Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "b37706389dde3e660a2f6f0ea53b4a9baaad9d4f"
Commit c2976d7a authored by ltqin's avatar ltqin
Browse files

Merge branch 'develop' into ck_conv_bww_fp16

parents e46ea9fd 0619ebf7
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1550 additions and 1287 deletions
composable_kernel/include/config.hpp
View file @ c2976d7a
...@@ -151,6 +151,12 @@ ...@@ -151,6 +151,12 @@
#define CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R5_TYPE_CONVERT_ISSUE 1 #define CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R5_TYPE_CONVERT_ISSUE 1
#endif #endif
// workaround for verifaction failure, due to compiler regression, for conv bwd-data fp16 using some
// tuning parameter
#ifndef CK_WORKAROUND_SWDEV_325164
#define CK_WORKAROUND_SWDEV_325164 1
#endif
namespace ck { namespace ck {
enum InMemoryDataOperationEnum_t enum InMemoryDataOperationEnum_t
......
composable_kernel/include/tensor_operation/element_wise_operation.hpp
View file @ c2976d7a
#ifndef CK_ELEMENT_WISE_OPERATION_HPP #ifndef CK_ELEMENT_WISE_OPERATION_HPP
#define CK_ELEMENT_WISE_OPERATION_HPP #define CK_ELEMENT_WISE_OPERATION_HPP
#include "data_type.hpp"
#include "data_type.hpp" #include "data_type.hpp"
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ c2976d7a
...@@ -5,6 +5,7 @@ ...@@ -5,6 +5,7 @@
#include "tensor_descriptor.hpp" #include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp" #include "tensor_descriptor_helper.hpp"
#include "static_tensor.hpp" #include "static_tensor.hpp"
#include "tensor_space_filling_curve.hpp"
namespace ck { namespace ck {
...@@ -123,73 +124,16 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -123,73 +124,16 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto src_scalar_per_access = generate_sequence( constexpr auto src_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
SrcDimAccessOrder,
constexpr auto src_dim_access_order = SrcDimAccessOrder{}; remove_cv_t<decltype(src_scalar_per_access)>>;
constexpr auto ordered_src_access_lengths =
container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
// make forward steps
const auto src_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(src_desc, forward_step_idx); // loop over space-filling curve
}, constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
Number<nDim>{});
// make backward steps
const auto src_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(src_desc, backward_step_idx);
},
Number<nDim>{});
// loop over tensor and copy // loop over tensor and copy
static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) { static_for<0, num_accesses, 1>{}([&](auto idx_1d) {
// judge move forward or move backward constexpr auto src_data_idx = SpaceFillingCurve::GetIndex(idx_1d);
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_src_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate src data index
constexpr auto src_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
: ordered_src_access_lengths[i] - 1 -
ordered_src_access_idx[i];
});
return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
src_scalar_per_access;
}();
constexpr auto src_data_idx_seq = generate_sequence_v2( constexpr auto src_data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{}); [&](auto i) { return Number<src_data_idx[i]>{}; }, Number<src_data_idx.Size()>{});
...@@ -218,39 +162,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -218,39 +162,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1
.template SetAsType<src_vector_t>( .template SetAsType<src_vector_t>(
src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]); src_data_idx_seq, src_vector_container.template AsType<src_vector_t>()[I0]);
constexpr auto move_on_dim = [&]() constexpr // move coordinate
{ if constexpr(idx_1d.value != num_accesses - 1)
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &=
ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move src coord
static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{ {
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate( move_tensor_coordinate(
src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]); src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
} }
else
{
move_tensor_coordinate(
src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
}
}
});
}); });
// move src coordinate back to slice origin (or not) // move src coordinate back to slice origin (or not)
...@@ -374,73 +292,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -374,73 +292,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto dst_scalar_per_access = generate_sequence( constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DstDimAccessOrder,
constexpr auto dst_dim_access_order = DstDimAccessOrder{}; remove_cv_t<decltype(dst_scalar_per_access)>>;
constexpr auto ordered_dst_access_lengths = constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
// make forward steps
const auto dst_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst_desc, forward_step_idx);
},
Number<nDim>{});
// make backward steps
const auto dst_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(dst_desc, backward_step_idx);
},
Number<nDim>{});
// loop over tensor and copy // loop over tensor and copy
static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) { static_for<0, num_accesses, 1>{}([&](auto idx_1d) {
// judge move forward or move backward constexpr auto dst_data_idx = SpaceFillingCurve::GetIndex(idx_1d);
