/*
* This file is part of the GROMACS molecular simulation package.
*
- * Copyright (c) 2017,2018, by the GROMACS development team, led by
+ * Copyright (c) 2017,2018,2019, by the GROMACS development team, led by
* Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
* and including many others, as listed in the AUTHORS file in the
* top-level source directory and at http://www.gromacs.org.
*/
#ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
-#include "gmxpre.h"
+# include "gmxpre.h"
#endif
#include "identifyavx512fmaunits.h"
#ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
-#include "config.h"
+# include "config.h"
#endif
#include <cstdint>
#include <mutex>
#ifndef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
-#include "gromacs/hardware/cpuinfo.h"
+# include "gromacs/hardware/cpuinfo.h"
#endif
namespace gmx
* execute 12 FMA and 12 shuffle instructions.
* \return Number of cycles used for the loop.
*/
-uint64_t
-timeFmaAndShuffleLoop(uint64_t loopCount)
+uint64_t timeFmaAndShuffleLoop(uint64_t loopCount)
{
uint64_t cycles;
// Unfortunately we need to resort to inline ASM since we are
// making a choice based on timing, and without efficient optimization
// (e.g. when doing debugging) the usual intrinsics are often implemented
// as independent load/store operations, which completely screws up timing.
- __asm__ __volatile__("\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
- "\tvmovaps %%zmm0, %%zmm1\n"
- "\tvmovaps %%zmm0, %%zmm2\n"
- "\tvmovaps %%zmm0, %%zmm3\n"
- "\tvmovaps %%zmm0, %%zmm4\n"
- "\tvmovaps %%zmm0, %%zmm5\n"
- "\tvmovaps %%zmm0, %%zmm6\n"
- "\tvmovaps %%zmm0, %%zmm7\n"
- "\tvmovaps %%zmm0, %%zmm8\n"
- "\tvmovaps %%zmm0, %%zmm9\n"
- "\tvmovaps %%zmm0, %%zmm10\n"
- "\tvmovaps %%zmm0, %%zmm11\n"
- "\tvpxord %%zmm12, %%zmm12, %%zmm12\n"
- "\tvmovaps %%zmm12, %%zmm13\n"
- "\tvmovaps %%zmm12, %%zmm14\n"
- "\tvmovaps %%zmm12, %%zmm15\n"
- "\tvmovaps %%zmm12, %%zmm16\n"
- "\tvmovaps %%zmm12, %%zmm17\n"
- "\tvmovaps %%zmm12, %%zmm18\n"
- "\tvmovaps %%zmm12, %%zmm19\n"
- "\tvmovaps %%zmm12, %%zmm20\n"
- "\tvmovaps %%zmm12, %%zmm21\n"
- "\tvmovaps %%zmm12, %%zmm22\n"
- "\tvmovaps %%zmm12, %%zmm23\n"
- "\tvmovaps %%zmm12, %%zmm30\n"
- "\trdtscp\n"
- "\tsalq $32, %%rdx\n"
- "\tmovl %%eax, %%eax\n"
- "\tmovq %%rdx, %%rbx\n"
- "\torq %%rax, %%rbx\n"
- "\tmovq %1, %%rdx\n"
- "1:\n"
- "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
- "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
- "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
- "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
- "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
- "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
- "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
- "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
- "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
- "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
- "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
- "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm12\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm13\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm14\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm15\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm16\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm17\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm18\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm19\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm20\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm21\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm22\n"
- "\tvpermd %%zmm30, %%zmm30, %%zmm23\n"
- "\tdec %%rdx\n"
- "\tjg 1b\n"
- "\trdtscp\n"
- "\tsalq $32, %%rdx\n"
- "\tmovl %%eax, %%eax\n"
- "\torq %%rax, %%rdx\n"
- "\tsubq %%rbx, %%rdx\n"
- "\tmovq %%rdx, %0\n"
- : "=r" (cycles) : "r" (loopCount)
- : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3",
- "zmm4", "zmm5", "zmm6", "zmm7", "zmm8", "zmm9", "zmm10",
