Update microbenchmarks 04/2020

.gitignore

@@ -3,3 +3,4 @@
 *.out
 *.log
 __pycache__
+*.sw[p,o]

kw/boot/Makefile

@@ -4,11 +4,11 @@
 # Builds the PULP bootloader without the run time. We need a separate link
 # script since the code is loacted at a different static location (boot ROM).
 
-PULP_APP = boot-oprecompkw
+PULP_APP = rom-oprecompkw
 PULP_APP_FC_SRCS = boot.c
 PULP_APP_FC_ASM_SRCS = crt0.S
 PULP_CFLAGS = -O3 -g -fno-jump-tables
-PULP_LDFLAGS = -Tlink.ld -nostdlib
-CONFIG_OPT = options/rt/no-crt0=true options/rt/no-link-script=true options/rt/no-rt=true
+PULP_LDFLAGS = -Tlink.ld -nostdlib -lgcc
+CONFIG_OPT = rt/no-crt0=true rt/no-link-script=true rt/no-rt=true
 
 include $(PULP_SDK_HOME)/install/rules/pulp_rt.mk

kw/boot/boot.c

@@ -26,7 +26,7 @@ inline uint64_t host2local(uint64_t host) {
 
 
 // Rudimentary putc/puts for output.
-#define boot_putc_address (STDOUT_BASE_ADDR + STDOUT_PUTC_OFFSET + (hal_core_id()<<3) + (hal_cluster_id()<<7))
+#define boot_putc_address (ARCHI_STDOUT_ADDR + STDOUT_PUTC_OFFSET + (hal_core_id()<<3) + (hal_cluster_id()<<7))
 
 static void boot_putc(unsigned enable, unsigned int c) {
 	if (enable) {

kw/boot/link.ld

@@ -6,7 +6,7 @@ __DYNAMIC  =  0;
 MEMORY
 {
 	ROM : ORIGIN = 0x1A000000, LENGTH = 0x2000
-	L1  : ORIGIN = 0x10000000, LENGTH = 0x10000
+	L1  : ORIGIN = 0x10000004, LENGTH = 0xFFFC
 	L2  : ORIGIN = 0x1C000000, LENGTH = 0x4000
 }
 

kw/libcxl/libcxl.c

@@ -9,7 +9,7 @@
 #include <pthread.h>
 
 #include "libcxl.h"
-#include "gv/gv_launcher.h"
+#include "vp/launcher.h"
 
 
 struct cxl_afu_h {
@@ -119,14 +119,14 @@ afu_open(struct cxl_afu_h *afu) {
 	}
 
 	// Configure the handles into the platform.
-	afu->capi_binding = gv_ioreq_binding(afu->gv, "soc/soc_ico", (void *)((uint64_t)1 << 48), (uint64_t)1 << 48, capi_callback, afu);
+	afu->capi_binding = gv_ioreq_binding(afu->gv, "soc/host_injector", (void *)((uint64_t)1 << 48), (uint64_t)1 << 48, capi_callback, afu);
 	// afu->capi_binding = gv_ioreq_binding(afu->gv, "soc/soc_ico", (void *)0x1a600000, 0x00100000, capi_callback, afu);
 	if (afu->capi_binding == NULL) {
 		fprintf(stderr, "libcxl error: unable to establish CAPI binding (gv_ioreq_binding)\n");
 		goto fail;
 	}
 
