     conv(BATCH_SIZE, IN_CHANNELS, IN_DIM,
             OUT_CHANNELS, KERNEL_DIM,
             OUT_DIM,
@@ -167,6 +183,7 @@ int main() {
             weights,
             bias,
             output);
+#endif
     uint64_t end_cpu = read_cycles();
     printf("CPU conv took %llu cycles\n", end_cpu - start_cpu);
 
@@ -191,15 +208,28 @@ int main() {
         NO_BIAS ? NULL : (acc_t*)bias,
         (elem_t*)output_mat,
 
-        NO_ACTIVATION, 0, 0, 0, 0, 0,
+        NO_ACTIVATION, ACC_SCALE_IDENTITY, 0, 0, 0, 0,
 
         WS);
     uint64_t end_gemmini = read_cycles();
     printf("Gemmini conv took %llu cycles\n", end_gemmini - start_gemmini);
 
     ASSERT(sizeof(output_mat) == sizeof(output), "output_mat size");
 
+#ifdef FAST
+    bool success = true;
+    for (int orow = 0; orow < BATCH_SIZE * OUT_DIM * OUT_DIM; orow++) {
+      for (int ocol = 0; ocol < OUT_CHANNELS; ocol++) {
+	elem_t v = output_mat[orow][ocol];
+	if (v != 21 && v != 31 && v != 46) {
+	  success = false;
+	  break;
+	}
+      }
+    }
+#else
     bool success = vec_is_equal(&output[0][0][0][0], &output_mat[0][0], sizeof(output) / sizeof(elem_t));
+#endif
 
     if (!success) {
         // return 1;
@@ -289,4 +319,3 @@ int main() {
 
     return 0;
 }
-