NN-512
Index
Files
Config Prefix=Example24 Platform=AVX512Float32 L1DataCachePerThread=32KiB L2CachePerThreadExL1=960KiB L3CachePerThreadExL1L2=1408KiB
Input ToTensor=in Channels=1041 Height=10 Width=34
Pooling FromTensor=in ToTensor=out Kind=AvgGlobal PaddingH=0 PaddingW=0
Output FromTensor=out
#pragma once
// NN-512 (https://NN-512.com)
//
// Copyright (C) 2019 [
// 37ef ced3 3727 60b4
// 3c29 f9c6 dc30 d518
// f4f3 4106 6964 cab4
// a06f c1a3 83fd 090e
// ]
//
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. 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.
//
// 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.
#include <pthread.h>
#include <stddef.h>
#ifdef __cplusplus
extern "C" { /**/
#endif
// All weights, biases, and other trained parameters are passed into
// the initialization code through the Params struct that is declared
// just below this comment. The corresponding struct definition can be
// found near the end of this header file.
//
// Each field of the Params struct is an array of float that holds a
// parameter tensor in NCHW format with no padding. The struct fields
// are ordered by name, lexically bytewise. If you concatenate all the
// trained parameter tensors to a file in this same format and order
// you can load the struct as follows (error checking omitted here):
//
// size_t size = sizeof(Example24Params);
// Example24Params* to = malloc(size);
// FILE* from = fopen("ParamsFile", "r");
// fread(to, size, 1, from);
// fclose(from);
//
// Be careful to match endianness (and floating point format).
typedef struct Example24Params Example24Params;
// The Net contains weights, biases, and other trained parameters in a
// form that enables efficient inference. It is created from the input
// parameter struct without modifying that struct. The input parameter
// struct is no longer needed once the Net has been created. Threads
// that are used to create the Net are temporary (in particular, those
// threads are not used for inference).
//
// Example24Params* params = malloc(sizeof(Example24Params));
//
// ... Load params (read from a file, perhaps) ...
//
// Example24Net* net; // For example, 4 threads:
// char* err = Example24NetCreate(&net, params, 4);
// free(params);
//
// if (err) { // Nonzero err indicates failure; net is unmodified.
// printf("%s\n", err); // Explain the failure, add a newline.
// free(err); // Free the error string to avoid a memory leak.
// exit(1); // Exit, or propagate the failure some other way.
// }
//
// ... Perform all inference that depends on net ...
//
// Example24NetDestroy(net);
//
// The Net can be shared and reused without restriction because it is
// never modified (not even temporarily) after being created. The Net
// should be destroyed (to free memory) once all dependent inference
// is complete.
typedef struct Example24Net Example24Net;
char* Example24NetCreate(
Example24Net**,
Example24Params*,
ptrdiff_t threads
);
void Example24NetDestroy(Example24Net*);
// An Engine performs inference. It contains inference threads, scratch
// memory, and a pointer to the Net. Any number of Engines can share the
// same Net (and perform inference in parallel) because the Net is never
// modified. For best performance the number of inference threads should
// not exceed the number of CPU cores.
//
// Example24Net* net;
//
// ... Create net ...
//
// Example24Engine* engine; // For example, 4 inference threads:
// char* err = Example24EngineCreate(&engine, net, 4);
//
// if (err) { // Nonzero err means failure; engine is unmodified.
// printf("%s\n", err); // Explain the failure, add a newline.
// free(err); // Free the error string to avoid a memory leak.
//
// ... Destroy net ...
//
// exit(1); // Exit, or propagate the failure some other way.
// }
//
// ... Use the POSIX threads API to adjust engine's threads ...
// ... Use engine to perform inference (dependent on net) ...
//
// Example24EngineDestroy(engine); // Terminate threads, free memory.
//
// ... Destroy net ...
//
// The POSIX threads API can be used to adjust an Engine's threads. If
// an Engine has N threads, those threads are indexed 0, 1, 2, ..., N-1
// and a pthread_t identifier is associated with each index. To set the
// CPU affinity mask for the first inference thread, for example:
//
// pthread_t thread; // The first thread has index 0:
// char* err = Example24EnginePthreadT(engine, 0, &thread);
//
// assert(!err); // Can only fail if the thread index is invalid.
//
// pthread_setaffinity_np(thread, ...); // Details omitted.
//
// The inference function reads floats from (one or more) input tensors
// and writes floats to (one or more) output tensors. All the input and
// output tensors are owned (allocated and freed) by the caller and are
// in CHW format, 32-bit floating point, fully packed (in other words,
// C has the largest pitch, W has the smallest pitch, and there is no
// padding anywhere).