- "zmm11", "zmm12", "zmm13", "zmm14", "zmm15", "zmm16", "zmm17",
- "zmm18", "zmm19", "zmm20", "zmm21", "zmm22", "zmm23", "zmm30");
+ __asm__ __volatile__(
+ "\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
+ "\tvmovaps %%zmm0, %%zmm1\n"
+ "\tvmovaps %%zmm0, %%zmm2\n"
+ "\tvmovaps %%zmm0, %%zmm3\n"
+ "\tvmovaps %%zmm0, %%zmm4\n"
+ "\tvmovaps %%zmm0, %%zmm5\n"
+ "\tvmovaps %%zmm0, %%zmm6\n"
+ "\tvmovaps %%zmm0, %%zmm7\n"
+ "\tvmovaps %%zmm0, %%zmm8\n"
+ "\tvmovaps %%zmm0, %%zmm9\n"
+ "\tvmovaps %%zmm0, %%zmm10\n"
+ "\tvmovaps %%zmm0, %%zmm11\n"
+ "\tvpxord %%zmm12, %%zmm12, %%zmm12\n"
+ "\tvmovaps %%zmm12, %%zmm13\n"
+ "\tvmovaps %%zmm12, %%zmm14\n"
+ "\tvmovaps %%zmm12, %%zmm15\n"
+ "\tvmovaps %%zmm12, %%zmm16\n"
+ "\tvmovaps %%zmm12, %%zmm17\n"
+ "\tvmovaps %%zmm12, %%zmm18\n"
+ "\tvmovaps %%zmm12, %%zmm19\n"
+ "\tvmovaps %%zmm12, %%zmm20\n"
+ "\tvmovaps %%zmm12, %%zmm21\n"
+ "\tvmovaps %%zmm12, %%zmm22\n"
+ "\tvmovaps %%zmm12, %%zmm23\n"
+ "\tvmovaps %%zmm12, %%zmm30\n"
+ "\trdtscp\n"
+ "\tsalq $32, %%rdx\n"
+ "\tmovl %%eax, %%eax\n"
+ "\tmovq %%rdx, %%rbx\n"
+ "\torq %%rax, %%rbx\n"
+ "\tmovq %1, %%rdx\n"
+ "1:\n"
+ "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
+ "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
+ "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
+ "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
+ "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
+ "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
+ "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
+ "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
+ "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
+ "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
+ "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
+ "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm12\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm13\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm14\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm15\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm16\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm17\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm18\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm19\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm20\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm21\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm22\n"
+ "\tvpermd %%zmm30, %%zmm30, %%zmm23\n"
+ "\tdec %%rdx\n"
+ "\tjg 1b\n"
+ "\trdtscp\n"
+ "\tsalq $32, %%rdx\n"
+ "\tmovl %%eax, %%eax\n"
+ "\torq %%rax, %%rdx\n"
+ "\tsubq %%rbx, %%rdx\n"
+ "\tmovq %%rdx, %0\n"
+ : "=r"(cycles)
+ : "r"(loopCount)
+ : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6",
+ "zmm7", "zmm8", "zmm9", "zmm10", "zmm11", "zmm12", "zmm13", "zmm14", "zmm15", "zmm16",
+ "zmm17", "zmm18", "zmm19", "zmm20", "zmm21", "zmm22", "zmm23", "zmm30");
return cycles;
}
* execute 12 FMA instructions.
* \return Number of cycles used for the loop.
*/
-uint64_t
-timeFmaOnlyLoop(uint64_t loopCount)
+uint64_t timeFmaOnlyLoop(uint64_t loopCount)
{
uint64_t cycles;
// Unfortunately we need to resort to inline ASM since we are
// making a choice based on timing, and without efficient optimization
// (e.g. when doing debugging) the usual intrinsics are often implemented
// as independent load/store operations, which completely screws up timing.
- __asm__ __volatile__("\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
- "\tvmovaps %%zmm0, %%zmm1\n"
- "\tvmovaps %%zmm0, %%zmm2\n"
- "\tvmovaps %%zmm0, %%zmm3\n"
- "\tvmovaps %%zmm0, %%zmm4\n"
- "\tvmovaps %%zmm0, %%zmm5\n"
- "\tvmovaps %%zmm0, %%zmm6\n"
- "\tvmovaps %%zmm0, %%zmm7\n"
- "\tvmovaps %%zmm0, %%zmm8\n"
- "\tvmovaps %%zmm0, %%zmm9\n"
- "\tvmovaps %%zmm0, %%zmm10\n"
- "\tvmovaps %%zmm0, %%zmm11\n"
- "\trdtscp\n"
- "\tsalq $32, %%rdx\n"
- "\tmovl %%eax, %%eax\n"
- "\tmovq %%rdx, %%rbx\n"
- "\torq %%rax, %%rbx\n"
- "\tmovq %1, %%rdx\n"
- "1:\n"
- "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