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_dst_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
: ordered_dst_access_lengths[i] - 1 -
ordered_dst_access_idx[i];
});
return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
dst_scalar_per_access;
}();
constexpr auto dst_data_idx_seq = generate_sequence_v2( constexpr auto dst_data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{}); [&](auto i) { return Number<dst_data_idx[i]>{}; }, Number<dst_data_idx.Size()>{});
...@@ -470,39 +330,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -470,39 +330,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1
is_dst_valid, is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]); dst_vector_container.template AsType<dst_vector_t>()[I0]);
constexpr auto move_on_dim = [&]() constexpr // move coordinate
{ if constexpr(idx_1d.value != num_accesses - 1)
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &=
ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move dst coord
static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
}
else
{ {
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate( move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]); dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
} }
}
});
}); });
// move dst coordinate back to slice origin (or not) // move dst coordinate back to slice origin (or not)
...@@ -522,55 +356,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -522,55 +356,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto src_scalar_per_access = generate_sequence( constexpr auto src_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
SrcDimAccessOrder,
constexpr auto src_dim_access_order = SrcDimAccessOrder{}; remove_cv_t<decltype(src_scalar_per_access)>>;
constexpr auto ordered_src_access_lengths =
container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) { constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
index_t tmp = ordered_src_access_lengths[I0] - 1; constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_accesses - 1>{}, Number<0>{});
static_for<1, i, 1>{}([&](auto j) { return reset_step;
tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate src data index after last iteration in RunRead(), if it has not being reset by
// RunRead()
constexpr auto src_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
src_scalar_per_access;
}();
//
constexpr auto reset_src_data_step = [&]() {
Index reset_src_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
return reset_src_data_step_;
}();
return reset_src_data_step;
} }
__device__ static constexpr auto GetDstCoordinateResetStep() __device__ static constexpr auto GetDstCoordinateResetStep()
...@@ -580,55 +374,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1 ...@@ -580,55 +374,15 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto dst_scalar_per_access = generate_sequence( constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DstDimAccessOrder,
constexpr auto dst_dim_access_order = DstDimAccessOrder{}; remove_cv_t<decltype(dst_scalar_per_access)>>;
constexpr auto ordered_dst_access_lengths =
container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_dst_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index after last iteration in RunWrite(), if it has not being reset by
// RunWrite()
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
dst_scalar_per_access;
}();
//
constexpr auto reset_dst_data_step = [&]() {
Index reset_dst_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
return reset_dst_data_step_; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
}(); constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_accesses - 1>{}, Number<0>{});
return reset_dst_data_step; return reset_step;
} }
// src_slice_origin_step_idx need to be known at compile-time, for performance reason // src_slice_origin_step_idx need to be known at compile-time, for performance reason
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r1.hpp
View file @ c2976d7a
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
#include "common_header.hpp" #include "common_header.hpp"
#include "tensor_descriptor.hpp" #include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp" #include "tensor_descriptor_helper.hpp"
#include "tensor_space_filling_curve.hpp"
namespace ck { namespace ck {
...@@ -40,9 +41,6 @@ struct ThreadwiseTensorSliceTransfer_v6r1 ...@@ -40,9 +41,6 @@ struct ThreadwiseTensorSliceTransfer_v6r1
using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{})); using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{})); using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
__device__ constexpr ThreadwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc, __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
...@@ -79,70 +77,14 @@ struct ThreadwiseTensorSliceTransfer_v6r1 ...@@ -79,70 +77,14 @@ struct ThreadwiseTensorSliceTransfer_v6r1
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
auto make_forward_steps = [&](auto desc) {
return generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(desc, forward_step_idx);
},
Number<nDim>{});
};
auto make_backward_steps = [&](auto desc) {
return generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) { // loop over space-filling curve
backward_step_idx(j) = (i.value == j.value) ? -scalar_per_access[i] : 0; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
});
return make_tensor_coordinate_step(desc, backward_step_idx);
},
Number<nDim>{});
