Sum of Absolute Differences Benchmark

examples/sum_abs_diff/Makefile

@@ -0,0 +1,161 @@
+# Copyright (c) 2019, University of Washington All rights reserved.
+# 
+# Redistribution and use in source and binary forms, with or without modification,
+# are permitted provided that the following conditions are met:
+# 
+# Redistributions of source code must retain the above copyright notice, this list
+# of conditions and the following disclaimer.
+# 
+# Redistributions in binary form must reproduce the above copyright notice, this
+# list of conditions and the following disclaimer in the documentation and/or
+# other materials provided with the distribution.
+# 
+# Neither the name of the copyright holder nor the names of its contributors may
+# be used to endorse or promote products derived from this software without
+# specific prior written permission.
+# 
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+# ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+# WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR
+# ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+# (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+# LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+# ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+# SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+################################################################################
+# Paths / Environment Configuration
+################################################################################
+_REPO_ROOT ?= $(shell git rev-parse --show-toplevel)
+CURRENT_PATH := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
+
+-include $(_REPO_ROOT)/environment.mk
+
+################################################################################
+# Define BSG_MACHINE_PATH, the location of the Makefile.machine.include file
+# that defines the machine to compile and simulate on. Using BSG_F1_DIR (which
+# is set in environment.mk) uses the same machine as in bsg_replicant.
+################################################################################
+
+BSG_MACHINE_PATH=$(BSG_F1_DIR)
+
+################################################################################
+# Define the range of versions
+################################################################################
+# Kernel versions. See kernel/README.md for more information.  Version names do
+# not need to use v* and can be any string
+VERSIONS = v0 v1 v2 v3 v4 v5 v6 v7 v8 v9 v10
+
+################################################################################
+# Define any sources that should be used compiled during kernel compilation,
+# including the source file with the kernel itself. kernel.riscv will
+# be the name of the compiled RISC-V Binary for the Manycore
+#
+# Use KERNEL_*LIBRARIES list sources that should be compiled and linked with all
+# kernel.cpp versions. However, if you have version-specific sources you must
+# come up with your own solution.
+# 
+# Use KERNEL_INCLUDES to specify the path to directories that contain headers.
+################################################################################
+
+# C Libraries
+KERNEL_CLIBRARIES   +=
+# C++ Libraries
+KERNEL_CXXLIBRARIES +=
+
+KERNEL_INCLUDES     += -I$(CURRENT_PATH)/kernel/include
+
+# Define the default kernel.cpp file. If KERNEL_DEFAULT is not defined it will
+# be set to kernel.cpp in the same directory as this Makefile.
+DEFAULT_VERSION     := v0
+KERNEL_DEFAULT      := kernel/$(DEFAULT_VERSION)/kernel.cpp
+
+################################################################################
+# Include the kernel build rules (This must be included after KERNEL_*LIBRARIES,
+# KERNEL_DEFAULT, KERNEL_INCLUDES, etc)
+################################################################################
+
+-include $(FRAGMENTS_PATH)/kernel/cudalite.mk
+
+################################################################################
+# END OF KERNEL-SPECIFIC RULES / START OF HOST-SPECIFIC RULES
+################################################################################
+
+
+################################################################################
+# Define the $(HOST_TARGET), the name of the host executable to generate. The
+# cosimulation host executable will be called
+# $(HOST_TARGET).cosim. HOST_*SOURCES list the host files that should be
+# compiled and linked into the executable.
+################################################################################
+
+HOST_TARGET         := sum_abs_diff
+HOST_CSOURCES       := 
+HOST_CXXSOURCES     := $(HOST_TARGET).cpp
+HOST_INCLUDES       := -I$(CURRENT_PATH)
+
+################################################################################
+# Include the Cosimulation host build rules (This must be included after
+# HOST_*SOURCES, HOST_TARGET, HOST_INCLUDES, etc)
+################################################################################
+
+-include $(FRAGMENTS_PATH)/host/cosim.mk
+
+################################################################################
+# Define the clean rules. clean calls the makefile-specific cleans, whereas
+# users can add commands and dependencies to custom.clean.
+################################################################################
+version.clean:
+	rm -rf kernel/*/*{.csv,.log,.rvo,.riscv,.vpd,.key,.png,.dis}
+	rm -rf kernel/*/{stats,pc_stats}
+
+custom.clean: version.clean
+
+clean: cosim.clean analysis.clean cudalite.clean custom.clean
+
+################################################################################
+# Define overall-goals. The all rule runs all kernel versions, and the default
+# kernel.
+################################################################################
+
+_HELP_STRING := "Makefile Rules\n"
+
+_HELP_STRING += "    default: \n"
+_HELP_STRING += "        - Run the default kernel ($KERNEL_DEFAULT) and generate all of the\n"
+_HELP_STRING += "          analysis products\n"
+default: stats graphs pc_stats
+
+_HELP_STRING += "    analysis: \n"
+_HELP_STRING += "        - Launch indpendent cosimulation executions of each kernel version.\n"
+_HELP_STRING += "          When execution finishes, it generates all the analysis products \n"
+_HELP_STRING += "          for each kernel in each respective kernel/<version_name>/ \n"
+_HELP_STRING += "          directory\n"
+analysis: $(foreach v,$(VERSIONS),kernel/$v/stats kernel/$v/graphs kernel/$v/pc_stats)
+
+_HELP_STRING += "    statistics: \n"
+_HELP_STRING += "        - Launch indpendent cosimulation executions of each kernel version.\n"
+_HELP_STRING += "          When execution finishes, it generates ONLY the parsed operation \n"
+_HELP_STRING += "          stats for each kernel in each respective kernel/<version_name>/ \n"
+_HELP_STRING += "          directory\n"
+statistics: $(foreach v,$(VERSIONS),kernel/$v/stats)
+
+_HELP_STRING += "    all: \n"
+_HELP_STRING += "        - Launch both the default and analysis target\n"
+all: analysis default
+
+.DEFAULT_GOAL = help
+_HELP_STRING += "    help: \n"
+_HELP_STRING += "        - Output a friendly help message.\n"
+help:
+	@echo -e $(HELP_STRING)
+
+# Always re-run, if asked.
+.PHONY: default analysis help
+
+# These last three lines ensure that _HELP_STRING is appended to the top of
+# whatever else comes before it.
+_HELP_STRING += "\n"
+_HELP_STRING += $(HELP_STRING)
+HELP_STRING := $(_HELP_STRING)

