Skip to content

GitLab

"git@developer.sourcefind.cn:modelzoo/solov2-pytorch.git" did not exist on "c10ae5d2ca158ae506b93198afa4926ea48bb40b"
Commit cb13b5d2 authored by aska-0096's avatar aska-0096
Browse files

ATI Conv Instances

parent 59f33851
Hide whitespace changes
Inline Side-by-side
Showing with 65 additions and 54 deletions
example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
View file @ cb13b5d2
......@@ -35,19 +35,8 @@ using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayo
// clang-format off
template <ck::index_t NDimSpatial>
using DeviceConvFwdInstances = std::tuple<
#if 0
#if 1
// Instances provide to AIT Fp16
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
512, 16, 4, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 1, // MRepeat x NRepeat
S<1, 256, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 16, 4>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 2, true,
2, 1, S<1, 256, 1, 1>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -55,21 +44,21 @@ using DeviceConvFwdInstances = std::tuple<
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
256, 64, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
2, 4, // MRepeat x NRepeat
S<1, 256, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
2, 2, S<1, 256, 1, 1>, 8>,
8, 1, // MRepeat x NRepeat
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
256, 64, 4, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
64, 256, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
2, 4, // MRepeat x NRepeat
S<1, 256, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
2, 2, S<1, 256, 1, 1>, 8>,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -77,10 +66,10 @@ using DeviceConvFwdInstances = std::tuple<
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
256, 128, 4, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 4, // MRepeat x NRepeat
S<1, 256, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
4, 2, S<1, 256, 1, 1>, 8>,
8, 2, // MRepeat x NRepeat
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -88,10 +77,10 @@ using DeviceConvFwdInstances = std::tuple<
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
256, 128, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 4, // MRepeat x NRepeat
S<1, 256, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
4, 2, S<1, 256, 1, 1>, 8>,
8, 2, // MRepeat x NRepeat
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -100,9 +89,9 @@ using DeviceConvFwdInstances = std::tuple<
128, 256, 4, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 4, // MRepeat x NRepeat
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
4, 2, S<1, 128, 1, 2>, 8>,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -111,20 +100,20 @@ using DeviceConvFwdInstances = std::tuple<
128, 256, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 4, // MRepeat x NRepeat
S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
4, 2, S<1, 128, 1, 2>, 8>,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 192, // BlockSize
192, 48, 6, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 256, // BlockSize
128, 128, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
6, 1, // MRepeat x NRepeat
S<1, 192, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<2, 48, 2>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 4, true,
3, 1, S<1, 96, 1, 2>, 8>,
4, 2, // MRepeat x NRepeat
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 8>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -132,21 +121,21 @@ using DeviceConvFwdInstances = std::tuple<
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 128, // BlockSize
128, 64, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
2, 4, // MRepeat x NRepeat
S<1, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<2, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
2, 2, S<1, 128, 1, 1>, 8>,
4, 2, // MRepeat x NRepeat
S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 4>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 96, // BlockSize
96, 48, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 128, // BlockSize
64, 128, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
6, 1, // MRepeat x NRepeat
S<1, 96, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<2, 48, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
3, 1, S<1, 48, 1, 2>, 8>,
2, 4, // MRepeat x NRepeat
S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 32, 1, 4>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
......@@ -155,11 +144,33 @@ using DeviceConvFwdInstances = std::tuple<
64, 64, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
4, 2, // MRepeat x NRepeat
S<1, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<1, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
4, 1, S<1, 64, 1, 1>, 8>
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 16, 1, 4>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 64, // BlockSize
128, 32, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
8, 1, // MRepeat x NRepeat
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 16, 1, 4>, 8>,
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
NDimSpatial,
InputLayout<NDimSpatial>, WeightLayout<NDimSpatial>, ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>, OutputLayout<NDimSpatial>,
InKernelDataType, WeiKernelDataType, ck::Tuple<BiasKernelDataType, ResidualKernelDataType>, OutKernelDataType, AccDataType, CShuffleDataType,
InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmSpec, 64, // BlockSize
32, 128, 8, 8, // MPerBlock x NPerBlock x K0PerBlock x K1
16, 16, // MPerWMMA x NPerWMMA
2, 4, // MRepeat x NRepeat
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true,
1, 1, S<1, 16, 1, 4>, 8>
#endif
#if 1
#if 0
// GEMM_N = 16
// K0 = 8
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment