Merge pull request #20 from mklarqvist/cleanup

General cleanup: function names as used in the paper. Fix compilation issues for Windows.

.gitignore

@@ -34,3 +34,6 @@
 example
 bench
 instrumented_benchmark
+.vscode
+CMakeFiles
+Testing

CMakeLists.txt

@@ -0,0 +1,20 @@
+
+cmake_minimum_required(VERSION 2.8)
+project(libppospopcnt C CXX)
+set(CMAKE_BUILD_TYPE Release)
+include_directories(. benchmark/linux/)
+
+set(CMAKE_CXX_STANDARD 11)  # enable C++11 standard
+set(CMAKE_CXX_FLAGS "-O3 -march=native")
+set(CMAKE_C_FLAGS "-O3 -march=native")
+
+add_executable(bench pospopcnt.c benchmark.cpp)
+add_executable(example pospopcnt.c example.c)
+add_executable(instrumented_benchmark pospopcnt.c benchmark/linux/instrumented_benchmark.cpp)
+
+enable_testing()
+add_test(NAME example COMMAND example)
+add_test(NAME bench COMMAND bench 100000 10)
+add_test(NAME instrumented_benchmark COMMAND instrumented_benchmark -n 100000 -i 10)
+
+install(FILES pospopcnt.h DESTINATION ${CMAKE_INSTALL_PREFIX}/include/pospopcnt)

Makefile

@@ -19,7 +19,7 @@
 OPTFLAGS  := -O3 -march=native
 CFLAGS     = -std=c99 $(OPTFLAGS) $(DEBUG_FLAGS)
 CPPFLAGS   = -std=c++0x $(OPTFLAGS) $(DEBUG_FLAGS)
-CPP_SOURCE = main.cpp benchmark/linux/instrumented_benchmark.cpp
+CPP_SOURCE = benchmark.cpp benchmark/linux/instrumented_benchmark.cpp
 C_SOURCE   = pospopcnt.c example.c
 OBJECTS    = $(CPP_SOURCE:.cpp=.o) $(C_SOURCE:.c=.o)
 
@@ -36,8 +36,8 @@ all: bench
 benchmark/linux/instrumented_benchmark.o : benchmark/linux/instrumented_benchmark.cpp
 	$(CXX) $(CPPFLAGS) -I. -Ibenchmark/linux -c -o $@ $<
 
-bench: pospopcnt.o main.o
-	$(CXX) $(CPPFLAGS) pospopcnt.o main.o -o bench
+bench: pospopcnt.o benchmark.o
+	$(CXX) $(CPPFLAGS) pospopcnt.o benchmark.o -o bench
 
 itest: instrumented_benchmark
 	$(CXX) --version

README.md

@@ -12,19 +12,19 @@ This benchmark shows the speedup of the 3 `pospopcnt` algorithms used on x86 CPU
 
 | Algorithm                         | 128  | 256  | 512  | 1024 | 2048 | 4096 | 8192 | 65536 |
 |-----------------------------------|------|------|------|------|------|------|------|-------|
-| pospopcnt_u16_sse_mula_unroll8    | **2.09** | 3.16 | 2.35 | 1.88 | 1.67 | 1.56 | 1.5  | 1.44  |
-| pospopcnt_u16_avx512_mula_unroll8 | 1.78 | **3.61** | **3.61** | 3.59 | 3.68 | 3.65 | 3.67 | 3.7   |
-| pospopcnt_u16_avx512_mula3        | 0.77 | 0.9  | 3.24 | **3.96** | **4.96** | 5.87 | 6.52 | 7.24  |
-| pospopcnt_u16_avx512_csa          | 0.52 | 0.74 | 1.83 | 2.64 | 4.06 | **6.43** | **9.41** | **16.28** |
+| pospopcnt_u16_sse_blend_popcnt_unroll8    | **2.09** | 3.16 | 2.35 | 1.88 | 1.67 | 1.56 | 1.5  | 1.44  |
+| pospopcnt_u16_avx512_blend_popcnt_unroll8 | 1.78 | **3.61** | **3.61** | 3.59 | 3.68 | 3.65 | 3.67 | 3.7   |
+| pospopcnt_u16_avx512_adder_forest        | 0.77 | 0.9  | 3.24 | **3.96** | **4.96** | 5.87 | 6.52 | 7.24  |
+| pospopcnt_u16_avx512_harvey_seal          | 0.52 | 0.74 | 1.83 | 2.64 | 4.06 | **6.43** | **9.41** | **16.28** |
 
 Compared to a naive unvectorized solution (`pospopcnt_u16_scalar_naive_nosimd`):
 
 | Algorithm                         | 128  | 256   | 512   | 1024  | 2048  | 4096  | 8192  | 65536  |
 |-----------------------------------|------|-------|-------|-------|-------|-------|-------|--------|
-| pospopcnt_u16_sse_mula_unroll8    | **8.28** | 9.84  | 10.55 | 11    | 11.58 | 11.93 | 12.13 | 12.28  |
-| pospopcnt_u16_avx512_mula_unroll8 | 7.07 | **11.25** | **16.21** | 21    | 25.49 | 27.91 | 29.73 | 31.55  |
-| pospopcnt_u16_avx512_mula3        | 3.05 | 2.82  | 14.53 | **23.13** | **34.37** | 44.91 | 52.78 | 61.68  |
-| pospopcnt_u16_avx512_csa          | 2.07 | 2.3   | 8.21  | 15.41 | 28.17 | **49.14** | **76.11** | **138.71** |
+| pospopcnt_u16_sse_blend_popcnt_unroll8    | **8.28** | 9.84  | 10.55 | 11    | 11.58 | 11.93 | 12.13 | 12.28  |
+| pospopcnt_u16_avx512_blend_popcnt_unroll8 | 7.07 | **11.25** | **16.21** | 21    | 25.49 | 27.91 | 29.73 | 31.55  |
+| pospopcnt_u16_avx512_adder_forest        | 3.05 | 2.82  | 14.53 | **23.13** | **34.37** | 44.91 | 52.78 | 61.68  |
+| pospopcnt_u16_avx512_harvey_seal          | 2.07 | 2.3   | 8.21  | 15.41 | 28.17 | **49.14** | **76.11** | **138.71** |
 
 The host architecture used is a 10 nm Cannon Lake [Core i3-8121U](https://ark.intel.com/content/www/us/en/ark/products/136863/intel-core-i3-8121u-processor-4m-cache-up-to-3-20-ghz.html) with gcc (GCC) 7.3.1 20180303 (Red Hat 7.3.1-5).
 

appveyor.yml

@@ -14,11 +14,9 @@ platform:
   - x64
 
 build_script:
-  - cmd: if [%BUILD_ENVIRONMENT%]==[cygwin] (
-      C:\cygwin\bin\bash -l -c "make;" )
-  - cmd: if [%BUILD_ENVIRONMENT%]==[cygwin64] (
-      C:\cygwin64\bin\bash -l -c "make;" )
-  - cmd: if [%BUILD_ENVIRONMENT%]==[mingw] (
-      C:\MinGW\msys\1.0\bin\bash -l -c "make;" )
-  - cmd: if [%BUILD_ENVIRONMENT%]==[mingw-w64] (
-      C:\msys64\usr\bin\bash -l -c "make;" )
+  - if "%platform%" == "x86" cmake -G "Visual Studio 15 2017" .
+  - if "%platform%" == "x64" cmake -G "Visual Studio 15 2017 Win64" .
+  - cmake --build . --config Release
+
+test_script:
+  - ctest -C Release

main.cpp → benchmark.cpp

@@ -4,6 +4,12 @@
 #include <cassert>//assert
 #include <cstring>//memset
 
+#ifdef _MSC_VER
+# include <intrin.h>
+#else
+# include <x86intrin.h>
+#endif
+
 #include "pospopcnt.h"
 
 inline void* aligned_malloc(size_t size, size_t align) {
@@ -43,8 +49,12 @@ struct bench_unit {
 
 uint64_t get_cpu_cycles() {
     uint64_t result;
+#ifndef _MSC_VER
     __asm__ volatile(".byte 15;.byte 49;shlq $32,%%rdx;orq %%rdx,%%rax":"=a"
                      (result)::"%rdx");
+#else
+    result = __rdtsc();
+#endif
     return result;
 };
 
@@ -104,6 +114,7 @@ int pospopcnt_u16_wrapper(pospopcnt_u16_method_type f, int id, int iterations,
     std::vector<uint64_t> clocks;
     std::vector<uint32_t> times;
 
+#ifndef _MSC_VER
 // Intel guide:
 // @see: https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/ia-32-ia-64-benchmark-code-execution-paper.pdf
 asm   volatile ("CPUID\n\t"
@@ -122,6 +133,7 @@ asm   volatile("RDTSCP\n\t"
                "mov %%edx, %0\n\t"
                "mov %%eax, %1\n\t"
                "CPUID\n\t": "=r" (cycles_high1), "=r" (cycles_low1):: "%rax", "%rbx", "%rcx", "%rdx");
+#endif
 
     for (int i = 0; i < iterations; ++i) {
         memset(counters, 0, sizeof(uint32_t)*16);
@@ -139,19 +151,21 @@ asm   volatile("RDTSCP\n\t"
     /*at this stage we exclusively own the CPU*/ 
 #endif
 
+#ifndef _MSC_VER 
     asm   volatile ("CPUID\n\t"
                     "RDTSC\n\t"
                     "mov %%edx, %0\n\t"
                     "mov %%eax, %1\n\t": "=r" (cycles_high), "=r" (cycles_low):: "%rax", "%rbx", "%rcx", "%rdx");
-
+#endif
     // Call argument subroutine pointer.
     (*f)(data, n, counters);
 
