mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Unified conv3D API + support for all data types. (#133)
* Convolution ND * Code unification across dimensions for generating tensor descriptors. * Example * Instances * Move convnd f32 instance file to comply with repo structure. * Conv 1D tensor layouts. * Formatting and use ReferenceConv * Reference ConvFwd supporting 1D and 2D convolution. * Debug printing TensorLayout name. * Conv fwd 1D instance f32 * Refactor conv ND example. Needed to support various conv dimensio. Needed to support various conv dimensions * Rename conv nd example director to prevent conflicts. * Refactor some common utility to single file. Plus some tests. * Refactor GetHostTensorDescriptor + UT. * Add 1D test case. * Test reference convolution 1d/2d * Remove some leftovers. * Fix convolution example error for 1D * Refactor test check errors utility function. * Test Conv2D Fwd XDL * More UT for 1D case. * Parameterize input & weight initializers. * Rename example to prevent conflicts. * Split convnd instance into separate files for 1d/2d * Address review comments. * Fix data type for flops/gbytes calculations. * Assign example number 11. * 3D cases for convolution utility functions. * 3D reference convolution. * Add support for 3D convolution. * Check for inputs bigger than 2GB. * Formatting * Support for bf16/f16/f32/i8 - conv instances + UT. * Use check_err from test_util.hpp. * Split convnd test into separate files for each dim. * Fix data generation and use proper instances. * Formatting * Skip tensor initialization if not necessary. * Fix CMakefiles. * Remove redundant conv2d_fwd test. * Lower problem size for conv3D UT. * 3D case for convnd example. * Remove leftovers after merge. * Add Conv Specialization string to GetTypeString * Skip instance causing numerical errors. * Small fixes. * Remove redundant includes. * Fix namespace name error. * Script for automatic testing and logging convolution fwd UTs * Comment out numactl cmd. * Refine weights initalization and relax rtol for fp16 * Fix weights initialization for int8. * Add type_convert when store output in ref conv 1D. * Get back old conv2d_fwd_xdl operation. * Silence conv debug print. * format * clean * clean * Fix merge. * Fix namespace for check_err Co-authored-by: Adam Osewski <aosewski@amd.com> Co-authored-by: Chao Liu <chao.liu2@amd.com>
This commit is contained in:
@@ -23,11 +23,16 @@ template <typename T>
|
||||
struct FillMonotonicSeq
|
||||
{
|
||||
T m_init_value{0};
|
||||
T m_step{1};
|
||||
|
||||
template <typename ForwardIter>
|
||||
void operator()(ForwardIter first, ForwardIter last) const
|
||||
{
|
||||
std::iota(first, last, m_init_value);
|
||||
std::generate(first, last, [=, n = m_init_value]() mutable {
|
||||
auto tmp = n;
|
||||
n += m_step;
|
||||
return tmp;
|
||||
});
|
||||
}
|
||||
};
|
||||
|
||||
@@ -53,7 +58,7 @@ template <ck::index_t NDim,
|
||||
typename FillInputOp = FillMonotonicSeq<InDataType>,
|
||||
typename FillWeightsOp = FillConstant<WeiDataType>>
|
||||
Tensor<OutDataType> RunReferenceConv(const ck::conv_util::ConvParams& params,
|
||||
const FillInputOp& fill_input_op = FillInputOp{0},
|
||||
const FillInputOp& fill_input_op = FillInputOp{},
|
||||
const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f})
|
||||
{
|
||||
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
|
||||
@@ -84,6 +89,9 @@ Tensor<OutDataType> RunReferenceConv(const ck::conv_util::ConvParams& params,
|
||||
fill_weights_op(weights.begin(), weights.end());
|
||||
std::fill(host_output.begin(), host_output.end(), OutDataType(0.f));
|
||||
|
||||
// std::cout <<"input: " << input.mDesc << std::endl << input.mData << std::endl;
|
||||
// std::cout <<"weight: " << weights.mDesc << std::endl << weights.mData << std::endl;
|
||||
|
||||
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
|
||||
WeiDataType,
|
||||
OutDataType,
|
||||
@@ -104,6 +112,7 @@ Tensor<OutDataType> RunReferenceConv(const ck::conv_util::ConvParams& params,
