// Each thread starts here and performs 16 hashes simultaneously. With four
// threads, there are 64 hashes in flight at a time. Each thread repeats this
// four times. The total number of hashes performed is 256.
int main()
{
__builtin_nyuzi_write_control_reg(30, 0xffffffff); // Start other threads
const int kSourceBlockSize = 128;
const int kHashSize = 32;
const int kNumBuffers = 2;
const int kNumLanes = 16;
unsigned int basePtr = 0x100000 + __builtin_nyuzi_read_control_reg(0) * (kHashSize * kNumLanes * kNumBuffers)
+ (kSourceBlockSize * kNumLanes);
const vecu16_t kStepVector = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
vecu16_t inputPtr = __builtin_nyuzi_makevectori(basePtr) + (kStepVector * __builtin_nyuzi_makevectori(kHashSize));
vecu16_t tmpPtr = inputPtr + __builtin_nyuzi_makevectori(kSourceBlockSize * kNumLanes);
vecu16_t outputPtr = tmpPtr + __builtin_nyuzi_makevectori(kHashSize * kNumLanes);
for (int i = 0; i < 4; i++)
{
// Double sha-2 hash
sha2Hash(inputPtr, kSourceBlockSize / kHashSize, outputPtr);
sha2Hash(tmpPtr, 1, outputPtr);
}
return 0;
}
void copyTest()
{
veci16_t *dest = (veci16_t*) region1Base + currentThread() * LOOP_UNROLL;
veci16_t *src = (veci16_t*) region2Base + currentThread() * LOOP_UNROLL;
veci16_t values = __builtin_nyuzi_makevectori(0xdeadbeef);
int transferCount = kTransferSize / (64 * NUM_THREADS * LOOP_UNROLL);
int unrollCount;
int startTime = getTime();
startParallel();
do
{
// The compiler will automatically unroll this
for (unrollCount = 0; unrollCount < LOOP_UNROLL; unrollCount++)
dest[unrollCount] = src[unrollCount];
dest += NUM_THREADS * LOOP_UNROLL;
src += NUM_THREADS * LOOP_UNROLL;
}
while (--transferCount);
endParallel();
if (currentThread() == 0)
{
int endTime = getTime();
printf("copy: %g bytes/cycle\n", (float) kTransferSize / (endTime - startTime));
}
}
int main()
{
veci16 *dest = (veci16*) region1Base + __builtin_nyuzi_read_control_reg(0) * LOOP_UNROLL;
veci16 values = __builtin_nyuzi_makevectori(0xdeadbeef);
int transferCount = kTransferSize / (64 * NUM_STRANDS * LOOP_UNROLL);
do
{
dest[0] = values;
dest[1] = values;
dest[2] = values;
dest[3] = values;
dest[4] = values;
dest[5] = values;
dest[6] = values;
dest[7] = values;
dest += NUM_STRANDS * LOOP_UNROLL;
}
while (--transferCount);
}
void* memset(void *_dest, int value, size_t length)
{
char *dest = (char*) _dest;
value &= 0xff;
// XXX Possibly fill bytes/words until alignment is hit
if ((((unsigned int) dest) & 63) == 0)
{
// Write 64 bytes at a time.
veci16_t reallyWideValue = __builtin_nyuzi_makevectori(value | (value << 8) | (value << 16)
| (value << 24));
while (length > 64)
{
*((veci16_t*) dest) = reallyWideValue;
length -= 64;
dest += 64;
}
}
if ((((unsigned int) dest) & 3) == 0)
{
// Write 4 bytes at a time.
unsigned wideVal = value | (value << 8) | (value << 16) | (value << 24);
while (length > 4)
{
*((unsigned int*) dest) = wideVal;
dest += 4;
length -= 4;
}
}
// Write one byte at a time
while (length > 0)
{
*dest++ = value;
length--;
}
return _dest;
}
int main(void)
{
veci16_t pointers = { &foo, &foo, &foo, &foo, &foo, &foo, 0x17, &foo, &foo, &foo, &foo, &foo, &foo, &foo, &foo, &foo };
__builtin_nyuzi_write_control_reg(CR_FAULT_HANDLER, (unsigned int) faultHandler);
__builtin_nyuzi_write_control_reg(CR_TLB_MISS_HANDLER, (unsigned int) tlb_miss_handler);
__builtin_nyuzi_write_control_reg(CR_FLAGS, FLAG_MMU_EN | FLAG_SUPERVISOR_EN);
// This ensures the libc functions are mapped into the TLB so we don't generate
// multiple TLB misses in the fault handler (doesn't break the test, just makes
// debugging cleaner)
printf("Starting test %d\n", 12);
// This will cause an alignment fault on the 6th lane and jump to 'faultHandler'.
// Use scatter store rather than a normal scalar store to ensure the
// subcycle counter is saved correctly.
__builtin_nyuzi_scatter_storei(pointers, __builtin_nyuzi_makevectori(0));
printf("should_not_be_here\n"); // CHECKN: should_not_be_here
return 0;
}
// Run 16 parallel hashes
void sha2Hash(vecu16_t pointers, int totalBlocks, vecu16_t outHashes)
{
// Initial H values
vecu16_t h0 = __builtin_nyuzi_makevectori(0x6A09E667);
vecu16_t h1 = __builtin_nyuzi_makevectori(0xBB67AE85);
vecu16_t h2 = __builtin_nyuzi_makevectori(0x3C6EF372);
vecu16_t h3 = __builtin_nyuzi_makevectori(0xA54FF53A);
vecu16_t h4 = __builtin_nyuzi_makevectori(0x510E527F);
vecu16_t h5 = __builtin_nyuzi_makevectori(0x9B05688C);
vecu16_t h6 = __builtin_nyuzi_makevectori(0x1F83D9AB);
vecu16_t h7 = __builtin_nyuzi_makevectori(0x5BE0CD19);
for (int i = 0; i < totalBlocks; i++)
{
vecu16_t w[64];
for (int index = 0; index < 16; index++)
{
w[index] = __builtin_nyuzi_gather_loadi(pointers);
pointers += __builtin_nyuzi_makevectori(4);
}
for (int index = 16; index < 64; index++)
w[index] = SIG1(w[index - 2]) + w[index - 7] + SIG0(w[index - 15]) + w[index - 16];
vecu16_t a = h0;
vecu16_t b = h1;
vecu16_t c = h2;
vecu16_t d = h3;
vecu16_t e = h4;
vecu16_t f = h5;
vecu16_t g = h6;
vecu16_t h = h7;
for (int round = 0; round < 64; round++)
{
vecu16_t temp1 = h + SIG1(e) + CH(e, f, g) + __builtin_nyuzi_makevectori(K[round]) + w[round];
vecu16_t temp2 = SIG0(a) + MA(a, b, c);
h = g;
g = f;
f = e;
e = d + temp1;
d = c;
c = b;
b = a;
a = temp1 + temp2;
}
h0 += a;
h1 += b;
h2 += c;
h3 += d;
h4 += e;
h5 += f;
h6 += g;
h7 += h;
}
// doesn't add padding or length fields to end...
__builtin_nyuzi_scatter_storei(outHashes, h0);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(4), h1);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(8), h2);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(12), h3);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(16), h4);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(20), h5);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(24), h6);
__builtin_nyuzi_scatter_storei(outHashes + __builtin_nyuzi_makevectori(28), h7);
}
inline vecu16_t ROTR(vecu16_t x, int y)
{
return (x >> __builtin_nyuzi_makevectori(y)) | (x << (__builtin_nyuzi_makevectori(32 - y)));
}