};
// make forward steps
const auto src_forward_steps = make_forward_steps(src_desc);
const auto dst_forward_steps = make_forward_steps(dst_desc);
// make backward steps
const auto src_backward_steps = make_backward_steps(src_desc);
const auto dst_backward_steps = make_backward_steps(dst_desc);
// loop over slice window
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
static_for<0, num_accesses, 1>{}([&](auto idx_1d) {
using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>; using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
using src_vector_t = typename src_vector_type::type; using src_vector_t = typename src_vector_type::type;
...@@ -168,60 +110,21 @@ struct ThreadwiseTensorSliceTransfer_v6r1 ...@@ -168,60 +110,21 @@ struct ThreadwiseTensorSliceTransfer_v6r1
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_); coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
// copy data from dst_vector into dst_buf // copy data from dst_vector into dst_buf
if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set) dst_buf.template Update<DstInMemOp, dst_vector_t>(
{
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(), dst_coord_.GetOffset(),
is_dst_valid, is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]); dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
{
dst_buf.template AtomicAdd<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move coordinate // move coordinate
static_for<0, nDim, 1>{}([&](auto i) { if constexpr(idx_1d.value != num_accesses - 1)
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(
src_desc, src_coord_, src_forward_steps[dim_access_order[i]]);
move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]);
}
else
{ {
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate( move_tensor_coordinate(
src_desc, src_coord_, src_backward_steps[dim_access_order[i]]); src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]); dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
}
} }
}); });
});
// move coordinate back to slice origin (or not) // move coordinate back to slice origin (or not)
if constexpr(SrcResetCoordinateAfterRun) if constexpr(SrcResetCoordinateAfterRun)
...@@ -243,59 +146,18 @@ struct ThreadwiseTensorSliceTransfer_v6r1 ...@@ -243,59 +146,18 @@ struct ThreadwiseTensorSliceTransfer_v6r1
__device__ static constexpr auto GetCoordinateResetStep() __device__ static constexpr auto GetCoordinateResetStep()
{ {
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate data index after last iteration in Run(), if it has not being reset
constexpr auto data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
scalar_per_access;
}();
//
constexpr auto reset_data_step = [&]() {
Index reset_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_data_step_(i) = -data_idx[i]; });
return reset_data_step_; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
}(); constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_accesses - 1>{}, Number<0>{});
return reset_data_step; return reset_step;
} }
// src_slice_origin_step_idx need to be known at compile-time, for performance reason // src_slice_origin_step_idx need to be known at compile-time, for performance reason
...@@ -332,7 +194,7 @@ struct ThreadwiseTensorSliceTransfer_v6r1 ...@@ -332,7 +194,7 @@ struct ThreadwiseTensorSliceTransfer_v6r1
SrcCoord src_coord_; SrcCoord src_coord_;
DstCoord dst_coord_; DstCoord dst_coord_;
const ElementwiseOperation element_op_; const ElementwiseOperation element_op_;
}; }; // namespace ck
} // namespace ck } // namespace ck
#endif #endif
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r2.hpp
View file @ c2976d7a
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
#include "common_header.hpp" #include "common_header.hpp"
#include "tensor_descriptor.hpp" #include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp" #include "tensor_descriptor_helper.hpp"
#include "tensor_space_filling_curve.hpp"
namespace ck { namespace ck {
...@@ -44,10 +45,6 @@ struct ThreadwiseTensorSliceTransfer_v6r2 ...@@ -44,10 +45,6 @@ struct ThreadwiseTensorSliceTransfer_v6r2
using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{})); using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{})); using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using Src0CoordStep = decltype(make_tensor_coordinate_step(Src0Desc{}, Index{}));
using Src1CoordStep = decltype(make_tensor_coordinate_step(Src1Desc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
__device__ constexpr ThreadwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc, __device__ constexpr ThreadwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
...@@ -96,72 +93,14 @@ struct ThreadwiseTensorSliceTransfer_v6r2 ...@@ -96,72 +93,14 @@ struct ThreadwiseTensorSliceTransfer_v6r2
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
auto make_forward_steps = [&](auto desc) { constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
return generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(desc, forward_step_idx);
},
Number<nDim>{});
};
auto make_backward_steps = [&](auto desc) {
return generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(desc, backward_step_idx);
},
Number<nDim>{});
};
// make forward steps
const auto src0_forward_steps = make_forward_steps(src0_desc);
const auto src1_forward_steps = make_forward_steps(src1_desc);
const auto dst_forward_steps = make_forward_steps(dst_desc);
// make backward steps