|
||||
OutElementOp{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
// std::cout <<"output: " << host_output.mDesc << std::endl << host_output.mData << std::endl;
|
||||
return host_output;
|
||||
}
|
||||
|
||||
@@ -139,10 +148,10 @@ bool TestConv2DNHWC()
|
||||
472.5,
|
||||
490.5,
|
||||
508.5};
|
||||
res = res && test_util::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test_util::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
|
||||
params.N = 1;
|
||||
params.K = 2;
|
||||
@@ -162,10 +171,10 @@ bool TestConv2DNHWC()
|
||||
747., 747., 1138.5, 1138.5, 1174.5, 1174.5, 1210.5, 1210.5, 1246.5, 1246.5,
|
||||
1035., 1035., 1570.5, 1570.5, 1606.5, 1606.5, 1642.5, 1642.5, 1678.5, 1678.5,
|
||||
1323., 1323., 2002.5, 2002.5, 2038.5, 2038.5, 2074.5, 2074.5, 2110.5, 2110.5};
|
||||
res = res && test_util::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test_util::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
|
||||
return res;
|
||||
}
|
||||
@@ -194,10 +203,10 @@ bool TestConv1DNWC()
|
||||
ck::tensor_layout::convolution::NWK>(params);
|
||||
std::vector<std::size_t> ref_dims{1, 1, 4};
|
||||
std::vector<float> ref_data{7.5, 13.5, 19.5, 25.5};
|
||||
res = res && test_util::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test_util::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
|
||||
params.num_dim_spatial = 1;
|
||||
params.N = 1;
|
||||
@@ -219,10 +228,10 @@ bool TestConv1DNWC()
|
||||
ck::tensor_layout::convolution::NWK>(params);
|
||||
ref_dims = std::vector<std::size_t>{1, 2, 5};
|
||||
ref_data = std::vector<float>{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
|
||||
res = res && test_util::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test_util::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
|
||||
|
||||
params.num_dim_spatial = 1;
|
||||
params.N = 2;
|
||||
@@ -235,16 +244,14 @@ bool TestConv1DNWC()
|
||||
params.input_left_pads = std::vector<ck::index_t>{1};
|
||||
params.input_right_pads = std::vector<ck::index_t>{1};
|
||||
|
||||
auto out_tensor2 =
|
||||
RunReferenceConv<1,
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
ck::tensor_layout::convolution::NWC,
|
||||
ck::tensor_layout::convolution::KXC,
|
||||
ck::tensor_layout::convolution::NWK>(params, [](auto first, auto last) {
|
||||
std::generate(first, last, [n = 0]() mutable { return float(n++) * float(0.1f); });
|
||||
});
|
||||
auto out_tensor2 = RunReferenceConv<1,
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
ck::tensor_layout::convolution::NWC,
|
||||
ck::tensor_layout::convolution::KXC,
|
||||
ck::tensor_layout::convolution::NWK>(
|
||||
params, FillMonotonicSeq<float>{0.f, 0.1f});
|
||||
|
||||
ref_dims = std::vector<std::size_t>{2, 16, 16};
|
||||
ref_data = std::vector<float>{
|
||||
@@ -312,10 +319,94 @@ bool TestConv1DNWC()
|
||||
72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9,
|
||||
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4,
|
||||
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4};
|
||||
res = res && test_util::check_err(out_tensor2.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test_util::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!");
|
||||
res = res && test::check_err(out_tensor2.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!");
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
bool TestConv3DNCDHW()
|
||||
{
|
||||
bool res{true};
|
||||
ck::conv_util::ConvParams params;
|
||||
params.num_dim_spatial = 3;
|
||||
params.N = 1;
|
||||
params.K = 1;
|
||||
params.C = 2;
|
||||
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 3};
|
||||
params.input_spatial_lengths = std::vector<ck::index_t>{6, 6, 6};
|
||||
params.conv_filter_strides = std::vector<ck::index_t>{1, 1, 1};
|
||||
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1};
|
||||
params.input_left_pads = std::vector<ck::index_t>{0, 0, 0};
|
||||
params.input_right_pads = std::vector<ck::index_t>{0, 0, 0};
|
||||
|
||||
auto out_tensor = RunReferenceConv<3,
