thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

UPdated to have shorter dispatch size and larger loop size

Browse source
This commit is contained in:
Alejandro Saucedo 2020-10-16 07:39:37 +01:00
parent 9e79b9f352
commit c69fcb7e60
1 changed files with 26 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

36
test/TestAsyncOperations.cpp
View file

@ -7,7 +7,7 @@
TEST(TestAsyncOperations, TestManagerAsync)
{
uint32_t size = 100000;
uint32_t size = 10;
uint32_t numParallel = 6;
@ -19,22 +19,38 @@ TEST(TestAsyncOperations, TestManagerAsync)
layout(set = 0, binding = 0) buffer a { float pa[]; };
layout(set = 0, binding = 1) buffer b { float pb[]; };
shared uint sharedTotal[1];
void main() {
uint index = gl_GlobalInvocationID.x;
for (int i = 0; i < 10000; i++)
sharedTotal[0] = 0;
barrier();
memoryBarrierShared();
for (int i = 0; i < 100000000; i++)
{
pa[index] += 1.0;
atomicAdd(sharedTotal[0], 1);
atomicAdd(sharedTotal[0], -1);
atomicAdd(sharedTotal[0], 1);
atomicAdd(sharedTotal[0], -1);
atomicAdd(sharedTotal[0], 1);
atomicAdd(sharedTotal[0], -1);
atomicAdd(sharedTotal[0], 1);
}
pb[index] = pa[index];
barrier();
memoryBarrierShared();
pb[index] = sharedTotal[0];
pa[index] = 0;
}
)");
std::vector<float> data(size, 0.0);
std::vector<float> resultSync(size, 10000);
std::vector<float> resultAsync(size, 10000);
std::vector<float> resultSync(size, 100000000);
std::vector<float> resultAsync(size, 100000000);
kp::Manager mgr;
@ -58,11 +74,11 @@ TEST(TestAsyncOperations, TestManagerAsync)
}
mgr.evalOpDefault<kp::OpTensorSyncLocal>(inputsSyncB);
auto endSync = std::chrono::high_resolution_clock::now();
auto durationSync = std::chrono::duration_cast<std::chrono::microseconds>(endSync - startSync).count();
mgr.evalOpDefault<kp::OpTensorSyncLocal>(inputsSyncB);
for (uint32_t i = 0; i < numParallel; i++) {
EXPECT_EQ(inputsSyncB[i]->data(), resultSync);
}
@ -97,11 +113,11 @@ TEST(TestAsyncOperations, TestManagerAsync)
mgrAsync.evalOpAwait("async" + std::to_string(i));
}
mgrAsync.evalOpDefault<kp::OpTensorSyncLocal>({ inputsAsyncB });
auto endAsync = std::chrono::high_resolution_clock::now();
auto durationAsync = std::chrono::duration_cast<std::chrono::microseconds>(endAsync - startAsync).count();
mgrAsync.evalOpDefault<kp::OpTensorSyncLocal>({ inputsAsyncB });
for (uint32_t i = 0; i < numParallel; i++) {
EXPECT_EQ(inputsAsyncB[i]->data(), resultAsync);
}