Skip to content

GitLab

"...resnet50_tensorflow.git" did not exist on "a55cf4d30d5a61ca94a933d725e79011a2e83693"
Commit 2dea900b authored by Chao Liu's avatar Chao Liu
Browse files

fixed GetSrcCoordinateResetStep and GetDstCoordinateResetStep in v1r3 and v3

parent 6fe9682a
Show whitespace changes
Inline Side-by-side
Showing with 183 additions and 88 deletions
composable_kernel/include/tensor_operation/threadwise_dynamic_tensor_slice_transfer.hpp
View file @ 2dea900b
...@@ -80,7 +80,8 @@ struct ThreadwiseDynamicTensorSliceTransfer_v1r3 ...@@ -80,7 +80,8 @@ struct ThreadwiseDynamicTensorSliceTransfer_v1r3
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{}; constexpr auto I1 = Number<1>{};
// TODO: don't use this // scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence( constexpr auto dst_scalar_per_access = generate_sequence(
lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{}); lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
...@@ -260,31 +261,64 @@ struct ThreadwiseDynamicTensorSliceTransfer_v1r3 ...@@ -260,31 +261,64 @@ struct ThreadwiseDynamicTensorSliceTransfer_v1r3
__device__ static constexpr auto GetDstCoordinateResetStep() __device__ static constexpr auto GetDstCoordinateResetStep()
{ {
constexpr auto dst_scalar_per_access = [&]() { constexpr auto I0 = Number<0>{};
Index dst_scalar_per_access;
static_for<0, nDim, 1>{}([&](auto i) { // scalar per access on each dim
dst_scalar_per_access(i) = (i == DstVectorDim) ? DstScalarPerVector : 1; // TODO: don't use lambda_scalar_per_access
}); constexpr auto dst_scalar_per_access = generate_sequence(
lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
return dst_scalar_per_access; constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
}();
MultiIndex<nDim> dst_reset_iterator; constexpr auto dim_access_order = DimAccessOrder{};
// TODO: this is wrong, need to consider DimAccessOrder constexpr auto ordered_access_lengths =
dst_reset_iterator(Number<0>{}) = dst_scalar_per_access[Number<0>{}] - SliceLengths{}[0]; 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) { static_for<1, nDim, 1>{}([&](auto i) {
constexpr auto i_m1 = i - Number<1>{}; index_t tmp = ordered_access_lengths[I0] - 1;
static_for<0, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
// TODO: this is wrong forward_sweep(i) = tmp % 2 == 0;
dst_reset_iterator(i) = (SliceLengths{}[i_m1] % (2 * dst_scalar_per_access[i_m1]) == 0)
? 0
: (dst_scalar_per_access[i] - SliceLengths{}[i]);
}); });
return dst_reset_iterator; 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_access_lengths[i] - 1 : 0;
});
auto dst_data_idx = container_reorder_given_old2new(ordered_idx, dim_access_order) *
dst_scalar_per_access;
return dst_data_idx;
}();
//
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;
}();
return reset_dst_data_step;
} }
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
...@@ -385,19 +419,20 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -385,19 +419,20 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{}; constexpr auto I1 = Number<1>{};
// TODO: don't use this // scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto src_scalar_per_access = generate_sequence( constexpr auto src_scalar_per_access = generate_sequence(
lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{}); lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
constexpr auto src_scalar_step_in_vector = constexpr auto src_scalar_step_in_vector =
generate_sequence(lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{}); generate_sequence(lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / src_scalar_per_access; constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
constexpr auto src_dim_access_order = SrcDimAccessOrder{}; constexpr auto src_dim_access_order = SrcDimAccessOrder{};
constexpr auto ordered_access_lengths = constexpr auto ordered_src_access_lengths =
container_reorder_given_new2old(access_lengths, src_dim_access_order); container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
// make forward iterators // make forward iterators
const auto src_forward_iterators = generate_tuple( const auto src_forward_iterators = generate_tuple(
...@@ -428,7 +463,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -428,7 +463,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
Number<nDim>{}); Number<nDim>{});
// loop over tensor and copy // loop over tensor and copy
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) { static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
// judge move forward or move backward // judge move forward or move backward
constexpr auto forward_sweep = [&]() { constexpr auto forward_sweep = [&]() {
...@@ -437,10 +472,10 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -437,10 +472,10 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
forward_sweep(I0) = true; forward_sweep(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) { static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0]; index_t tmp = ordered_src_access_idx[I0];