-	afu->wed_binding = gv_ioreq_binding(afu->gv, "job_fifo", NULL, 0, NULL, NULL);
+	afu->wed_binding = gv_ioreq_binding(afu->gv, "job_fifo_injector", NULL, 0, NULL, NULL);
 	if (afu->wed_binding == NULL) {
 		fprintf(stderr, "libcxl error: unable to establish WED FIFO binding (gv_ioreq_binding)\n");
 		goto fail;

kw/liboprecomp/Makefile

@@ -38,4 +38,5 @@ install:
 	install -D build/liboprecomp.a $(PULP_SDK_HOME)/install/ws/lib
 	install -D build/liboprecomp.so $(PULP_SDK_HOME)/install/ws/lib
 	install -D liboprecomp.h $(PULP_SDK_HOME)/install/ws/include
-	plpinfo tree --config-options="gvsoc/load-binary_eval=os.environ.get('PULP_SDK_HOME') + '/install/bin/boot-oprecompkw'" > $(PULP_SDK_HOME)/install/cfg/oprecompkw_config.json
+	mkdir -p $(PULP_SDK_HOME)/install/cfg
+	plpinfo tree --config-options="plt_loader/load-binary_eval=os.environ.get('PULP_SDK_HOME') + '/install/bin/rom-oprecompkw'" --config-options="soc/cluster/pe0/fetch_enable=true" --config-options="chip/boot_from_rom=true" --config-options="loader/boot/mode=rom" > $(PULP_SDK_HOME)/install/cfg/oprecompkw_config.json

kw/mb/blstm/Makefile

@@ -6,4 +6,4 @@ all:
 	$(MAKE) -C host
 
 run: all
-	host/build/blstm pulp/build/system.oprecompkw/blstm/blstm
+	host/build/blstm pulp/build/oprecompkw/blstm/blstm

kw/mb/blstm/host/Makefile

@@ -5,7 +5,7 @@ CFLAGS ?= -O3 -g -Wall -Wextra -std=c99
 CPPFLAGS ?= -O3 -g -Wall -Wextra -std=c++11 -fopenmp
 CFLAGS += -I$(PULP_SDK_HOME)/install/ws/include
 CPPFLAGS += -I$(PULP_SDK_HOME)/install/ws/include -I./include
-LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lgvlauncher -lpthread
+LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lpulpvplauncher -lpthread
 
 all: clean build/neuron build/blstm
 

kw/samples/nop/Makefile

@@ -6,4 +6,4 @@ all:
 	$(MAKE) -C host
 
 run: all
-	host/build/nop pulp/build/system.oprecompkw/nop/nop
+	host/build/nop pulp/build/oprecompkw/nop/nop

kw/samples/nop/host/Makefile

@@ -3,7 +3,7 @@
 
 CFLAGS ?= -O3 -g -Wall -Wextra
 CFLAGS += -I$(PULP_SDK_HOME)/install/ws/include
-LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lgvlauncher -lpthread
+LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lpulpvplauncher -lpthread
 
 all:: build/nop
 

kw/samples/nop/testset.cfg

@@ -0,0 +1,25 @@
+from plptest import *
+
+TestConfig = c = {}
+
+def check_output(config, output):
+
+  if output.find('Hello from the loaded PULP binary!') == -1:
+    return (False, "Didn't find output string")
+
+  return (True, None)
+
+
+def get_test():
+  return Test(
+    name = 'nop',
+    commands = [
+      Shell('build', 'make all run'),
+      Check('check', check_output)
+    ],
+    timeout=1000000000
+  )
+
+c['tests'] = [ ]
+
+c['tests'].append(get_test())

kw/samples/square/Makefile

@@ -6,4 +6,4 @@ all:
 	$(MAKE) -C host
 
 run: all
-	host/build/square pulp/build/system.oprecompkw/square/square
+	host/build/square pulp/build/oprecompkw/square/square

kw/samples/square/host/Makefile

@@ -3,7 +3,7 @@
 
 CFLAGS ?= -O3 -g -Wall -Wextra -std=c99
 CFLAGS += -I$(PULP_SDK_HOME)/install/ws/include
-LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lgvlauncher -lpthread
+LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lpulpvplauncher -lpthread
 