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
ck::tensor_layout::convolution::NCDHW,
|
||||
ck::tensor_layout::convolution::KCZYX,
|
||||
ck::tensor_layout::convolution::NKDHW>(
|
||||
params, FillMonotonicSeq<float>{0.f, 0.1f});
|
||||
std::vector<std::size_t> ref_dims{1, 1, 4, 4, 4};
|
||||
std::vector<float> ref_data{
|
||||
407.7, 410.40002, 413.09998, 415.80002, 423.90002, 426.6, 429.30002, 432.,
|
||||
440.1, 442.80002, 445.5, 448.2, 456.30002, 459., 461.7, 464.40002,
|
||||
504.90002, 507.6, 510.30002, 513., 521.1, 523.8, 526.5, 529.2001,
|
||||
537.3, 540., 542.7001, 545.4, 553.5, 556.2001, 558.9, 561.6,
|
||||
602.10004, 604.8, 607.5, 610.2, 618.3, 621., 623.7, 626.4,
|
||||
634.5, 637.2, 639.9, 642.60004, 650.7, 653.4, 656.10004, 658.8,
|
||||
699.3, 702., 704.7, 707.4, 715.5, 718.2, 720.9, 723.60004,
|
||||
731.7, 734.4001, 737.10004, 739.8, 747.9001, 750.60004, 753.3, 756.};
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error [case 1]: wrong output tensor dimensions!");
|
||||
res = res && test::check_err(out_tensor.mData, ref_data, "Error [case 1]: incorrect results!");
|
||||
|
||||
params.N = 1;
|
||||
params.K = 2;
|
||||
params.C = 2;
|
||||
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 3};
|
||||
params.input_spatial_lengths = std::vector<ck::index_t>{12, 12, 12};
|
||||
params.conv_filter_strides = std::vector<ck::index_t>{3, 3, 3};
|
||||
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1};
|
||||
params.input_left_pads = std::vector<ck::index_t>{0, 0, 0};
|
||||
params.input_right_pads = std::vector<ck::index_t>{0, 0, 0};
|
||||
|
||||
out_tensor = RunReferenceConv<3,
|
||||
float,
|
||||
float,
|
||||
float,
|
||||
ck::tensor_layout::convolution::NCDHW,
|
||||
ck::tensor_layout::convolution::KCZYX,
|
||||
ck::tensor_layout::convolution::NKDHW>(
|
||||
params, FillMonotonicSeq<float>{0.f, 0.1f});
|
||||
ref_dims = std::vector<std::size_t>{1, 2, 4, 4, 4};
|
||||
ref_data = std::vector<float>{
|
||||
2756.7002, 2764.7998, 2772.9001, 2781., 2853.9001, 2862., 2870.1, 2878.2002,
|
||||
2951.1, 2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5, 3072.6,
|
||||
3923.1, 3931.2, 3939.2998, 3947.4, 4020.2998, 4028.4001, 4036.5002, 4044.5999,
|
||||
4117.5, 4125.6, 4133.7, 4141.8, 4214.7, 4222.8, 4230.9004, 4239.,
|
||||
5089.5, 5097.5996, 5105.7, 5113.8, 5186.7, 5194.8, 5202.9, 5211.,
|
||||
5283.9004, 5292., 5300.0996, 5308.2, 5381.0996, 5389.2, 5397.3, 5405.4004,
|
||||
6255.9004, 6264.0005, 6272.1, 6280.2, 6353.1, 6361.2, 6369.301, 6377.4,
|
||||
6450.301, 6458.4, 6466.5, 6474.6, 6547.5, 6555.6, 6563.699, 6571.801,
|
||||
2756.7002, 2764.7998, 2772.9001, 2781., 2853.9001, 2862., 2870.1, 2878.2002,
|
||||
2951.1, 2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5, 3072.6,
|
||||
3923.1, 3931.2, 3939.2998, 3947.4, 4020.2998, 4028.4001, 4036.5002, 4044.5999,
|
||||
4117.5, 4125.6, 4133.7, 4141.8, 4214.7, 4222.8, 4230.9004, 4239.,
|
||||
5089.5, 5097.5996, 5105.7, 5113.8, 5186.7, 5194.8, 5202.9, 5211.,
|
||||
5283.9004, 5292., 5300.0996, 5308.2, 5381.0996, 5389.2, 5397.3, 5405.4004,
|
||||
6255.9004, 6264.0005, 6272.1, 6280.2, 6353.1, 6361.2, 6369.301, 6377.4,
|
||||
6450.301, 6458.4, 6466.5, 6474.6, 6547.5, 6555.6, 6563.699, 6571.801};
|
||||
res = res && test::check_err(out_tensor.mDesc.GetLengths(),
|
||||
ref_dims,
|
||||
"Error [case 2]: wrong output tensor dimensions!");
|
||||
res =
|
||||
res && test::check_err(
|
||||
out_tensor.mData, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f);
|
||||
|
||||
return res;
|
||||
}
|
||||
@@ -329,5 +420,7 @@ int main(void)
|
||||
std::cout << "TestConv2DNHWC ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
|
||||
res = TestConv1DNWC();
|
||||
std::cout << "TestConv1DNHWC ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
|
||||
res = TestConv3DNCDHW();
|
||||
std::cout << "TestConv3DNCDHW ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
|
||||
return 0;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user