ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Overhaul to Reducton and its dependants (#237)

Browse Source 
* Tiny fix in dynamic_buffer.hpp to support vectorized AtomicAdd for double type

* Update to host layer and host reduction

* Merge and remove reduction kernels

* Merge and remove reduction device interfaces and update pooling device interface

* Merge and remove useless reduction device instances

* Update to reduction profiler and reduction ctests

* Update to reduction and pooling examples and add one reduction example

* Change to reduction examples to let them testable by ctest

* Add explicit pass checking for reduction and pooling examples

* Explicit assignment of tensor shapes in example reduce_blockwise_two_call

* Use atomic_add to repace atomicAdd and add atomic_add for double type

* Add reduce ctest support for double data type

* Replace to_int_vector() by using c++ std::vector::assign()

* Keep DeviceReduceThreadWise separated from DeviceReduceBlockWise

* Merge DeviceReduceBlockWise and DeviceReduceMultiBlockAtomicAdd into DeviceReduceMultiBlock

* Add GetAtomicOperationZeroValue() support for AtomicMax

* Tiny change to reduce example README.md

* Fix some tiny issues due to branch merging

* Revoke previous change in dynamic_buffer.hpp and add atomic_add for double2_t

* Add reduce multiblock_atomic_add instances for fp64 to verify vectorized atomic_add on fp64

* Renaming

* Clean the header includings in device_reduce instances header files
This commit is contained in:
Qianfeng
2022-05-25 01:19:12 +08:00
committed by GitHub
parent 1085794df3
commit 63eee2d999
94 changed files with 2429 additions and 6785 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
example/13_pool2d_fwd/README.md
View File

@@ -4,9 +4,9 @@
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: run kernel # of times (>1)
#arg3: time kernel (0=no, 1=yes)
#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_pool2d_fwd 1 1 10
./bin/example_pool2d_fwd 1 1 1
```
Result
@@ -14,9 +14,7 @@ Result
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1}
Warm up
Warm up 1 time
Start running 10 times...
Perf: 0.415453 ms, 1.37996 TFlops, 749.726 GB/s
error: 0
max_diff: 0, 1, 1
Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
```

114
example/13_pool2d_fwd/pool2d_fwd.cpp
View File

@@ -20,6 +20,8 @@ using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
using IndexDataType = int32_t;
using InLayout = ck::tensor_layout::convolution::NHWC;
using OutLayout = ck::tensor_layout::convolution::NHWC;
@@ -29,7 +31,7 @@ static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
static constexpr bool NeedIndices = false;
static constexpr bool OutputIndex = false;
static constexpr bool PropagateNan = false;
using DevicePoolFwdInstance =
@@ -38,7 +40,7 @@ using DevicePoolFwdInstance =
OutDataType, // OutDataType
AccDataType, // AccDataType
ReduceOpId,
NeedIndices,
OutputIndex,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
@@ -51,10 +53,10 @@ template <typename InDataType,
typename AccDataType,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool NeedIndices>
bool OutputIndex>
static void pool_host_verify(const Tensor<InDataType>& in,
Tensor<OutDataType>& out,
Tensor<int>& out_indices,
Tensor<IndexDataType>& out_indices,
const std::array<ck::index_t, 2>& window_spatial_lengths,
const std::array<ck::index_t, 2>& window_strides,
const std::array<ck::index_t, 2>& in_left_pads,
@@ -62,26 +64,26 @@ static void pool_host_verify(const Tensor<InDataType>& in,
{
using namespace ck::host_reduce;
const int divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const int32_t divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const auto PreUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
const auto PosUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
if constexpr(!NeedIndices)
if constexpr(!OutputIndex)
{
auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
for(int y = 0; y < window_spatial_lengths[0]; ++y)
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < ck::type_convert<int>(in.mDesc.GetLengths()[2]) && wi >= 0 &&
wi < ck::type_convert<int>(in.mDesc.GetLengths()[3]))
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
@@ -108,24 +110,24 @@ static void pool_host_verify(const Tensor<InDataType>& in,
auto opReduce = ReduceOpFn2<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
int accuIndex = 0;
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
IndexDataType accuIndex = 0;
for(int y = 0; y < window_spatial_lengths[0]; ++y)
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
int currIndex = y * window_spatial_lengths[1] + x;
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
IndexDataType currIndex = y * window_spatial_lengths[1] + x;
PreUnaryOp(currVal);
binop_with_nan_check2<AccDataType, PropagateNan>(
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce, accuVal, currVal, accuIndex, currIndex);
}
}
@@ -149,9 +151,9 @@ int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool do_verification;
int init_method;
bool time_kernel;
// Pool shape
ck::index_t N = 128;
@@ -167,17 +169,23 @@ int main(int argc, char* argv[])
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
if(argc == 1)
{
do_verification = true;
init_method = 1;
time_kernel = true;
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
N = std::stoi(argv[4]);
C = std::stoi(argv[5]);
@@ -196,7 +204,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
@@ -228,9 +236,11 @@ int main(int argc, char* argv[])
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
@@ -245,25 +255,25 @@ int main(int argc, char* argv[])
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_c_ho_wo_device.mDesc.GetElementSpace());
DeviceMem out_indices_device_buf(sizeof(int) *
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
out_indices_n_c_ho_wo_device.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
auto pool = DevicePoolFwdInstance{};
auto invoker_ptr = pool.MakeInvokerPointer();
auto argument_ptr =
pool.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<int*>(out_indices_device_buf.GetDeviceBuffer()),
N,
C,
std::array<ck::index_t, 2>{{Hi, Wi}},
std::array<ck::index_t, 2>{{Y, X}},
std::array<ck::index_t, 2>{{Ho, Wo}},
window_strides,
input_left_pads,
input_right_pads);
auto pool = DevicePoolFwdInstance{};
auto invoker_ptr = pool.MakeInvokerPointer();
auto argument_ptr = pool.MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
N,
C,
std::array<ck::index_t, 2>{{Hi, Wi}},
std::array<ck::index_t, 2>{{Y, X}},
std::array<ck::index_t, 2>{{Ho, Wo}},
window_strides,
input_left_pads,
input_right_pads);
if(!pool.IsSupportedArgument(argument_ptr.get()))
{
@@ -286,6 +296,7 @@ int main(int argc, char* argv[])
<< std::endl;
bool pass = true;
if(do_verification)
{
pool_host_verify<InDataType,
@@ -293,7 +304,7 @@ int main(int argc, char* argv[])
AccDataType,
ReduceOpId,
PropagateNan,
NeedIndices>(in_n_c_hi_wi,
OutputIndex>(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
@@ -303,15 +314,16 @@ int main(int argc, char* argv[])
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
pass &= ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
pass = pass && ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
if constexpr(NeedIndices)
if constexpr(OutputIndex)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
pass &= ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
};
}
return pass ? 0 : 1;
return (pass ? 0 : 1);
}