 all:: build/square
 

kw/samples/square/testset.cfg

@@ -0,0 +1,25 @@
+from plptest import *
+
+TestConfig = c = {}
+
+def check_output(config, output):
+
+  if output.find('all correct') == -1:
+    return (False, "Didn't find output string")
+
+  return (True, None)
+
+
+def get_test():
+  return Test(
+    name = 'square',
+    commands = [
+      Shell('build', 'make all run'),
+      Check('check', check_output)
+    ],
+    timeout=1000000
+  )
+
+c['tests'] = [ ]
+
+c['tests'].append(get_test())

mb/cnn/kw/Makefile

@@ -6,4 +6,4 @@ all:
 	$(MAKE) -C host
 
 run: all
-	host/build/conv pulp/build/system.oprecompkw/conv/conv
+	host/build/conv pulp/build/oprecompkw/conv/conv

mb/cnn/kw/host/Makefile

@@ -3,7 +3,7 @@
 
 CFLAGS ?= -O3 -g -Wall -Wextra -std=c99
 CFLAGS += -I$(PULP_SDK_HOME)/install/ws/include
-LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lgvlauncher -lpthread
+LDFLAGS ?= -L$(PULP_SDK_HOME)/install/ws/lib -loprecomp -lcxl -lpulpvplauncher -lpthread
 
 all:: build/conv
 

mb/cnn/kw/pulp/pulp.c

@@ -56,7 +56,7 @@ int main(uint64_t wedptr) {
 
 	// Determine the tiling of the input data via a binary search. This is quite
 	// inefficient and should be improved.
-	const uint32_t max_size = 65536;
+	const uint32_t max_size = 65536 - 0x5000;
 	const uint32_t max_volume = (max_size - (16 << 10)) / 4;
 	printf("Determining Tiling\n");
 	printf("  max_volume = %lu\n", max_volume);
@@ -96,6 +96,11 @@ int main(uint64_t wedptr) {
 	float *tile_x = rt_alloc(RT_ALLOC_CL_DATA, bufsz_x);
 	float *tile_w = rt_alloc(RT_ALLOC_CL_DATA, bufsz_w);
 	float *tile_y = rt_alloc(RT_ALLOC_CL_DATA, bufsz_y);
+	if (tile_x == NULL || tile_y == NULL || tile_w == NULL)
+	{
+	  printf("Failed to allocate memory\n");
+	  return -1;
+	}
 
 	// Perform the outer iterations.
 	for (int32_t ko1 = 0; ko1 < wed.KO; ko1 += TKO) {
@@ -123,25 +128,31 @@ int main(uint64_t wedptr) {
 					const int32_t XM1 = m_hi-m_lo;
 
 					printf("Loading x tile ki1=%lu..%lu, n=%lu..%lu (%ld), m=%lu..%lu (%ld)\n", ki1, ki1+KI1, n_lo, n_hi, XN1, m_lo, m_hi, XM1);
+					int merge = 0;
 					for (int32_t ki2 = 0; ki2 < KI1; ++ki2) {
 						for (int32_t n2 = 0; n2 < XN1; ++n2) {
-							plp_dma_memcpy(
+							plp_dma_memcpy_merge(
 								host2local(wed.in_x + (ki1+ki2)*wed.N*wed.M*4 + (n2+n_lo)*wed.M*4 + m_lo*4),
 								(uint32_t)tile_x + ki2*XN1*XM1*4 + n2*XM1*4,
 								XM1*4,
-								PLP_DMA_EXT2LOC
+								PLP_DMA_EXT2LOC,
+								merge
 							);
+							merge = 1;
 						}
 					}
 					plp_dma_barrier();
+					merge = 0;
 					for (int32_t ko2 = 0; ko2 < KO1; ++ko2) {
 						for (int32_t ki2 = 0; ki2 < KI1; ++ki2) {
-							plp_dma_memcpy(
+							plp_dma_memcpy_merge(
 								host2local(wed.in_w + (ko1+ko2)*wed.KI*wed.U*wed.V*4 + (ki1+ki2)*wed.U*wed.V*4),
 								(uint32_t)tile_w + ko2*TKI*wed.U*wed.V*4 + ki2*wed.U*wed.V*4,
 								wed.U*wed.V*4,
-								PLP_DMA_EXT2LOC
+								PLP_DMA_EXT2LOC,
+								merge
 							);
+							merge = 1;
 						}
 					}
 					plp_dma_barrier();
@@ -171,14 +182,17 @@ int main(uint64_t wedptr) {
 				}
 
 				// Write result tile back.
+				int merge = 0;
 				for (int32_t ko2 = 0; ko2 < KO1; ++ko2) {
 					for (int32_t n2 = 0; n2 < N1; ++n2) {
-						plp_dma_memcpy(
+						plp_dma_memcpy_merge(
 							host2local(wed.out_y + (ko1+ko2)*wed.N*wed.M*4 + (n1+n2)*wed.M*4 + m1*4),
 							(uint32_t)tile_y + ko2*N1*M1*4 + n2*M1*4,
 							M1*4,
-							PLP_DMA_LOC2EXT
+							PLP_DMA_LOC2EXT,
+							merge
 						);
+						merge = 1;
 					}
 				}
 			}

mb/sparsesolve/Makefile

@@ -11,7 +11,7 @@ LDFLAGS = -lm -lrt
 
 build: sparsesolve
 
-sparsesolve: sparsesolve.o mmio.o matrix.o cg.o ir.o oprecomp.o
+sparsesolve: sparsesolve.o mmio.o matrix.o oprecomp.o
 	$(CC) $^ $(LDFLAGS) -o $@
 
 clean:
@@ -21,9 +21,20 @@ tags: *.c *.h
 	ctags *.c *.h
 
 test: sparsesolve
-	./sparsesolve data/prepared/mb/sparsesolve/bcsstk01.mtx 1 1e-7 10000 1e-7 100
+	./sparsesolve data/prepared/mb/sparsesolve/bcsstk01.mtx 1 1e-7 10000 1e-7 52
 
-%.o: %.c cg.h vector.h matrix.h
+plots: bcsstk01.pdf gr_30_30.pdf msc10848.pdf
+
+%.pdf: %.eps
+	epstopdf $<
+
+%.eps: sparsesolve.plt %.dat
+	gnuplot -c $^ $* > $@
+
+%.dat: sparsesolve
+	for i in `seq 1 52`; do ./sparsesolve ../data/prepared/mb/sparsesolve/$*.mtx 10000 1e-7 8 $$i ; done | grep -v '#' > $@
+
+%.o: %.c floatm.h vector.h matrix.h
 	$(CC) $(CFLAGS) -c $< -o $@
 
 oprecomp.o: $(VPATH)../common/oprecomp.c

mb/sparsesolve/floatm.h

@@ -0,0 +1,30 @@
+#ifndef FLOATM
+#define FLOATM double
+#endif
+
+#ifndef DOUBLE
+#define DOUBLE double
+#endif
+
+static inline double double_truncate(uint8_t m, double y)
+{
+    if (m >= 52) return y; // nothing to do
+    if (!isfinite(y)) return y; // NaN and Inf does not need conversion
+    union { double d; uint64_t i; } c;
+    c.d = y;
+    c.i &= ~(((1UL << (52 - m)) - 1UL));
+    return c.d;
+}
+
+static inline float float_truncate(uint8_t m, float y)
+{
+    if (m >= 23) return y; // nothing to do
+    if (!isfinite(y)) return y; // NaN and Inf does not need conversion
+    union { float f; uint32_t i; } c;
+    c.f = y;
+    c.i &= ~(((1U << (23 - m)) - 1U));
+    return c.f;
+}
+
+#define TRUNCATE2(t, m, y) t ## _truncate(m, y)
+#define TRUNCATE(t, m, y) TRUNCATE2(t, m, y)