static_for<0, i, 1>{}([&](auto j) { static_for<0, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j]; tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
}); });
forward_sweep(i) = tmp % 2 == 0; forward_sweep(i) = tmp % 2 == 0;
...@@ -450,19 +485,20 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -450,19 +485,20 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
}(); }();
// calculate src data index // calculate src data index
constexpr auto data_idx = [&]() { constexpr auto src_data_idx = [&]() {
Index ordered_idx; Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) { static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
? ordered_access_idx[i] : ordered_src_access_lengths[i] - 1 -
: ordered_access_lengths[i] - 1 - ordered_access_idx[i]; ordered_src_access_idx[i];
}); });
auto data_idx = container_reorder_given_old2new(ordered_idx, src_dim_access_order) * auto src_data_idx =
container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
src_scalar_per_access; src_scalar_per_access;
return data_idx; return src_data_idx;
}(); }();
// copy data // copy data
...@@ -486,7 +522,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -486,7 +522,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) { static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t buffer_offset = constexpr index_t buffer_offset =
buffer_desc_.CalculateOffset(data_idx + i * src_scalar_step_in_vector); buffer_desc_.CalculateOffset(src_data_idx + i * src_scalar_step_in_vector);
buffer_(Number<buffer_offset>{}) = src_vector[i]; buffer_(Number<buffer_offset>{}) = src_vector[i];
}); });
...@@ -502,7 +538,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -502,7 +538,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) { static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
constexpr index_t buffer_offset = constexpr index_t buffer_offset =
buffer_desc_.CalculateOffset(data_idx + i * src_scalar_step_in_vector); buffer_desc_.CalculateOffset(src_data_idx + i * src_scalar_step_in_vector);
buffer_(Number<buffer_offset>{}) = src_vector[i]; buffer_(Number<buffer_offset>{}) = src_vector[i];
}); });
...@@ -513,10 +549,11 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -513,10 +549,11 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
StaticallyIndexedArray<bool, nDim> move_on_dim; StaticallyIndexedArray<bool, nDim> move_on_dim;
static_for<0, nDim, 1>{}([&](auto i) { static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1; move_on_dim(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) { static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1; move_on_dim(i) &=
ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
}); });
}); });
...@@ -563,6 +600,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -563,6 +600,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
constexpr auto I0 = Number<0>{}; constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{}; constexpr auto I1 = Number<1>{};
// src scalar per access on each dim
// TODO: don't use this // TODO: don't use this
constexpr auto dst_scalar_per_access = generate_sequence( constexpr auto dst_scalar_per_access = generate_sequence(
lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{}); lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
...@@ -570,12 +608,12 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -570,12 +608,12 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
constexpr auto dst_scalar_step_in_vector = constexpr auto dst_scalar_step_in_vector =
generate_sequence(lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{}); generate_sequence(lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access; constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dst_dim_access_order = DstDimAccessOrder{}; constexpr auto dst_dim_access_order = DstDimAccessOrder{};
constexpr auto ordered_access_lengths = constexpr auto ordered_dst_access_lengths =
container_reorder_given_new2old(access_lengths, dst_dim_access_order); container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
// make forward iterators // make forward iterators
const auto dst_forward_iterators = generate_tuple( const auto dst_forward_iterators = generate_tuple(
...@@ -610,7 +648,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -610,7 +648,7 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
Number<nDim>{}); Number<nDim>{});
// loop over tensor and copy // loop over tensor and copy
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) { static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
// judge move forward or move backward // judge move forward or move backward
constexpr auto forward_sweep = [&]() { constexpr auto forward_sweep = [&]() {
...@@ -619,10 +657,10 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -619,10 +657,10 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