const auto src0_backward_steps = make_backward_steps(src0_desc);
const auto src1_backward_steps = make_backward_steps(src1_desc);
const auto dst_backward_steps = make_backward_steps(dst_desc);
// loop over slice window
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// loop over space-filling curve
static_for<0, num_accesses, 1>{}([&](auto idx_1d) {
using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>; using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
using src0_vector_t = typename src0_vector_type::type; using src0_vector_t = typename src0_vector_type::type;
...@@ -197,65 +136,22 @@ struct ThreadwiseTensorSliceTransfer_v6r2 ...@@ -197,65 +136,22 @@ struct ThreadwiseTensorSliceTransfer_v6r2
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_); coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
// copy data from dst_vector into dst_buf // copy data from dst_vector into dst_buf
if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set) dst_buf.template Update<DstInMemOp, dst_vector_t>(
{
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
{
dst_buf.template AtomicAdd<dst_vector_t>(
dst_coord_.GetOffset(), dst_coord_.GetOffset(),
is_dst_valid, is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]); dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move coordinate // move coordinate
static_for<0, nDim, 1>{}([&](auto i) { if constexpr(idx_1d.value != num_accesses - 1)
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(
src0_desc, src0_coord_, src0_forward_steps[dim_access_order[i]]);
move_tensor_coordinate(
src1_desc, src1_coord_, src1_forward_steps[dim_access_order[i]]);
move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]);
}
else
{ {
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate( move_tensor_coordinate(
src0_desc, src0_coord_, src0_backward_steps[dim_access_order[i]]); src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
src1_desc, src1_coord_, src1_backward_steps[dim_access_order[i]]); src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]); dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
} }
}
});
}); });
// move coordinate back to slice origin (or not) // move coordinate back to slice origin (or not)
...@@ -286,59 +182,18 @@ struct ThreadwiseTensorSliceTransfer_v6r2 ...@@ -286,59 +182,18 @@ struct ThreadwiseTensorSliceTransfer_v6r2
__device__ static constexpr auto GetCoordinateResetStep() __device__ static constexpr auto GetCoordinateResetStep()
{ {
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate data index after last iteration in Run(), if it has not being reset
constexpr auto data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
scalar_per_access;
}();
//
constexpr auto reset_data_step = [&]() {
Index reset_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_data_step_(i) = -data_idx[i]; });
return reset_data_step_; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
}(); constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_accesses - 1>{}, Number<0>{});
return reset_data_step; return reset_step;
} }
// src_slice_origin_step_idx need to be known at compile-time, for performance reason // src_slice_origin_step_idx need to be known at compile-time, for performance reason
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r3.hpp
View file @ c2976d7a
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
#include "common_header.hpp" #include "common_header.hpp"
#include "tensor_descriptor.hpp" #include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp" #include "tensor_descriptor_helper.hpp"
#include "tensor_space_filling_curve.hpp"
namespace ck { namespace ck {
...@@ -48,11 +49,6 @@ struct ThreadwiseTensorSliceTransfer_v6r3 ...@@ -48,11 +49,6 @@ struct ThreadwiseTensorSliceTransfer_v6r3
using Src2Coord = decltype(make_tensor_coordinate(Src2Desc{}, Index{})); using Src2Coord = decltype(make_tensor_coordinate(Src2Desc{}, Index{}));
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{})); using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using Src0CoordStep = decltype(make_tensor_coordinate_step(Src0Desc{}, Index{}));
using Src1CoordStep = decltype(make_tensor_coordinate_step(Src1Desc{}, Index{}));
using Src2CoordStep = decltype(make_tensor_coordinate_step(Src2Desc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
__device__ constexpr ThreadwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc, __device__ constexpr ThreadwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
...@@ -112,74 +108,14 @@ struct ThreadwiseTensorSliceTransfer_v6r3 ...@@ -112,74 +108,14 @@ struct ThreadwiseTensorSliceTransfer_v6r3
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
auto make_forward_steps = [&](auto desc) {
return generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(desc, forward_step_idx);
},
Number<nDim>{});
};
auto make_backward_steps = [&](auto desc) {
return generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(desc, backward_step_idx);
},
Number<nDim>{});
};
// make forward steps
const auto src0_forward_steps = make_forward_steps(src0_desc);
const auto src1_forward_steps = make_forward_steps(src1_desc);
const auto src2_forward_steps = make_forward_steps(src2_desc);
const auto dst_forward_steps = make_forward_steps(dst_desc);
// make backward steps
const auto src0_backward_steps = make_backward_steps(src0_desc);
const auto src1_backward_steps = make_backward_steps(src1_desc);
const auto src2_backward_steps = make_backward_steps(src2_desc);