examples/sum_abs_diff/README.md

@@ -0,0 +1,106 @@
+# Sum of Absolute Differences
+
+This example runs sum of absolute differences on a refernce matrix
+and a frame matrix. Asuming parameters for the height and width of 
+reference and frame matrixes, the result is a matrix with dimensions
+(reference width - frame width + 1) x (reference height - frame height +1).
+
+Element (x,y) in the result matrix, is the sum of absolute difference between
+the frame, and the sub-matrix of the reference starting from position (x,y) with 
+the same height and width as the frame. The purpose of this kernel is to find the 
+porition of the reference matrix that has the most resemblence to the frame matrix.
+
+The kernel code is located in the subdirectories of [kernel](kernel). The actual
+sum of absolute differences code is in the header file
+[kernel/include/sum_abs_diff.hpp](kernel/include/sum_abs_diff.hpp). 
+
+# Makefile Targets
+
+For a list of all Makefile targets, run `make help`.
+
+## Versions
+
+There are several different versions of this kernel. Each is a subdirectory in
+the [kernel](kernel) directory.
+
+### Version 0
+In this version, each tile only performs one unit of work, i.e. calculating one
+element of the result matrix. Based on the size of the result matrix, and the 
+tile group dimensions, enough tile groups will be launched to populate the 
+entire result matrix.
+
+
+### Version 1
+
+In this version, each tile performs multiple units of work. The workload share 
+of each tile group is defined by block_size_x/y and passed to the kernel. Each 
+tile then performs the kernel in a loop, with each iteration belonging to a 
+unit of work (or vitual thread).
+
+
+### Version 2
+
+In this version, frame dimensions are fixed at 4 by 4.
+
+
+### Version 3
+
+In this version, the frame dimensions are templatized in the kernel, instead 
+of being passed in as an input argument to the kernel. This gives the compiler
+the opportunity to optimize based on frame dimensions known at compiler time.
+
+
+### Version 4
+
+In this version, the reference dimensions are templatized in the kernel, instead 
+of being passed in as an input argument to the kernel. This gives the compiler
+the opportunity to optimize based on reference dimensions known at compiler time.
+
+
+### Version 5
+
+Version 5 - Refernce and frame dimensions templatized
+In this version, both the reference and frame dimensions are templatized 
+in the kernel, instead of being passed in as an input argument to the kernel.
+This gives the compiler the opportunity to optimize based on reference 
+dimensions known at compiler time.
+
+
+### Version 6
+
+This version loads the frame into tile group shared memory using tile group
+shared memory macros, and uses the shared memory for redundant accesses to 
+the frame.
+
+
+### Version 7
+
+This version uses tile group shared memory macros for storing refernce & frame 
+Due to redundant accesses to DRAM, performance can be improved by loading reference 
+and frame matrxix into shared meomry and using that for compuation.
+
+This version is currently not working.
+
+
+### Version 8
+
+This version loads frame into tile group shared memory similar to version 7,
+however it uses the new replacement for the deprecated tile group shared memory
+macros, the striped array.
+
+
+### Version 9
+
+Similar to version 8, but uses the new templatized barrier library instead of 
+the deprecated barrier macros.
+
+
+### Version 10
+
+This version combines all optimizations together:
+ * Templatized reference and frame dimensions
+ * Frame is loaded into tile group shared striped memory
+ * Templatized barrier library is used for synchronization
+
+
+