//
// float* inData = malloc(sizeof(float)*1041*10*34);
// float* outData = malloc(sizeof(float)*1041*1*1);
//
// for (...) { // Reuse the input and output tensors.
//
// ... Write the input floats ...
//
// Example24EngineInference( // This function cannot fail.
// engine, // Pass an Engine as the first argument.
// inData, // The tensor arguments are sorted by name.
// outData
// );
//
// ... Read the output floats ...
//
// }
//
// free(inData);
// free(outData);
//
// The tensor parameters of the inference function are ordered by name,
// lexically bytewise. In other words, the function parameters have been
// sorted by name using Go's "<" string comparison operator (a bytewise
// lexical string sort).
typedef struct Example24Engine Example24Engine;
char* Example24EngineCreate(
Example24Engine**,
Example24Net*,
ptrdiff_t threads
);
char* Example24EnginePthreadT(
Example24Engine*,
ptrdiff_t threadIdx,
pthread_t* to
);
void Example24EngineInference(
Example24Engine*,
float* inData,
float* outData
);
void Example24EngineDestroy(Example24Engine*);
// The fields of the following struct have been sorted by name using
// Go's "<" string comparison operator (bytewise lexical string sort).
// Tensor dimensions are NxCxHxW where N is the outermost/slowest and
// W is the innermost/fastest. There is no padding anywhere.
struct Example24Params {
} __attribute__((packed));
#ifdef __cplusplus
/**/ }
#endif
// End of file.
// To build an object file:
// gcc -c -w -std=c99 -pthread -Ofast -mavx512f Example24.c
// NN-512 (https://NN-512.com)
//
// Copyright (C) 2019 [
// 37ef ced3 3727 60b4
// 3c29 f9c6 dc30 d518
// f4f3 4106 6964 cab4
// a06f c1a3 83fd 090e
// ]
//
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
//
// 1. Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// 2. 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.
//
// 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.
#include <errno.h>
#include <stdarg.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <immintrin.h>
#include "Example24.h"
static char* Example24Errmsg1(ptrdiff_t lineNum1, char* format1, ...) {
char* msg1 = malloc(277);
int step1 = sprintf(msg1, "Example24: line %td: ", lineNum1);
va_list ap1;
va_start(ap1, format1);
vsnprintf(msg1+step1, 277-step1, format1, ap1);
va_end(ap1);
return msg1;
}
typedef struct Example24ThreaderTask1 Example24ThreaderTask1;
typedef void (*Example24ThreaderCallee1)(Example24ThreaderTask1*, int64_t*);
typedef struct Example24ThreaderHub1 Example24ThreaderHub1;
typedef struct Example24ThreaderNode1 Example24ThreaderNode1;
typedef struct Example24ThreaderUnwind1 Example24ThreaderUnwind1;
typedef struct Example24ThreaderTeam1 Example24ThreaderTeam1;
struct Example24ThreaderTask1 {
Example24ThreaderCallee1 callee1;
void* any1;
ptrdiff_t nd1;
int64_t hull1[4];
};
struct Example24ThreaderHub1 {
pthread_mutex_t mut1;
pthread_cond_t cond1;
ptrdiff_t pending1;
ptrdiff_t offset1;
long mask1;
long status1[];
};
struct Example24ThreaderNode1 {
pthread_mutex_t mut2;
int64_t np1;
int64_t pt1[4];
Example24ThreaderTask1* task1;
pthread_cond_t cond2;
Example24ThreaderTeam1* team1;
pthread_t thr1;
} __attribute__((aligned(64)));
struct Example24ThreaderUnwind1 {
ptrdiff_t join1;
ptrdiff_t nodeConds1;
ptrdiff_t nodeMuts1;
ptrdiff_t hubCond1;
ptrdiff_t hubMut1;
void* nodes1;
void* hub1;
};
struct Example24ThreaderTeam1 {
ptrdiff_t nt1;
Example24ThreaderHub1* hub2;
Example24ThreaderNode1* nodes2;
Example24ThreaderUnwind1 unwind1;
};
static void Example24ThreaderInc1(
ptrdiff_t nd2,
int64_t*restrict hull2,
int64_t*restrict pt2
) {
for (ptrdiff_t i1 = 0; i1 < nd2; ++i1) {
int64_t elem1 = pt2[i1];
if (++elem1 == hull2[i1]) {
pt2[i1] = 0;
} else {
pt2[i1] = elem1;
break;
}
}
}
static void Example24ThreaderPut1(
ptrdiff_t nd3,
int64_t*restrict hull3,
int64_t*restrict pt3,
int64_t val1