+#ifndef _MSC_VER 
     asm   volatile("RDTSCP\n\t"
                    "mov %%edx, %0\n\t"
                    "mov %%eax, %1\n\t"
                    "CPUID\n\t": "=r" (cycles_high1), "=r" (cycles_low1):: "%rax", "%rbx", "%rcx", "%rdx");
-
+#endif
 #ifdef __linux__ 
         // raw_local_irq_restore(flags);/*we enable hard interrupts on our CPU*/
         // preempt_enable();/*we enable preemption*/

benchmark/linux/instrumented_benchmark.cpp

@@ -29,36 +29,36 @@ pospopcnt_u16_method_type pospopcnt_u16_methods[] = {
     pospopcnt_u16_scalar_umul128,
     pospopcnt_u16_scalar_umul128_unroll2,
     pospopcnt_u16_sse_single,
-    pospopcnt_u16_sse_mula,
-    pospopcnt_u16_sse_mula_unroll4,
-    pospopcnt_u16_sse_mula_unroll8,
-    pospopcnt_u16_sse_mula_unroll16,
+    pospopcnt_u16_sse_blend_popcnt,
+    pospopcnt_u16_sse_blend_popcnt_unroll4,
+    pospopcnt_u16_sse_blend_popcnt_unroll8,
+    pospopcnt_u16_sse_blend_popcnt_unroll16,
     pospopcnt_u16_sse_sad,
-    pospopcnt_u16_sse_csa,
+    pospopcnt_u16_sse_harvey_seal,
     pospopcnt_u16_avx2_popcnt,
     pospopcnt_u16_avx2,
     pospopcnt_u16_avx2_naive_counter,
     pospopcnt_u16_avx2_single,
     pospopcnt_u16_avx2_lemire,
     pospopcnt_u16_avx2_lemire2,
-    pospopcnt_u16_avx2_mula,
-    pospopcnt_u16_avx2_mula_unroll4,
-    pospopcnt_u16_avx2_mula_unroll8,
-    pospopcnt_u16_avx2_mula_unroll16,
-    pospopcnt_u16_avx2_mula3,
-    pospopcnt_u16_avx2_csa,
+    pospopcnt_u16_avx2_blend_popcnt,
+    pospopcnt_u16_avx2_blend_popcnt_unroll4,
+    pospopcnt_u16_avx2_blend_popcnt_unroll8,
+    pospopcnt_u16_avx2_blend_popcnt_unroll16,
+    pospopcnt_u16_avx2_adder_forest,
+    pospopcnt_u16_avx2_harvey_seal,
     pospopcnt_u16_avx512,
     pospopcnt_u16_avx512bw_popcnt32_mask,
     pospopcnt_u16_avx512bw_popcnt64_mask,
     pospopcnt_u16_avx512_masked_ops,
     pospopcnt_u16_avx512_popcnt,
-    pospopcnt_u16_avx512bw_mula,
-    pospopcnt_u16_avx512bw_mula_unroll4,
-    pospopcnt_u16_avx512bw_mula_unroll8,
+    pospopcnt_u16_avx512bw_blend_popcnt,
+    pospopcnt_u16_avx512bw_blend_popcnt_unroll4,
+    pospopcnt_u16_avx512bw_blend_popcnt_unroll8,
     pospopcnt_u16_avx512_mula2,
-    pospopcnt_u16_avx512bw_mula3,
-    pospopcnt_u16_avx512bw_csa,
-    pospopcnt_u16_avx512vbmi_csa};
+    pospopcnt_u16_avx512bw_adder_forest,
+    pospopcnt_u16_avx512bw_harvey_seal,
+    pospopcnt_u16_avx512vbmi_harvey_seal};
 
 void print16(uint32_t *flags) {
     for (int k = 0; k < 16; k++)
@@ -188,6 +188,7 @@ bool benchmark(uint32_t n, uint32_t iterations, pospopcnt_u16_method_type fn, bo
     return isok;
 }
 
+#if POSPOPCNT_SIMD_VERSION >= 5
 void measurepopcnt(uint32_t n, uint32_t iterations, bool verbose) {
     std::vector<int> evts;
     std::vector<uint16_t> vdata(n);
@@ -252,6 +253,7 @@ void measurepopcnt(uint32_t n, uint32_t iterations, bool verbose) {
         printf("cycles per 16-bit word:  %4.3f; ref cycles per 16-bit word: %4.3f \n", double(mins[0]) / n, double(mins[5]) / n);
     }
 }
+#endif
 
 static void print_usage(char *command) {
     printf(" Try %s -n 100000 -i 15 -v \n", command);
@@ -315,8 +317,10 @@ int main(int argc, char **argv) {
       printf("array size: %.3f MB\n", array_in_bytes / (1024 * 1024.));
     }
 
+#if POSPOPCNT_SIMD_VERSION >= 5
     measurepopcnt(n, iterations, verbose);
-
+#endif
+
     for (size_t k = 0; k < PPOPCNT_NUMBER_METHODS; k++) {
         printf("%-40s\t", pospopcnt_u16_method_names[k]);
         fflush(NULL);

example.c

@@ -1,5 +1,6 @@
 #include <stdio.h>
 #include <stdlib.h>
+#include <string.h> // memcpy
 
 #include "pospopcnt.h"