Changeset 4232

Timestamp:

Oct 12, 2014, 5:14:35 PM (4 years ago)

Author:

linmengl

Message:

micro benchmark with Agner's testp ready

Location:

trunk/lib_ir

Files:

• AgnerTestP/DriverSrcLinux (modified) (1 prop)
• AgnerTestP/PMCTest/Makefile (modified) (2 diffs)
• AgnerTestP/PMCTest/PMCTestB.pytemplate.cpp (modified) (2 diffs)
• AgnerTestP/PMCTest/gen.py (modified) (2 diffs)
• CMakeLists.txt (modified) (2 diffs)
• xmlwf/src/xmlwf.cpp (modified) (1 diff)

trunk/lib_ir/AgnerTestP/PMCTest/Makefile

@@ -1 @@
-all: llc_func.ll PMCTestA.cpp PMCTestB.cpp
-        clang PMCTestB.cpp -c -emit-llvm -msse2 -m64 -o B.bc -I ../../ -lpthread
+all: llc_func.ll PMCTestA.cpp B.bc
         llvm-as-3.5 llc_func.ll -o llc_func.bc
         llvm-link B.bc llc_func.bc -o B_llc.bc
@@ -7 +6 @@
         g++ -O2 -m64 PMCTestA.cpp B_llc_opt.o -lpthread -o PMCTest
 
+asm: B_llc_opt.bc
+        llc-svn -O3 -mattr=+sse2 B_llc_opt.bc
+        vim B_llc_opt.s
+
+single: single_llc_func.ll PMCTestA.cpp B.bc
+        llvm-as-3.5 single_llc_func.ll -o llc_func.bc
+        llvm-link B.bc llc_func.bc -o B_llc.bc
+        opt-3.5 -O3 -std-compile-opts -mattr=+sse2 B_llc.bc -o B_llc_opt.bc
+        llc-3.5 -O3 -mattr=+sse2 -filetype=obj B_llc_opt.bc -o B_llc_opt.o
+        g++ -O2 -m64 PMCTestA.cpp B_llc_opt.o -lpthread -o PMCTest
+
+B.bc: PMCTestB.cpp
+        clang PMCTestB.cpp -c -emit-llvm -msse2 -m64 -o B.bc -I ../../ -lpthread
+
 clean:
         rm -f *.bc *.o a.out PMCTest PMCTestB.cpp

trunk/lib_ir/AgnerTestP/PMCTest/PMCTestB.pytemplate.cpp

@@ -131 +131 @@
 
 SIMD_type UserData[NUM_THREADS][USER_DATA_SIZE];
-SIMD_type one_constant = simd<4>::himask();
+SIMD_type one_constant = simd<2>::himask();
 
 
@@ -231 +231 @@
         // Put the code to test here,
         // or a call to a function defined in a separate module        
-
+        {% if NoLoop %}
+          {% for ii in range(0, TestInstrCount) %}
+          UserData[thread][{{ii}}] = {{ TestInstr | replace("[i]", "[" ~ ii ~ "]") }};
+          {% endfor %}
+        {% else %}
         for (i = 0; i < {{TestInstrCount}}; i++) UserData[thread][i] = {{TestInstr}};
+        {% endif %}
 
 

trunk/lib_ir/AgnerTestP/PMCTest/gen.py

@@ -1 +1 @@
 from jinja2 import Template, Environment, FileSystemLoader
+import subprocess
+import argparse
+import sys
 
 Template    = "PMCTestB.pytemplate.cpp"
 Output      = "PMCTestB.cpp"
+
+FunctionListI1 = ["add", "sub", "mult", "eq", "lt", "gt", "ult", "ugt"]
+FunctionListI2 = ["add", "sub", "mult", "eq", "lt", "gt", "ult", "ugt", "vsll", "vsrl", "vsra"]
+FunctionListI4 = ["add", "sub", "mult", "eq", "lt", "gt", "ult", "ugt", "vsll", "vsrl", "vsra"]
+FunctionListI8 = ["add", "sub", "mult", "eq", "lt", "gt", "ult", "ugt", "vsll", "vsrl", "vsra"]
+
+FW = 2
+
+def C(name):
+    return "{name}_{fw}".format(name=name, fw=FW)
+
+def CXX(name):
+    if name in ["vsll", "vsrl", "vsra"]:
+        return "simd<{fw}>::{name}i".format(fw=FW, name=name[1:])
+    return "simd<{fw}>::{name}".format(fw=FW, name=name)
 
 #configuring Jinja2
@@ -14 +32 @@
         f.write(content)
 
-def genFile():
+def genFile(params_dict):
     fillTemplate(template=Template, output=Output,
-                 params=dict(DeclaredIRFunc = "add_1",
-                             DataSize = 1000,
-                             UserDataInit = "simd<32>::constant<0>()",
-                             TestInstrCount = 500,
-                             TestInstr = "simd<1>::add(UserData[thread][i], one_constant)"
-                             ))
+                 params=params_dict)
+
+header = "Scale, Core, Core cyc, Instruct, Uops, L1D Miss"
+
+def extract(func):
+    if subprocess.call("llvm-extract-3.5 -func {0} -S llc_func.ll > single_llc_func.ll".format(func),
+            shell=True) != 0:
+        raise Exception("Extract IR func failed.")
+
+def run(single=False):
+    if single:
+        if subprocess.call("timeout 30s make single", shell=True) != 0:
+            raise Exception("Make failed.")
+    else:
+        if subprocess.call("make", shell=True) != 0:
+            raise Exception("Make failed.")
+    output = subprocess.check_output("./PMCTest")
+    output = output.splitlines()
+    return output[6].split()
 