const auto dst_backward_steps = make_backward_steps(dst_desc);
// loop over slice window
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
}();
// loop over space-filling curve
static_for<0, num_accesses, 1>{}([&](auto idx_1d) {
using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>; using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
using src0_vector_t = typename src0_vector_type::type; using src0_vector_t = typename src0_vector_type::type;
...@@ -224,73 +160,25 @@ struct ThreadwiseTensorSliceTransfer_v6r3 ...@@ -224,73 +160,25 @@ struct ThreadwiseTensorSliceTransfer_v6r3
const bool is_dst_valid = const bool is_dst_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_); coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
// copy data from dst_vector into dst_buf dst_buf.template Update<DstInMemOp, dst_vector_t>(
if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set)
{
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
{
dst_buf.template AtomicAdd<dst_vector_t>(
dst_coord_.GetOffset(), dst_coord_.GetOffset(),
is_dst_valid, is_dst_valid,
dst_vector_container.template AsType<dst_vector_t>()[I0]); dst_vector_container.template AsType<dst_vector_t>()[I0]);
}
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move coordinate // move coordinate
static_for<0, nDim, 1>{}([&](auto i) { if constexpr(idx_1d.value != num_accesses - 1)
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{ {
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate( move_tensor_coordinate(
src0_desc, src0_coord_, src0_forward_steps[dim_access_order[i]]); src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
src1_desc, src1_coord_, src1_forward_steps[dim_access_order[i]]); src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
src2_desc, src2_coord_, src2_forward_steps[dim_access_order[i]]); src2_desc, src2_coord_, make_tensor_coordinate_step(src2_desc, forward_step));
move_tensor_coordinate( move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]); dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
}
else
{
move_tensor_coordinate(
src0_desc, src0_coord_, src0_backward_steps[dim_access_order[i]]);
move_tensor_coordinate(
src1_desc, src1_coord_, src1_backward_steps[dim_access_order[i]]);
move_tensor_coordinate(
src2_desc, src2_coord_, src2_backward_steps[dim_access_order[i]]);
move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]);
}
} }
}); });
});
// move coordinate back to slice origin (or not) // move coordinate back to slice origin (or not)
if constexpr(Src0ResetCoordinateAfterRun) if constexpr(Src0ResetCoordinateAfterRun)
...@@ -328,59 +216,18 @@ struct ThreadwiseTensorSliceTransfer_v6r3 ...@@ -328,59 +216,18 @@ struct ThreadwiseTensorSliceTransfer_v6r3
__device__ static constexpr auto GetCoordinateResetStep() __device__ static constexpr auto GetCoordinateResetStep()
{ {
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto scalar_per_access = generate_sequence( constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{}); detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / scalar_per_access; using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
DimAccessOrder,
constexpr auto dim_access_order = DimAccessOrder{}; remove_cv_t<decltype(scalar_per_access)>>;
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate data index after last iteration in Run(), if it has not being reset
constexpr auto data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
scalar_per_access;
}();
//
constexpr auto reset_data_step = [&]() {
Index reset_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_data_step_(i) = -data_idx[i]; });
return reset_data_step_; constexpr auto num_accesses = SpaceFillingCurve::GetNumOfAccess();
}(); constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_accesses - 1>{}, Number<0>{});
return reset_data_step; return reset_step;
} }
// src_slice_origin_step_idx need to be known at compile-time, for performance reason // src_slice_origin_step_idx need to be known at compile-time, for performance reason
......
composable_kernel/include/utility/dynamic_buffer.hpp
View file @ c2976d7a
...@@ -3,6 +3,7 @@ ...@@ -3,6 +3,7 @@
#include "amd_buffer_addressing.hpp" #include "amd_buffer_addressing.hpp"
#include "c_style_pointer_cast.hpp" #include "c_style_pointer_cast.hpp"
#include "config.hpp"
#include "enable_if.hpp" #include "enable_if.hpp"
namespace ck { namespace ck {
...@@ -108,6 +109,30 @@ struct DynamicBuffer ...@@ -108,6 +109,30 @@ struct DynamicBuffer
} }
} }
template <InMemoryDataOperationEnum_t Op,
typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value,
bool>::type = false>
__host__ __device__ void Update(index_t i, bool is_valid_element, const X& x)
{
if constexpr(Op == InMemoryDataOperationEnum_t::Set)
{
this->template Set<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum_t::AtomicAdd)
{
this->template AtomicAdd<X>(i, is_valid_element, x);
}
else if constexpr(Op == InMemoryDataOperationEnum_t::Add)
{
auto tmp = this->template Get<X>(i, is_valid_element);
this->template Set<X>(i, is_valid_element, x + tmp);
// tmp += x;
// this->template Set<X>(i, is_valid_element, tmp);
}
}
template <typename X, template <typename X,
typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type, typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
typename scalar_type<remove_cvref_t<T>>::type>::value, typename scalar_type<remove_cvref_t<T>>::type>::value,
......
composable_kernel/include/utility/magic_division.hpp
View file @ c2976d7a
...@@ -25,7 +25,11 @@ struct MagicDivision ...@@ -25,7 +25,11 @@ struct MagicDivision
// uint32_t // uint32_t
__host__ __device__ static constexpr auto CalculateMagicNumbers(uint32_t divisor) __host__ __device__ static constexpr auto CalculateMagicNumbers(uint32_t divisor)
{ {
// assert(divisior >= 1 && divisior <= INT32_MAX); // WARNING: magic division is only applicable for division inside this range.
// You should use the return value of CalculateMagicNumbers, if division is not inside this
// range. The "else" logic below is to quiet down run-time error.
if(divisor >= 1 && divisor <= INT32_MAX)
{
uint32_t shift = 0; uint32_t shift = 0;
for(shift = 0; shift < 32; ++shift) for(shift = 0; shift < 32; ++shift)
{ {
...@@ -41,6 +45,11 @@ struct MagicDivision ...@@ -41,6 +45,11 @@ struct MagicDivision
return make_tuple(uint32_t(multiplier), shift); return make_tuple(uint32_t(multiplier), shift);
} }
else
{
return make_tuple(uint32_t(0), uint32_t(0));
}
}
__host__ __device__ static constexpr uint32_t CalculateMagicMultiplier(uint32_t divisor) __host__ __device__ static constexpr uint32_t CalculateMagicMultiplier(uint32_t divisor)
{ {
......
composable_kernel/include/utility/tensor_space_filling_curve.hpp
View file @ c2976d7a
#ifndef TENSOR_SPACE_FILLING_CURVE_HPP
#define TENSOR_SPACE_FILLING_CURVE_HPP
#include "math.hpp" #include "math.hpp"
#include "sequence.hpp" #include "sequence.hpp"
#include "sequence_helper.hpp"
#include "tensor_adaptor.hpp" #include "tensor_adaptor.hpp"
#include "statically_indexed_array_multi_index.hpp" #include "statically_indexed_array_multi_index.hpp"
#include "tuple_helper.hpp" #include "tuple_helper.hpp"
...@@ -37,13 +41,25 @@ struct SpaceFillingCurve ...@@ -37,13 +41,25 @@ struct SpaceFillingCurve
ScalarPerVector; ScalarPerVector;
} }
template <index_t AccessIdx1dBegin, index_t AccessIdx1dEnd>
static __device__ __host__ constexpr auto GetStepBetween(Number<AccessIdx1dBegin>,
Number<AccessIdx1dEnd>)
{
static_assert(AccessIdx1dBegin >= 0, "1D index should be non-negative");
static_assert(AccessIdx1dBegin < GetNumOfAccess(), "1D index should be larger than 0");
static_assert(AccessIdx1dEnd >= 0, "1D index should be non-negative");
static_assert(AccessIdx1dEnd < GetNumOfAccess(), "1D index should be larger than 0");
constexpr auto idx_begin = GetIndex(Number<AccessIdx1dBegin>{});
constexpr auto idx_end = GetIndex(Number<AccessIdx1dEnd>{});
return idx_end - idx_begin;
}
template <index_t AccessIdx1d> template <index_t AccessIdx1d>
static __device__ __host__ constexpr auto GetForwardStep(Number<AccessIdx1d>) static __device__ __host__ constexpr auto GetForwardStep(Number<AccessIdx1d>)
{ {
static_assert(AccessIdx1d < GetNumOfAccess(), "1D index should be larger than 0");
constexpr auto idx_curr = GetIndex(Number<AccessIdx1d>{}); return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d + 1>{});
constexpr auto idx_next = GetIndex(Number<AccessIdx1d + 1>{});
return idx_next - idx_curr;
} }
template <index_t AccessIdx1d> template <index_t AccessIdx1d>
...@@ -51,9 +67,7 @@ struct SpaceFillingCurve ...@@ -51,9 +67,7 @@ struct SpaceFillingCurve
{ {
static_assert(AccessIdx1d > 0, "1D index should be larger than 0"); static_assert(AccessIdx1d > 0, "1D index should be larger than 0");
constexpr auto idx_curr = GetIndex(Number<AccessIdx1d>{}); return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d - 1>{});
constexpr auto idx_prev = GetIndex(Number<AccessIdx1d - 1>{});
return idx_prev - idx_curr;
} }
template <index_t AccessIdx1d> template <index_t AccessIdx1d>
...@@ -129,3 +143,4 @@ struct SpaceFillingCurve ...@@ -129,3 +143,4 @@ struct SpaceFillingCurve
}; };
} // namespace ck } // namespace ck
#endif
device_operation/CMakeLists.txt
View file @ c2976d7a
...@@ -101,8 +101,15 @@ set(DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE ...@@ -101,8 +101,15 @@ set(DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_atomic_add_nhwc_kyxc_nhwk_f16_instance.cpp; ${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_fwd_xdl_c_shuffle_bias_relu_atomic_add_nhwc_kyxc_nhwk_f16_instance.cpp;
) )
# device_conv2d_bwd_data_instance
set(DEVICE_CONV2D_BWD_DATA_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp;
${PROJECT_SOURCE_DIR}/device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp;
)
add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE}) add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE})
add_library(device_gemm_bias_2d_instance SHARED ${DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE}) add_library(device_gemm_bias_relu_instance SHARED ${DEVICE_GEMM_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE}) add_library(device_gemm_bias_relu_add_instance SHARED ${DEVICE_GEMM_BIAS_RELU_ADD_INSTANCE_SOURCE})
add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE}) add_library(device_batched_gemm_instance SHARED ${DEVICE_BATCHED_GEMM_INSTANCE_SOURCE})
...@@ -111,6 +118,8 @@ add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURC ...@@ -111,6 +118,8 @@ add_library(device_conv2d_fwd_instance SHARED ${DEVICE_CONV2D_FWD_INSTANCE_SOURC
