liuliu

static void _task_entry_point(uint32_t part0, uint32_t part1)
{
	union ptr_splitter p;
	p.part[0] = part0;
	p.part[1] = part1;
	task_t *task = (task_t*)p.ptr;
	task->fn(task);
	task->done = 1;
	swapcontext(&task->schd->callee, &task->schd->caller);
}

liuliu

static void deltask(schd_t* const schd, task_t* const t)
{
	if (t->prev)
		t->prev->next = t->next;
	else
		schd->head = t->next;
	if (t->next)
		t->next->prev = t->prev;
	else
		schd->tail = t->prev;

liuliu

union ptr_splitter {
	void *ptr;
	uint32_t part[2];
};

static const int default_stack_size = 65536;

typedef struct schd_s schd_t;
typedef struct task_s task_t;
typedef void (*task_fn_t)(task_t *task);

liuliu

#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <ucontext.h>
#include <pthread.h>
#include "nnc/gpu/ccv_nnc_compat.h"

union ptr_splitter {
	void *ptr;
	uint32_t part[2];

liuliu

#include <stdio.h>
#include <string.h>

static inline int _ccv_nnc_try_sort_and_insert(int* md, const int ins, const int c)
{
	if (!c)
	{
		md[0] = ins;
		return 1;
	}

liuliu

My 4-GPU Titan / Titan Black Setup is like this:

PSU: www.newegg.com/Product/Product.aspx?Item=N82E16817182251

Motherboard: http://www.newegg.com/Product/Product.aspx?Item=N82E16813132262 (You have to switch the PCI-e port to at x8 speed and change boot parameter to include "pci=nomsi irqpoll" to get stable performance).

CPU: http://www.newegg.com/Product/Product.aspx?Item=N82E16819117402

Memory: http://www.newegg.com/Product/Product.aspx?Item=N82E16820231793

liuliu

#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <zlib.h>
#include <assert.h>

static size_t get_atom_size(uint8_t* atom)
{
	return (atom[0] << 24) | (atom[1] << 16) | (atom[2] << 8) | atom[3];
}

liuliu

static void ccv_convnet_read_extra(ccv_convnet_t* convnet, const char* filename)
{
	sqlite3* db = 0;
	if (SQLITE_OK == sqlite3_open(filename, &db))
	{
		sqlite3_stmt* layer_quant_stmt = 0;
		const char layer_quant_qs[] =
			"SELECT layer, quant FROM layer_quant;";
		if (SQLITE_OK == sqlite3_prepare_v2(db, layer_quant_qs, sizeof(layer_quant_qs), &layer_quant_stmt, 0))
		{

liuliu

sqlite3* db = 0;
assert(SQLITE_OK == sqlite3_open(argv[2], &db));
const char create_table_qs[] =
	"CREATE TABLE IF NOT EXISTS layer_quant "
	"(layer INTEGER PRIMARY KEY ASC, quant BLOB);";
assert(SQLITE_OK == sqlite3_exec(db, create_table_qs, 0, 0, 0));
const char data_insert_qs[] =
	"REPLACE INTO layer_quant "
	"(layer, quant) VALUES ($layer, $quant);";
sqlite3_stmt* data_insert_stmt = 0;

liuliu

int f;
for (f = 10; f < 13; f++)
{
	int i, j, k;
	ccv_convnet_layer_t *layer = convnet->layers + f;
	char filename[256];
	float kmean[256];
	sprintf(filename, "lut-layer-%d.txt", f);
	FILE *r = fopen(filename, "r");
	for (i = 0; i < 0x100; i++)