- "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
- "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
- "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
- "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
- "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
- "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
- "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
- "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
- "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
- "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
- "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
- "\tdec %%rdx\n"
- "\tjg 1b\n"
- "\trdtscp\n"
- "\tsalq $32, %%rdx\n"
- "\tmovl %%eax, %%eax\n"
- "\torq %%rax, %%rdx\n"
- "\tsubq %%rbx, %%rdx\n"
- "\tmovq %%rdx, %0\n"
- : "=r" (cycles) : "r" (loopCount)
- : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3",
- "zmm4", "zmm5", "zmm6", "zmm7", "zmm8", "zmm9", "zmm10", "zmm11");
+ __asm__ __volatile__(
+ "\tvpxord %%zmm0, %%zmm0, %%zmm0\n"
+ "\tvmovaps %%zmm0, %%zmm1\n"
+ "\tvmovaps %%zmm0, %%zmm2\n"
+ "\tvmovaps %%zmm0, %%zmm3\n"
+ "\tvmovaps %%zmm0, %%zmm4\n"
+ "\tvmovaps %%zmm0, %%zmm5\n"
+ "\tvmovaps %%zmm0, %%zmm6\n"
+ "\tvmovaps %%zmm0, %%zmm7\n"
+ "\tvmovaps %%zmm0, %%zmm8\n"
+ "\tvmovaps %%zmm0, %%zmm9\n"
+ "\tvmovaps %%zmm0, %%zmm10\n"
+ "\tvmovaps %%zmm0, %%zmm11\n"
+ "\trdtscp\n"
+ "\tsalq $32, %%rdx\n"
+ "\tmovl %%eax, %%eax\n"
+ "\tmovq %%rdx, %%rbx\n"
+ "\torq %%rax, %%rbx\n"
+ "\tmovq %1, %%rdx\n"
+ "1:\n"
+ "\tvfmadd231pd %%zmm0, %%zmm0, %%zmm0\n"
+ "\tvfmadd231pd %%zmm1, %%zmm1, %%zmm1\n"
+ "\tvfmadd231pd %%zmm2, %%zmm2, %%zmm2\n"
+ "\tvfmadd231pd %%zmm3, %%zmm3, %%zmm3\n"
+ "\tvfmadd231pd %%zmm4, %%zmm4, %%zmm4\n"
+ "\tvfmadd231pd %%zmm5, %%zmm5, %%zmm5\n"
+ "\tvfmadd231pd %%zmm6, %%zmm6, %%zmm6\n"
+ "\tvfmadd231pd %%zmm7, %%zmm7, %%zmm7\n"
+ "\tvfmadd231pd %%zmm8, %%zmm8, %%zmm8\n"
+ "\tvfmadd231pd %%zmm9, %%zmm9, %%zmm9\n"
+ "\tvfmadd231pd %%zmm10, %%zmm10, %%zmm10\n"
+ "\tvfmadd231pd %%zmm11, %%zmm11, %%zmm11\n"
+ "\tdec %%rdx\n"
+ "\tjg 1b\n"
+ "\trdtscp\n"
+ "\tsalq $32, %%rdx\n"
+ "\tmovl %%eax, %%eax\n"
+ "\torq %%rax, %%rdx\n"
+ "\tsubq %%rbx, %%rdx\n"
+ "\tmovq %%rdx, %0\n"
+ : "=r"(cycles)
+ : "r"(loopCount)
+ : "rax", "rbx", "rcx", "rdx", "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6",
+ "zmm7", "zmm8", "zmm9", "zmm10", "zmm11");
return cycles;
}
-bool
-checkDualAvx512FmaUnits()
+bool checkDualAvx512FmaUnits()
{
- uint64_t timeFmaAndShuf = static_cast<uint64_t>(1e9); // Large value
+ uint64_t timeFmaAndShuf = static_cast<uint64_t>(1e9); // Large value
// Make sure the CPU is in AVX512 mode by executing a fairly long loop.
// Use the return value to make sure it is not optimized away. Later invocations
// use fewer iterations, so they should always be faster.
- uint64_t timeFmaOnly = timeFmaOnlyLoop(100000);
+ uint64_t timeFmaOnly = timeFmaOnlyLoop(100000);
// Execute the loops three times
for (int i = 0; i < 3; i++)
{
- timeFmaAndShuf = std::min(timeFmaAndShuf, timeFmaAndShuffleLoop(1000) );
- timeFmaOnly = std::min(timeFmaOnly, timeFmaOnlyLoop(1000) );
+ timeFmaAndShuf = std::min(timeFmaAndShuf, timeFmaAndShuffleLoop(1000));
+ timeFmaOnly = std::min(timeFmaOnly, timeFmaOnlyLoop(1000));
}
return timeFmaAndShuf > 1.5 * timeFmaOnly;
}
-#endif // GMX_X86_GCC_INLINE_ASM && SIMD_AVX_512_CXX_SUPPORTED
+#endif // GMX_X86_GCC_INLINE_ASM && SIMD_AVX_512_CXX_SUPPORTED
/*! \brief Mutex to guard the execution of the timing test
*/
std::mutex initMutex;
-} // namespace
+} // namespace
-int
-identifyAvx512FmaUnits()
+int identifyAvx512FmaUnits()
{
static bool initialized = false;
static int result = 0;
if (!initialized)
{
- std::lock_guard<std::mutex> lock(initMutex);
+ std::lock_guard<std::mutex> lock(initMutex);
if (!initialized)
{
} // namespace gmx
#ifdef GMX_IDENTIFY_AVX512_FMA_UNITS_STANDALONE
-int
-main()
+int main()
{
printf("%d\n", gmx::identifyAvx512FmaUnits());
return 0;
biod.pnpi.spb.ru / alexxy / gromacs.git / blobdiff