add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE}) add_library(device_conv2d_fwd_bias_relu_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE}) add_library(device_conv2d_fwd_bias_relu_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ADD_INSTANCE_SOURCE})
add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE}) add_library(device_conv2d_fwd_bias_relu_atomic_add_instance SHARED ${DEVICE_CONV2D_FWD_BIAS_RELU_ATOMIC_ADD_INSTANCE_SOURCE})
add_library(device_gemm_bias_2d_instance SHARED ${DEVICE_GEMM_BIAS_2D_INSTANCE_SOURCE})
add_library(device_conv2d_bwd_data_instance SHARED ${DEVICE_CONV2D_BWD_DATA_INSTANCE_SOURCE})
target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>) target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_gemm_bias_2d_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>) target_include_directories(device_gemm_bias_2d_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
...@@ -122,6 +131,7 @@ target_include_directories(device_conv2d_fwd_instance SYSTEM PUBLIC $<BUILD_INTE ...@@ -122,6 +131,7 @@ target_include_directories(device_conv2d_fwd_instance SYSTEM PUBLIC $<BUILD_INTE
target_include_directories(device_conv2d_fwd_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>) target_include_directories(device_conv2d_fwd_bias_relu_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>) target_include_directories(device_conv2d_fwd_bias_relu_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_fwd_bias_relu_atomic_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>) target_include_directories(device_conv2d_fwd_bias_relu_atomic_add_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_include_directories(device_conv2d_bwd_data_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_compile_features(device_gemm_instance PUBLIC) target_compile_features(device_gemm_instance PUBLIC)
target_compile_features(device_gemm_bias_2d_instance PUBLIC) target_compile_features(device_gemm_bias_2d_instance PUBLIC)
...@@ -133,6 +143,7 @@ target_compile_features(device_conv2d_fwd_instance PUBLIC) ...@@ -133,6 +143,7 @@ target_compile_features(device_conv2d_fwd_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_instance PUBLIC) target_compile_features(device_conv2d_fwd_bias_relu_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_add_instance PUBLIC) target_compile_features(device_conv2d_fwd_bias_relu_add_instance PUBLIC)
target_compile_features(device_conv2d_fwd_bias_relu_atomic_add_instance PUBLIC) target_compile_features(device_conv2d_fwd_bias_relu_atomic_add_instance PUBLIC)
target_compile_features(device_conv2d_bwd_data_instance PUBLIC)
set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_gemm_bias_2d_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_gemm_bias_2d_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
...@@ -144,6 +155,7 @@ set_target_properties(device_conv2d_fwd_instance PROPERTIES POSITION_INDEPENDENT ...@@ -144,6 +155,7 @@ set_target_properties(device_conv2d_fwd_instance PROPERTIES POSITION_INDEPENDENT
set_target_properties(device_conv2d_fwd_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_conv2d_fwd_bias_relu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_conv2d_fwd_bias_relu_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_fwd_bias_relu_atomic_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_conv2d_fwd_bias_relu_atomic_add_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
set_target_properties(device_conv2d_bwd_data_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_gemm_instance LIBRARY DESTINATION lib) install(TARGETS device_gemm_instance LIBRARY DESTINATION lib)
install(TARGETS device_gemm_bias_2d_instance LIBRARY DESTINATION lib) install(TARGETS device_gemm_bias_2d_instance LIBRARY DESTINATION lib)
...@@ -155,3 +167,4 @@ install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib) ...@@ -155,3 +167,4 @@ install(TARGETS device_conv2d_fwd_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib) install(TARGETS device_conv2d_fwd_bias_relu_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib) install(TARGETS device_conv2d_fwd_bias_relu_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_fwd_bias_relu_atomic_add_instance LIBRARY DESTINATION lib) install(TARGETS device_conv2d_fwd_bias_relu_atomic_add_instance LIBRARY DESTINATION lib)
install(TARGETS device_conv2d_bwd_data_instance LIBRARY DESTINATION lib)
device_operation/include/convolution_backward_data_specialization.hpp 0 → 100644
View file @ c2976d7a
#ifndef CONVOLUTION_BACKWARD_DATA_SPECIALIZATION
#define CONVOLUTION_BACKWARD_DATA_SPECIALIZATION
namespace ck {
namespace tensor_operation {
namespace device {
enum ConvolutionBackwardDataSpecialization_t
{
Default,
Filter1x1Stride1Pad0,
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
device_operation/include/device_conv_bwd_data.hpp 0 → 100644
View file @ c2976d7a
#ifndef DEVICE_CONV_BWD_DATA_HPP
#define DEVICE_CONV_BWD_DATA_HPP
#include <iostream>
#include "device_base.hpp"
#include "element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation>
struct DeviceConvBwdData : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(void* p_in,
const void* p_wei,
const void* p_out,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation>
using DeviceConvBwdDataPtr = std::unique_ptr<