) {
ptrdiff_t i2 = 0;
for (; i2 < nd3 && val1; ) {
int64_t wrap1 = hull3[i2];
int64_t carry1 = val1/wrap1;
pt3[i2++] = val1-carry1*wrap1;
val1 = carry1;
}
for (; i2 < nd3; pt3[i2++] = 0);
}
static void Example24ThreaderAdd1(
ptrdiff_t nd4,
int64_t*restrict hull4,
int64_t*restrict pt4,
int64_t*restrict plus1,
int64_t carry2
) {
for (ptrdiff_t i3 = 0; i3 < nd4; ++i3) {
int64_t wrap2 = hull4[i3];
int64_t sum1 = pt4[i3]+plus1[i3]+carry2;
if (sum1 < wrap2) {
pt4[i3] = sum1;
carry2 = 0;
} else {
pt4[i3] = sum1-wrap2;
carry2 = 1;
}
}
}
static void* Example24ThreaderMain1(void* arg1) {
Example24ThreaderNode1* node1 = arg1;
Example24ThreaderTeam1* team2 = node1->team1;
ptrdiff_t nt2 = team2->nt1;
Example24ThreaderHub1* hub3 = team2->hub2;
Example24ThreaderNode1* nodes3 = team2->nodes2;
size_t role1 = node1-nodes3;
for (; __builtin_expect(pthread_mutex_lock(&node1->mut2), 0); );
for (; ; ) {
Example24ThreaderTask1* task2 = node1->task1;
if (!task2) {
for (; __builtin_expect(pthread_cond_wait(&node1->cond2, &node1->mut2), 0); );
continue;
}
int64_t np2 = node1->np1;
if (np2 < 0) {
for (; __builtin_expect(pthread_mutex_unlock(&node1->mut2), 0); );
return 0;
}
node1->task1 = 0;
Example24ThreaderCallee1 callee2 = task2->callee1;
ptrdiff_t nd5 = task2->nd1;
int64_t pt5[4];
for (; np2; np2 = node1->np1) {
memcpy(pt5, node1->pt1, sizeof(pt5));
node1->np1 = np2-1;
Example24ThreaderInc1(nd5, task2->hull1, node1->pt1);
for (; __builtin_expect(pthread_mutex_unlock(&node1->mut2), 0); );
callee2(task2, pt5);
for (; __builtin_expect(pthread_mutex_lock(&node1->mut2), 0); );
}
for (; __builtin_expect(pthread_mutex_unlock(&node1->mut2), 0); );
for (; __builtin_expect(pthread_mutex_lock(&hub3->mut1), 0); );
hub3->status1[role1/(sizeof(long)*8)] &= ~((long)1<<role1%(sizeof(long)*8));
ptrdiff_t offset2 = hub3->offset1;
long mask2 = hub3->mask1;
ptrdiff_t wrapped1 = 0;
for (; ; ) {
long hand1 = hub3->status1[offset2]&mask2;
if (!hand1) {
++offset2;
mask2 = -1;
continue;
}
ptrdiff_t target1 = offset2*(sizeof(long)*8)+__builtin_ctzl(hand1);
if (target1 == nt2) {
if (wrapped1) break;
offset2 = 0;
mask2 = -1;
wrapped1 = 1;
continue;
}
hand1 &= -hand1;
hub3->offset1 = offset2;
hub3->mask1 = mask2-hand1;
for (; __builtin_expect(pthread_mutex_unlock(&hub3->mut1), 0); );
Example24ThreaderNode1* node2 = nodes3+target1;
for (; __builtin_expect(pthread_mutex_lock(&node2->mut2), 0); );
for (np2 = node2->np1; np2; np2 = node2->np1) {
memcpy(pt5, node2->pt1, sizeof(pt5));
node2->np1 = np2-1;
Example24ThreaderInc1(nd5, task2->hull1, node2->pt1);
for (; __builtin_expect(pthread_mutex_unlock(&node2->mut2), 0); );
callee2(task2, pt5);
for (; __builtin_expect(pthread_mutex_lock(&node2->mut2), 0); );
}
for (; __builtin_expect(pthread_mutex_unlock(&node2->mut2), 0); );
for (; __builtin_expect(pthread_mutex_lock(&hub3->mut1), 0); );
hub3->status1[offset2] &= ~hand1;
offset2 = hub3->offset1;
mask2 = hub3->mask1;
wrapped1 = 0;
}
ptrdiff_t pending2 = --hub3->pending1;
for (; __builtin_expect(pthread_mutex_unlock(&hub3->mut1), 0); );
if (!pending2) for (; __builtin_expect(pthread_cond_signal(&hub3->cond1), 0); );
for (; __builtin_expect(pthread_mutex_lock(&node1->mut2), 0); );
}
}
static void Example24ThreaderDestroy1(Example24ThreaderTeam1* team3) {
if (!team3) return;
Example24ThreaderNode1* nodes4 = team3->nodes2;
Example24ThreaderNode1* stop1 = nodes4+team3->unwind1.join1;
for (Example24ThreaderNode1* node3 = nodes4; node3 != stop1; ++node3) {
for (; __builtin_expect(pthread_mutex_lock(&node3->mut2), 0); );
node3->np1 = -1;
node3->task1 = (Example24ThreaderTask1*)1;
for (; __builtin_expect(pthread_mutex_unlock(&node3->mut2), 0); );
for (; __builtin_expect(pthread_cond_signal(&node3->cond2), 0); );
}
for (Example24ThreaderNode1* node3 = nodes4; node3 != stop1; ++node3) {
for (; __builtin_expect(pthread_join(node3->thr1, 0), 0); );