examples/sum_abs_diff/kernel/include/sum_abs_diff.hpp

@@ -0,0 +1,10 @@
+#ifndef __SUM_ABS_DIFF_HPP
+#define __SUM_ABS_DIFF_HPP
+#include <cstdint>
+
+#define ABS(x) ((x) >= 0) ? (x) : (-(x))
+#define MIN(x,y) (((x) <= (y)) ? (x) : (y))
+
+
+
+#endif //__SUM_ABS_DIFF_HPP

examples/sum_abs_diff/kernel/v0/kernel.cpp

@@ -0,0 +1,81 @@
+/*
+ * This kernel performs sum of absolute differences (SAD) between a 
+ * Reference matrix and a frame matrix. 
+ * The kernel searches over the reference matrix, and calculates 
+ * sum of absolute differnces for every sub-matrix that matches 
+ * the sizes of the frame matrix, and stores the result for that 
+ * location into result matrix. 
+ */
+
+// BSG_TILE_GROUP_X_DIM and BSG_TILE_GROUP_Y_DIM must be defined
+// before bsg_manycore.h and bsg_tile_group_barrier.h are
+// included. bsg_tiles_X and bsg_tiles_Y must also be defined for
+// legacy reasons, but they are deprecated.
+#define BSG_TILE_GROUP_X_DIM 4
+#define BSG_TILE_GROUP_Y_DIM 4
+#define bsg_tiles_X BSG_TILE_GROUP_X_DIM
+#define bsg_tiles_Y BSG_TILE_GROUP_Y_DIM
+#include <bsg_manycore.h>
+#include <bsg_tile_group_barrier.h>
+#include <cstdint>
+
+#include <sum_abs_diff.hpp>
+
+INIT_TILE_GROUP_BARRIER(r_barrier, c_barrier,
+                        0, BSG_TILE_GROUP_X_DIM-1,
+                        0, BSG_TILE_GROUP_Y_DIM-1);
+
+
+/* 
+ * Version 0 - Singler work per tile 
+ * In this version, each tile only performs one unit of work, i.e. calculating one
+ * element of the result matrix. Based on the size of the result matrix, and the 
+ * tile group dimensions, enough tile groups will be launched to populate the 
+ * entire result matrix.
+ */
+int  __attribute__ ((noinline)) sum_abs_diff_single_work_per_tile (int *REF, int *FRAME, int *RES,
+                                                                   uint32_t ref_height, uint32_t ref_width,
+                                                                   uint32_t frame_height, uint32_t frame_width,
+                                                                   uint32_t res_height, uint32_t res_width) {
+
+        int start_y = __bsg_tile_group_id_y * bsg_tiles_Y + bsg_y;
+        int end_y   = MIN (start_y + frame_height, res_height);
+        int start_x = __bsg_tile_group_id_x * bsg_tiles_X + bsg_x;
+        int end_x   = MIN (start_x + frame_width, res_width);
+
+
+        int sad = 0;
+        for (int y = start_y; y < end_y; y ++) {
+                for (int x = start_x; x < end_x; x ++) {
+                        sad += ABS ( (REF[y * ref_width + x] - FRAME[(y - start_y) * frame_width + (x - start_x)]) );
+                }
+        }
+
+        RES[start_y * res_width + start_x] = sad;
+        return 0;
+}
+
+
+
+extern "C" {
+        int  __attribute__ ((noinline)) kernel_sum_abs_diff(
+                      int *REF, int *FRAME, int *RES,
+                      uint32_t REF_HEIGHT, uint32_t REF_WIDTH,
+                      uint32_t FRAME_HEIGHT, uint32_t FRAME_WIDTH,
+                      uint32_t RES_HEIGHT, uint32_t RES_WIDTH,
+                      uint32_t block_size_y, uint32_t block_size_x) {
+                int rc;
+                bsg_cuda_print_stat_kernel_start();
+                bsg_cuda_print_stat_start(0);
+                rc = sum_abs_diff_single_work_per_tile  (REF, FRAME, RES,
+                                                         REF_HEIGHT, REF_WIDTH,
+                                                         FRAME_HEIGHT, FRAME_WIDTH,
+                                                         RES_HEIGHT, RES_WIDTH);
+                bsg_cuda_print_stat_end(0);
+
+                bsg_tile_group_barrier(&r_barrier, &c_barrier);
+
+                bsg_cuda_print_stat_kernel_end();
+                return rc;
+        }
+}