DeviceConvBwdData<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp 0 → 100644
View file @ c2976d7a
#include <stdlib.h>
#include "config.hpp"
#include "device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using BF16 = ushort;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>
// clang-format on
>;
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances =
std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>
// clang-format on
>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances{});
add_device_operation_instances(
instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
}
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp 0 → 100644
View file @ c2976d7a
#include <stdlib.h>
#include "config.hpp"
#include "device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
#if !CK_WORKAROUND_SWDEV_325164
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
#endif
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>
// clang-format on
>;
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances =
std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 8, true, 7, 1>
// clang-format on
>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances{});
add_device_operation_instances(
instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
}
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp 0 → 100644
View file @ c2976d7a
#include <stdlib.h>
#include "config.hpp"
#include "device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>
// clang-format on
>;
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances =
std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 4, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 4, true, 7, 1>
// clang-format on
>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances{});
add_device_operation_instances(
instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
}
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
device_operation/src/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp 0 → 100644
View file @ c2976d7a
#include <stdlib.h>
#include "config.hpp"
#include "device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using DataType = int8_t;
using AccType = int32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Default;
static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Filter1x1Stride1Pad0;
// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
// clang-format off
//####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 256, 4, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 128, 4, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 64, 128, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 64, 4, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 128, 64, 4, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 256, 64, 128, 4, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 128, 32, 4, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 128, 32, 128, 4, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 64, 32, 4, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault, 64, 32, 64, 4, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>
// clang-format on
>;
using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances =
std::tuple<
// clang-format off
//#####################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//#####################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Data| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcDstVectorDim| DstScalar|
//#####################################################################| | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//#####################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 256, 4, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 128, 4, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 128, 4, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 64, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 64, 128, 4, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 64, 4, 16, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 128, 64, 4, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 256, 64, 128, 4, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 64, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 128, 32, 4, 16, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 1, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 128, 32, 128, 4, 16, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 32, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 64, 32, 4, 16, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 2, 16, true, 7, 1>,
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< DataType, DataType, DataType, AccType, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0, 64, 32, 64, 4, 16, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, true, S<4, 16, 1>, S<2, 0, 1>, S<0, 2, 1>, 1, 4, 16, true, 7, 1>
// clang-format on
>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances,
device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances{});
add_device_operation_instances(
instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
}
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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