}
stop1 = nodes4+team3->unwind1.nodeConds1;
for (Example24ThreaderNode1* node3 = nodes4; node3 != stop1; ++node3) {
for (; __builtin_expect(pthread_cond_destroy(&node3->cond2), 0); );
}
stop1 = nodes4+team3->unwind1.nodeMuts1;
for (Example24ThreaderNode1* node3 = nodes4; node3 != stop1; ++node3) {
for (; __builtin_expect(pthread_mutex_destroy(&node3->mut2), 0); );
}
Example24ThreaderHub1* hub4 = team3->hub2;
if (team3->unwind1.hubCond1) {
for (; __builtin_expect(pthread_cond_destroy(&hub4->cond1), 0); );
}
if (team3->unwind1.hubMut1) {
for (; __builtin_expect(pthread_mutex_destroy(&hub4->mut1), 0); );
}
free(team3->unwind1.nodes1);
free(team3->unwind1.hub1);
free(team3);
}
static char* Example24ThreaderCreate1Up4(Example24ThreaderTeam1* team8, ptrdiff_t nt7) {
Example24ThreaderNode1* nodes5 = team8->nodes2;
for (Example24ThreaderNode1* node4 = nodes5; node4 != nodes5+nt7; ++node4) {
int err2 = pthread_mutex_init(&node4->mut2, 0);
if (__builtin_expect(err2, 0)) {
char* msg2 = Example24Errmsg1(__LINE__, "errno %d", err2);
team8->unwind1.nodeMuts1 = node4-nodes5;
team8->unwind1.nodeConds1 = node4-nodes5;
team8->unwind1.join1 = node4-nodes5;
return msg2;
}
node4->task1 = 0;
int err3 = pthread_cond_init(&node4->cond2, 0);
if (__builtin_expect(err3, 0)) {
char* msg3 = Example24Errmsg1(__LINE__, "errno %d", err3);
team8->unwind1.nodeMuts1 = node4-nodes5+1;
team8->unwind1.nodeConds1 = node4-nodes5;
team8->unwind1.join1 = node4-nodes5;
return msg3;
}
node4->team1 = team8;
int err4 = pthread_create(&node4->thr1, 0, Example24ThreaderMain1, node4);
if (__builtin_expect(err4, 0)) {
char* msg4 = Example24Errmsg1(__LINE__, "errno %d", err4);
team8->unwind1.nodeMuts1 = node4-nodes5+1;
team8->unwind1.nodeConds1 = node4-nodes5+1;
team8->unwind1.join1 = node4-nodes5;
return msg4;
}
}
team8->unwind1.nodeMuts1 = nt7;
team8->unwind1.nodeConds1 = nt7;
team8->unwind1.join1 = nt7;
return 0;
}
static char* Example24ThreaderCreate1Up3(Example24ThreaderTeam1* team7, ptrdiff_t nt6) {
Example24ThreaderHub1* hub5 = team7->hub2;
int err5 = pthread_mutex_init(&hub5->mut1, 0);
if (__builtin_expect(err5, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", err5);
}
team7->unwind1.hubMut1 = 1;
int err6 = pthread_cond_init(&hub5->cond1, 0);
if (__builtin_expect(err6, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", err6);
}
team7->unwind1.hubCond1 = 1;
return Example24ThreaderCreate1Up4(team7, nt6);
}
static char* Example24ThreaderCreate1Up2(Example24ThreaderTeam1* team6, ptrdiff_t nt5) {
size_t size2 = nt5*sizeof(Example24ThreaderNode1);
if (__builtin_expect(size2/sizeof(Example24ThreaderNode1) != (size_t)nt5, 0)) {
return Example24Errmsg1(__LINE__, "too many threads");
}
void* addr3 = malloc(size2+63);
if (__builtin_expect(!addr3, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", errno);
}
team6->unwind1.nodes1 = addr3;
team6->nodes2 = (void*)(((size_t)addr3+63)&-64);
return Example24ThreaderCreate1Up3(team6, nt5);
}
static char* Example24ThreaderCreate1Up1(Example24ThreaderTeam1* team5, ptrdiff_t nt4) {
team5->nt1 = nt4;
size_t size1 = sizeof(Example24ThreaderHub1);
size1 += sizeof(long)*((size_t)nt4/(sizeof(long)*8)+1);
size1 = (size1+63)&-64;
void* addr2 = malloc(size1+63);
if (__builtin_expect(!addr2, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", errno);
}
team5->unwind1.hub1 = addr2;
team5->hub2 = (void*)(((size_t)addr2+63)&-64);
return Example24ThreaderCreate1Up2(team5, nt4);
}
static char* Example24ThreaderCreate1(Example24ThreaderTeam1** team4, ptrdiff_t nt3) {
if (__builtin_expect(nt3 < 1, 0)) {
return Example24Errmsg1(__LINE__, "too few threads");
}
void* addr1 = calloc(1, sizeof(Example24ThreaderTeam1));
if (__builtin_expect(!addr1, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", errno);
}
char* err1 = Example24ThreaderCreate1Up1(addr1, nt3);
if (__builtin_expect(!!err1, 0)) {
Example24ThreaderDestroy1(addr1);
} else {
*team4 = addr1;
}
return err1;
}
static char* Example24ThreaderPthreadT1(
pthread_t* thr2,
Example24ThreaderTeam1* team9,
ptrdiff_t idx1
) {
if (__builtin_expect(idx1 < 0 || idx1 >= team9->nt1, 0)) {
return Example24Errmsg1(__LINE__, "bad thread idx");
}
*thr2 = team9->nodes2[idx1].thr1;
return 0;
}
static void Example24ThreaderDo1(Example24ThreaderTeam1* team10, Example24ThreaderTask1* task3) {
ptrdiff_t nd6 = task3->nd1;
if (nd6 < 1) return;
int64_t tot1 = task3->hull1[0];
for (ptrdiff_t i4 = 1; i4 < nd6; tot1 *= task3->hull1[i4++]);
ptrdiff_t nt8 = team10->nt1;
int64_t each1 = tot1/nt8;
ptrdiff_t more1 = tot1%nt8;
int64_t plus2[4];
Example24ThreaderPut1(nd6, task3->hull1, plus2, each1);
int64_t pt6[4] = {0};
Example24ThreaderHub1* hub6 = team10->hub2;
for (; __builtin_expect(pthread_mutex_lock(&hub6->mut1), 0); );
Example24ThreaderNode1* node5 = team10->nodes2;
for (ptrdiff_t i4 = 0; ; ++node5) {
for (; __builtin_expect(pthread_mutex_lock(&node5->mut2), 0); );
int64_t carry3 = i4 < more1;
node5->np1 = each1+carry3;
memcpy(node5->pt1, pt6, sizeof(pt6));
node5->task1 = task3;
for (; __builtin_expect(pthread_mutex_unlock(&node5->mut2), 0); );
for (; __builtin_expect(pthread_cond_signal(&node5->cond2), 0); );
if (++i4 == nt8) break;
Example24ThreaderAdd1(nd6, task3->hull1, pt6, plus2, carry3);
}
hub6->offset1 = 0;
hub6->mask1 = -1;
for (ptrdiff_t i4 = (size_t)nt8/(sizeof(long)*8); i4 >= 0; ) {
hub6->status1[i4--] = -1;
}
for (hub6->pending1 = nt8; hub6->pending1; ) {
for (; __builtin_expect(pthread_cond_wait(&hub6->cond1, &hub6->mut1), 0); );
}
for (; __builtin_expect(pthread_mutex_unlock(&hub6->mut1), 0); );
}
static __m512 Example24Exp1(__m512 x1) {
x1 = _mm512_max_ps(x1, _mm512_set1_ps(-8.733654e+01f));
x1 = _mm512_min_ps(x1, _mm512_set1_ps(8.872284e+01f));
__m512 t1 = _mm512_mul_ps(x1, _mm512_set1_ps(1.442695e+00f));
__m512 r1 = _mm512_roundscale_ps(t1, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
__m512 f1 = _mm512_fmadd_ps(r1, _mm512_set1_ps(-6.9314575e-01f), x1);
f1 = _mm512_fmadd_ps(r1, _mm512_set1_ps(-1.4286068e-06f), f1);
__m512 g1 = _mm512_set1_ps(4.194439e-02f);
g1 = _mm512_fmadd_ps(g1, f1, _mm512_set1_ps(1.6800667e-01f));
g1 = _mm512_fmadd_ps(g1, f1, _mm512_set1_ps(4.9999994e-01f));
g1 = _mm512_fmadd_ps(g1, f1, _mm512_set1_ps(9.999569e-01f));
g1 = _mm512_fmadd_ps(g1, f1, _mm512_set1_ps(9.9999964e-01f));
__m512i y1 = _mm512_slli_epi32(_mm512_cvtps_epi32(t1), 23);
return _mm512_castsi512_ps(_mm512_add_epi32(y1, _mm512_castps_si512(g1)));
}
static __m512 Example24Rsqrt1(__m512 x2) {
__m512 y2 = _mm512_rsqrt14_ps(x2);
__m512 z1 = _mm512_mul_ps(x2, y2);
__m512 a1 = _mm512_mul_ps(y2, _mm512_set1_ps(5e-01f));
__m512 b1 = _mm512_fnmadd_ps(y2, z1, _mm512_set1_ps(3e+00f));
return _mm512_mul_ps(a1, b1);
}
static void Example24Glopl1Callee1(Example24ThreaderTask1* task4, int64_t* pt7) {
char** tensors2 = task4->any1;
ptrdiff_t c1 = pt7[0];
char*restrict ptr1 = tensors2[0]+(ptrdiff_t)32640*c1;
char*restrict ptr2 = tensors2[1]+(ptrdiff_t)96*c1;
if (c1 < 43) {
__m512 buf1 = _mm512_setzero_ps();
__mmask16 mask3 = 65535;
for (ptrdiff_t i5 = 0; i5 < 6; ++i5) {
__m512 acc1 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc2 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)64+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc3 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)1360+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc4 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)1424+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc5 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)2720+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc6 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)2784+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc7 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)4080+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
__m512 acc8 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)4144+(ptrdiff_t)5440*i5+(ptrdiff_t)128*0);
for (ptrdiff_t j1 = 1; j1 < 10; ++j1) {
__m512 dat1 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat2 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)64+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat3 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)1360+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat4 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)1424+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat5 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)2720+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat6 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)2784+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat7 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)4080+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
__m512 dat8 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)4144+(ptrdiff_t)5440*i5+(ptrdiff_t)128*j1);
acc1 = _mm512_mask_add_ps(acc1, 65535, acc1, dat1);
acc2 = _mm512_mask_add_ps(acc2, 65535, acc2, dat2);
acc3 = _mm512_mask_add_ps(acc3, 65535, acc3, dat3);
acc4 = _mm512_mask_add_ps(acc4, 65535, acc4, dat4);
acc5 = _mm512_mask_add_ps(acc5, 65535, acc5, dat5);
acc6 = _mm512_mask_add_ps(acc6, 65535, acc6, dat6);
acc7 = _mm512_mask_add_ps(acc7, 65535, acc7, dat7);
acc8 = _mm512_mask_add_ps(acc8, 65535, acc8, dat8);
}
__m512 dat9 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat10 = _mm512_maskz_loadu_ps(15, ptr1+(ptrdiff_t)64+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat11 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)1360+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat12 = _mm512_maskz_loadu_ps(15, ptr1+(ptrdiff_t)1424+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat13 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)2720+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat14 = _mm512_maskz_loadu_ps(15, ptr1+(ptrdiff_t)2784+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat15 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)4080+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
__m512 dat16 = _mm512_maskz_loadu_ps(15, ptr1+(ptrdiff_t)4144+(ptrdiff_t)5440*i5+(ptrdiff_t)128*10);
acc1 = _mm512_mask_add_ps(acc1, 65535, acc1, dat9);
acc2 = _mm512_mask_add_ps(acc2, 15, acc2, dat10);
acc3 = _mm512_mask_add_ps(acc3, 65535, acc3, dat11);
acc4 = _mm512_mask_add_ps(acc4, 15, acc4, dat12);
acc5 = _mm512_mask_add_ps(acc5, 65535, acc5, dat13);
acc6 = _mm512_mask_add_ps(acc6, 15, acc6, dat14);
acc7 = _mm512_mask_add_ps(acc7, 65535, acc7, dat15);
acc8 = _mm512_mask_add_ps(acc8, 15, acc8, dat16);
acc1 = _mm512_mask_add_ps(acc1, 65535, acc1, acc2);
acc3 = _mm512_mask_add_ps(acc3, 65535, acc3, acc4);
acc5 = _mm512_mask_add_ps(acc5, 65535, acc5, acc6);
acc7 = _mm512_mask_add_ps(acc7, 65535, acc7, acc8);
__m512i pm1lo1 = _mm512_set_epi32(18, 2, 16, 0, 18, 2, 16, 0, 18, 2, 16, 0, 18, 2, 16, 0);
__m512i pm1hi1 = _mm512_set_epi32(19, 3, 17, 1, 19, 3, 17, 1, 19, 3, 17, 1, 19, 3, 17, 1);
__m512 hi1 = _mm512_shuffle_f32x4(acc1, acc1, 238);
__m512 hi6 = _mm512_shuffle_f32x4(acc3, acc3, 238);
__m512 hi3 = _mm512_shuffle_f32x4(acc5, acc5, 238);
__m512 hi8 = _mm512_shuffle_f32x4(acc7, acc7, 238);
acc1 = _mm512_mask_add_ps(acc1, 255, acc1, hi1);
acc3 = _mm512_mask_add_ps(acc3, 255, acc3, hi6);
acc5 = _mm512_mask_add_ps(acc5, 255, acc5, hi3);
acc7 = _mm512_mask_add_ps(acc7, 255, acc7, hi8);
__m512 hi2 = _mm512_shuffle_f32x4(acc1, acc1, 1);
__m512 hi7 = _mm512_shuffle_f32x4(acc3, acc3, 1);
__m512 hi4 = _mm512_shuffle_f32x4(acc5, acc5, 1);
__m512 hi9 = _mm512_shuffle_f32x4(acc7, acc7, 1);
acc1 = _mm512_mask_add_ps(acc1, 15, acc1, hi2);