 if __name__ == '__main__':
-    genFile()
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-i", "--instrcount", action="store_true",
+                    help="Calc same operation with increasing instr count.")
+    parser.add_argument("-f", "--iterfunc", action="store_true",
+                    help="Calc with different function")
+    parser.add_argument("-m", "--master-iterfunc", action="store_true",
+                    help="Calc with different function, with LLVM 3.5")
+    parser.add_argument("-c", "--count", type=int,
+                    help="Give instruction count")
+    parser.add_argument("-e", "--func",
+                    help="Specify one function to generate")
+    parser.add_argument("-w", "--fw", type=int,
+                    help="Field width used with --func and --iterfunc")
+    parser.add_argument("-a", "--idisa", action="store_true",
+                    help="Call IDISA function (default)")
+    parser.add_argument("-b", "--ir", action="store_true",
+                    help="Call IR function")
+    parser.add_argument("-n", "--noloop", action="store_true",
+                    help="Do not use loop in perf program")
+    parser.add_argument("-s", "--simple", action="store_true",
+                    help="Simple run with add_1 for testing")
+    parser.add_argument("-o", "--output",
+                    help="Output perf file")
+    args = parser.parse_args()
 
+    params = dict(DeclaredIRFunc = "add_1",
+                  DataSize = 2000,
+                  UserDataInit = "simd<32>::constant<0>()",
+                  TestInstrCount = 1000,
+                  TestInstr = "add_1(UserData[thread][i], one_constant)",
+                  NoLoop = False
+                  )
+    if args.ir and args.idisa:
+        raise Exception("Cannot select both IR and IDISA")
+    if args.noloop:
+        params["NoLoop"] = True
+    if args.count:
+        params["TestInstrCount"] = args.count
+    if args.master_iterfunc and args.idisa:
+        raise Exception("Cannot select IDISA with --master-iterfunc")
+
+    perf_output = "perf_chat.csv"
+    if args.output:
+        perf_output = args.output
+
+    EXEC = CXX
+    if args.ir:
+        EXEC = C
+
+    if args.simple:
+        params["TestInstrCount"] = args.count
+        genFile(params)
+        print header
+        print run()
+        sys.exit(0)
+
+    if args.func:
+        if args.fw:
+            FW = args.fw
+
+        name = args.func
+        params["DeclaredIRFunc"] = C(name)
+        params["TestInstr"] = EXEC(name) + "(UserData[thread][i], one_constant)"
+        genFile(params)
+        print run()
+        sys.exit(0)
+
+    if args.instrcount:
+        with open(perf_output, "w") as f:
+            f.write(header + "\n")
+
+        for count in xrange(100, 2000, 100):
+            params["TestInstrCount"] = count
+
+            genFile(params)
+
+            with open(perf_output, "a") as f:
+                f.write(str(count) + "," + ",".join(run()) + "\n")
+        sys.exit(0)
+
+    if args.iterfunc:
+        with open(perf_output, "w") as f:
+            f.write(header + "\n")
+        if args.fw:
+            FW = args.fw
+        for name in FunctionListI1:
+            params["DeclaredIRFunc"] = C(name)
+            params["TestInstr"] = EXEC(name) + "(UserData[thread][i], one_constant)"
+
+            genFile(params)
+            with open(perf_output, "a") as f:
+                f.write(name + "," + ",".join(run()) + "\n")
+        sys.exit(0)
+
+    if args.master_iterfunc:
+        with open(perf_output, "w") as f:
+            f.write(header + "\n")
+        if args.fw:
+            FW = args.fw
+        for name in FunctionListI1:
+            params["DeclaredIRFunc"] = C(name)
+            params["TestInstr"] = C(name) + "(UserData[thread][i], one_constant)"
+
+            genFile(params)
+            try:
+                extract(C(name))
+                with open(perf_output, "a") as f:
+                    f.write(name + "," + ",".join(run(single=True)) + "\n")
+            except Exception as e:
+                with open(perf_output, "a") as f:
+                    f.write(name + ", " + str(e) + "\n")
+        sys.exit(0)
+

trunk/lib_ir/CMakeLists.txt

@@ -7 +7 @@
 if (USE_AVX2)
     message("---- Compiling for AVX2")
-    set(CXX_SSE_FLAGS "-msse2 -m64") #"-march=core-avx2")
-    set(LLVM_SSE_FLAGS "-mattr=+sse,+sse2")
-    #set(LLVM_SSE_FLAGS -mattr=+sse,+sse2,+avx2,+bmi2,+bmi -march=x86-64)
+    #set(CXX_SSE_FLAGS "-msse2 -m64")
+    set(CXX_SSE_FLAGS "-mavx2 -m64 -march=core-avx2")
+    #set(LLVM_SSE_FLAGS "-mattr=+sse,+sse2")
+    set(LLVM_SSE_FLAGS -mattr=+sse,+sse2,+avx2,+bmi2,+bmi -march=x86-64)
 else (USE_AVX2)
     set(CXX_SSE_FLAGS "-msse2")
@@ -35 +36 @@
 # gen/ is used to test llc backend.
 add_subdirectory(gen)
+
+add_subdirectory(AgnerTestP/PMCTest)
 
 # test packh_2,4,8 only on avx2

trunk/lib_ir/xmlwf/src/xmlwf.cpp

@@ -17 +17 @@
 #include <sys/stat.h>
 
-#include <simd-lib/s2p.h>
 #include <simd-lib/buffer.hpp>
 #include <simd-lib/bitblock_iterator.hpp>