forward_sweep(I0) = true; forward_sweep(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) { static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0]; index_t tmp = ordered_dst_access_idx[I0];
static_for<0, i, 1>{}([&](auto j) { static_for<0, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j]; tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
}); });
forward_sweep(i) = tmp % 2 == 0; forward_sweep(i) = tmp % 2 == 0;
...@@ -636,9 +674,9 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -636,9 +674,9 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
Index ordered_idx; Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) { static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
? ordered_access_idx[i] : ordered_dst_access_lengths[i] - 1 -
: ordered_access_lengths[i] - 1 - ordered_access_idx[i]; ordered_dst_access_idx[i];
}); });
auto dst_data_idx = auto dst_data_idx =
...@@ -674,10 +712,11 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -674,10 +712,11 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
StaticallyIndexedArray<bool, nDim> move_on_dim; StaticallyIndexedArray<bool, nDim> move_on_dim;
static_for<0, nDim, 1>{}([&](auto i) { static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1; move_on_dim(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) { static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1; move_on_dim(i) &=
ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
}); });
}); });
...@@ -745,70 +784,126 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3 ...@@ -745,70 +784,126 @@ struct ThreadwiseDynamicTensorSliceTransfer_v3
__device__ static constexpr auto GetSrcCoordinateResetStep() __device__ static constexpr auto GetSrcCoordinateResetStep()
{ {
constexpr auto src_scalar_per_access = [&]() { constexpr auto I0 = Number<0>{};
Index src_scalar_per_access;
static_for<0, nDim, 1>{}([&](auto i) { // scalar per access on each dim
if constexpr(i == SrcVectorDim) // TODO: don't use lambda_scalar_per_access
{ constexpr auto src_scalar_per_access = generate_sequence(
src_scalar_per_access(i) = SrcScalarPerVector; lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
}
else
{
src_scalar_per_access(i) = 1;
}
});
return src_scalar_per_access; constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
}();
MultiIndex<nDim> src_reset_iterator; constexpr auto src_dim_access_order = SrcDimAccessOrder{};
src_reset_iterator(Number<0>{}) = src_scalar_per_access[Number<0>{}] - SliceLengths{}[0]; 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) { static_for<1, nDim, 1>{}([&](auto i) {
constexpr auto i_m1 = i - Number<1>{}; index_t tmp = ordered_src_access_lengths[I0] - 1;
src_reset_iterator(i) = (SliceLengths{}[i_m1] % (2 * src_scalar_per_access[i_m1]) == 0) static_for<0, i, 1>{}([&](auto j) {
? 0 tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
: (src_scalar_per_access[i] - SliceLengths{}[i]);
}); });
return src_reset_iterator; forward_sweep(i) = tmp % 2 == 0;
} });
__device__ static constexpr auto GetDstCoordinateResetStep() return forward_sweep;
{ }();
constexpr auto dst_scalar_per_access = [&]() {
Index dst_scalar_per_access; // 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) { static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(i == DstVectorDim) ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
{
dst_scalar_per_access(i) = DstScalarPerVector;
}
else
{
dst_scalar_per_access(i) = 1;
}
}); });
return dst_scalar_per_access; auto src_data_idx = container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
src_scalar_per_access;
return src_data_idx;
}(); }();
MultiIndex<nDim> dst_reset_iterator; //
constexpr auto reset_src_data_step = [&]() {
Index reset_src_data_step;
dst_reset_iterator(Number<0>{}) = dst_scalar_per_access[Number<0>{}] - SliceLengths{}[0]; 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()
{
constexpr auto I0 = Number<0>{};
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence(
lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dst_dim_access_order = DstDimAccessOrder{};
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) { static_for<1, nDim, 1>{}([&](auto i) {
constexpr auto i_m1 = i - Number<1>{}; index_t tmp = ordered_dst_access_lengths[I0] - 1;
static_for<0, i, 1>{}([&](auto j) {
tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
});
dst_reset_iterator(i) = (SliceLengths{}[i_m1] % (2 * dst_scalar_per_access[i_m1]) == 0) forward_sweep(i) = tmp % 2 == 0;
? 0
: (dst_scalar_per_access[i] - SliceLengths{}[i]);
}); });
return dst_reset_iterator; 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;
});
auto dst_data_idx = container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
dst_scalar_per_access;
return dst_data_idx;
}();
//
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;
}();
return reset_dst_data_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
......
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