acc3 = _mm512_mask_add_ps(acc3, 15, acc3, hi7);
acc5 = _mm512_mask_add_ps(acc5, 15, acc5, hi4);
acc7 = _mm512_mask_add_ps(acc7, 15, acc7, hi9);
__m512 hi5 = _mm512_shuffle_ps(acc1, acc5, 238);
__m512 hi10 = _mm512_shuffle_ps(acc3, acc7, 238);
acc1 = _mm512_shuffle_ps(acc1, acc5, 68);
acc3 = _mm512_shuffle_ps(acc3, acc7, 68);
acc1 = _mm512_mask_add_ps(acc1, 15, acc1, hi5);
acc3 = _mm512_mask_add_ps(acc3, 15, acc3, hi10);
__m512 hi11 = _mm512_permutex2var_ps(acc1, pm1hi1, acc3);
acc1 = _mm512_permutex2var_ps(acc1, pm1lo1, acc3);
acc1 = _mm512_mask_add_ps(acc1, 65535, acc1, hi11);
buf1 = _mm512_mask_mov_ps(buf1, mask3, acc1);
mask3 &= mask3<<4;
if (__builtin_expect(!mask3, 0)) {
mask3 = 65535;
buf1 = _mm512_mul_ps(buf1, _mm512_set1_ps(2.9411765e-03f));
_mm512_mask_storeu_ps(ptr2+(ptrdiff_t)4*((ptrdiff_t)4*i5-12), 65535, buf1);
}
}
buf1 = _mm512_mul_ps(buf1, _mm512_set1_ps(2.9411765e-03f));
_mm512_mask_storeu_ps(ptr2+(ptrdiff_t)4*16, 255, buf1);
return;
}
__m512 buf2 = _mm512_setzero_ps();
__mmask16 mask4 = 65535;
for (ptrdiff_t i6 = 0; i6 < 9; ++i6) {
__m512 acc9 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc10 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)64+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc11 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)128+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc12 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)192+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc13 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)256+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc14 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)320+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc15 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)384+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
__m512 acc16 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)448+(ptrdiff_t)1360*i6+(ptrdiff_t)512*0);
for (ptrdiff_t j2 = 1; j2 < 2; ++j2) {
__m512 dat17 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat18 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)64+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat19 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)128+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat20 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)192+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat21 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)256+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat22 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)320+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat23 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)384+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
__m512 dat24 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)448+(ptrdiff_t)1360*i6+(ptrdiff_t)512*j2);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, dat17);
acc10 = _mm512_mask_add_ps(acc10, 65535, acc10, dat18);
acc11 = _mm512_mask_add_ps(acc11, 65535, acc11, dat19);
acc12 = _mm512_mask_add_ps(acc12, 65535, acc12, dat20);
acc13 = _mm512_mask_add_ps(acc13, 65535, acc13, dat21);
acc14 = _mm512_mask_add_ps(acc14, 65535, acc14, dat22);
acc15 = _mm512_mask_add_ps(acc15, 65535, acc15, dat23);
acc16 = _mm512_mask_add_ps(acc16, 65535, acc16, dat24);
}
__m512 dat25 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)0+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
__m512 dat26 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)64+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
__m512 dat27 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)128+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
__m512 dat28 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)192+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
__m512 dat29 = _mm512_maskz_loadu_ps(65535, ptr1+(ptrdiff_t)256+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
__m512 dat30 = _mm512_maskz_loadu_ps(15, ptr1+(ptrdiff_t)320+(ptrdiff_t)1360*i6+(ptrdiff_t)512*2);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, dat25);
acc10 = _mm512_mask_add_ps(acc10, 65535, acc10, dat26);
acc11 = _mm512_mask_add_ps(acc11, 65535, acc11, dat27);
acc12 = _mm512_mask_add_ps(acc12, 65535, acc12, dat28);
acc13 = _mm512_mask_add_ps(acc13, 65535, acc13, dat29);
acc14 = _mm512_mask_add_ps(acc14, 15, acc14, dat30);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, acc13);
acc10 = _mm512_mask_add_ps(acc10, 65535, acc10, acc14);
acc11 = _mm512_mask_add_ps(acc11, 65535, acc11, acc15);
acc12 = _mm512_mask_add_ps(acc12, 65535, acc12, acc16);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, acc11);
acc10 = _mm512_mask_add_ps(acc10, 65535, acc10, acc12);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, acc10);
__m512i pm1lo2 = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
__m512i pm1hi2 = _mm512_set_epi32(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1);
__m512 hi12 = _mm512_shuffle_f32x4(acc9, acc9, 238);
acc9 = _mm512_mask_add_ps(acc9, 255, acc9, hi12);
__m512 hi13 = _mm512_shuffle_f32x4(acc9, acc9, 1);
acc9 = _mm512_mask_add_ps(acc9, 15, acc9, hi13);
__m512 hi14 = _mm512_shuffle_ps(acc9, acc9, 238);
acc9 = _mm512_mask_add_ps(acc9, 3, acc9, hi14);
__m512 hi15 = _mm512_permutexvar_ps(pm1hi2, acc9);
acc9 = _mm512_permutexvar_ps(pm1lo2, acc9);
acc9 = _mm512_mask_add_ps(acc9, 65535, acc9, hi15);
buf2 = _mm512_mask_mov_ps(buf2, mask4, acc9);
mask4 &= mask4<<1;
}
buf2 = _mm512_mul_ps(buf2, _mm512_set1_ps(2.9411765e-03f));
_mm512_mask_storeu_ps(ptr2+(ptrdiff_t)4*0, 511, buf2);
}
static void Example24Glopl1(Example24ThreaderTeam1* team13, char** tensors1) {
Example24ThreaderTask1 task5;
task5.callee1 = Example24Glopl1Callee1;
task5.any1 = tensors1;
task5.nd1 = 1;
task5.hull1[0] = 44;
Example24ThreaderDo1(team13, &task5);
}
struct Example24Net {
char* alloc1;
char* align1;
};
void Example24NetDestroy(Example24Net* net2) {
free(net2->alloc1);
free(net2);
}
char* Example24NetCreate(
Example24Net** net1,
Example24Params* params1,
ptrdiff_t threads1
) {
(void)params1;
(void)threads1;
if (__builtin_expect(!__builtin_cpu_supports("avx512f"), 0)) {
return Example24Errmsg1(__LINE__, "CPU does not support AVX512F");
}
Example24Net* net5 = malloc(sizeof(Example24Net));
if (__builtin_expect(!net5, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", errno);
}
net5->alloc1 = 0;
net5->align1 = 0;
*net1 = net5;
return 0;
}
struct Example24Engine {
Example24Net* net3;
Example24ThreaderTeam1* team11;
char* alloc2;
char* align2;
};
char* Example24EnginePthreadT(
Example24Engine* eng2,
ptrdiff_t idx2,
pthread_t* to1
) {
return Example24ThreaderPthreadT1(to1, eng2->team11, idx2);
}
void Example24EngineDestroy(Example24Engine* eng3) {
Example24ThreaderDestroy1(eng3->team11);
free(eng3->alloc2);
free(eng3);
}
char* Example24EngineCreate(
Example24Engine** eng4,
Example24Net* net4,
ptrdiff_t threads2
) {
Example24Engine* eng5 = malloc(sizeof(Example24Engine));
if (__builtin_expect(!eng5, 0)) {
return Example24Errmsg1(__LINE__, "errno %d", errno);
}
char* alloc3 = malloc(63);
if (__builtin_expect(!alloc3, 0)) {
char* msg5 = Example24Errmsg1(__LINE__, "errno %d", errno);
free(eng5);
return msg5;
}
eng5->alloc2 = alloc3;
eng5->align2 = (void*)(((size_t)alloc3+63)&-64);
char* err7 = Example24ThreaderCreate1(&eng5->team11, threads2);
if (__builtin_expect(!!err7, 0)) {
free(eng5);
free(alloc3);
return err7;
}
eng5->net3 = net4;
*eng4 = eng5;
return 0;
}
void Example24EngineInference(
Example24Engine* eng1,
float* inData,
float* outData
) {
Example24ThreaderTeam1* team12 = eng1->team11;
{
char* tensors3[] = {
(char*)inData,
(char*)outData
};
Example24Glopl1(team12, tensors3);
}
}
// End of file.