v3.9 (#2185)
* v3.8 update x * fix blackwell gg * doc change * doc change * doc change --------- Co-authored-by: yuzhai <yuzhai@nvidia.com> Co-authored-by: Haicheng Wu <haichengw@nvidia.com> Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
This commit is contained in:
Yujia Zhai
co-authored by
yuzhai
Haicheng Wu
Haicheng Wu
yuzhai
Haicheng Wu
Haicheng Wu
parent
8c4d1dc47d
commit
62750a2b75
@@ -275,7 +275,7 @@ private:
|
||||
};
|
||||
|
||||
// In the mainloop, PRMT selects 1 byte from only 8 bytes so the sign bit is handled in an extra PRMT.
|
||||
// Here the encodings of positive values and negative values are unified (except for the sign bit).
|
||||
// Here the encodings of positive values and negative values are unified (except for the sign bit).
|
||||
// For instance, 1 becomes 0b0111, which is the same encoding as -1 (0b1111).
|
||||
static bool unified_encode_int4b(cutlass::int4b_t const *block_in, cutlass::int4b_t *block_out, const size_t block_size) {
|
||||
|
||||
@@ -430,7 +430,7 @@ void reorder_tensor(
|
||||
};
|
||||
static_assert(has_major_mode(stride<0>(LayoutDst{})) ^ has_major_mode(stride<1>(LayoutDst{})),
|
||||
"Could not find stride-1 mode in destination layout");
|
||||
constexpr int N = shape_div(Int<8>{}, sizeof_bits<T>{});
|
||||
constexpr int N = shape_div(Int<8>{}, Int<sizeof_bits_v<T>>{});
|
||||
auto val_layout = conditional_return<has_major_mode(stride<0>(LayoutDst{}))>(
|
||||
make_layout(make_shape(Int<N>{}, Int<1>{}), GenColMajor{}),
|
||||
make_layout(make_shape(Int<1>{}, Int<N>{}), GenRowMajor{}));
|
||||
|
||||
@@ -55,6 +55,124 @@ namespace host {
|
||||
|
||||
namespace detail {
|
||||
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
struct TensorGreatestErrorFunc {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
TensorView<Element, Layout> lhs;
|
||||
TensorView<Element, Layout> rhs;
|
||||
double result;
|
||||
|
||||
/// Ctor
|
||||
TensorGreatestErrorFunc(
|
||||
TensorView<Element, Layout> const &lhs_,
|
||||
TensorView<Element, Layout> const &rhs_
|
||||
) :
|
||||
lhs(lhs_),
|
||||
rhs(rhs_),
|
||||
result(0.0) { }
|
||||
|
||||
/// Visits a coordinate
|
||||
void operator()(Coord<Layout::kRank> const &coord) {
|
||||
|
||||
Element lhs_ = lhs.at(coord);
|
||||
Element rhs_ = rhs.at(coord);
|
||||
|
||||
result = std::max(result, std::abs(double(lhs_) - double(rhs_)));
|
||||
}
|
||||
|
||||
/// Returns true if equal
|
||||
operator double() const {
|
||||
return result;
|
||||
}
|
||||
};
|
||||
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
struct TensorMREFunc {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
TensorView<Element, Layout> lhs;
|
||||
TensorView<Element, Layout> rhs;
|
||||
double sum;
|
||||
uint64_t count;
|
||||
static constexpr double epsilon = 1e-6;
|
||||
|
||||
/// Ctor
|
||||
TensorMREFunc(
|
||||
TensorView<Element, Layout> const &lhs_,
|
||||
TensorView<Element, Layout> const &rhs_
|
||||
) :
|
||||
lhs(lhs_),
|
||||
rhs(rhs_),
|
||||
sum(0.0),
|
||||
count(0) { }
|
||||
|
||||
/// Visits a coordinate
|
||||
void operator()(Coord<Layout::kRank> const &coord) {
|
||||
|
||||
Element lhs_ = lhs.at(coord);
|
||||
Element rhs_ = rhs.at(coord);
|
||||
|
||||
sum += std::abs(double(lhs_) - double(rhs_) / (double(rhs_) + epsilon));
|
||||
++count;
|
||||
}
|
||||
|
||||
/// Returns true if equal
|
||||
operator double() const {
|
||||
return sum / double(count);
|
||||
}
|
||||
};
|
||||
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
struct TensorMSEFunc {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
TensorView<Element, Layout> lhs;
|
||||
TensorView<Element, Layout> rhs;
|
||||
double sum;
|
||||
uint64_t count;
|
||||
|
||||
/// Ctor
|
||||
TensorMSEFunc(
|
||||
TensorView<Element, Layout> const &lhs_,
|
||||
TensorView<Element, Layout> const &rhs_
|
||||
) :
|
||||
lhs(lhs_),
|
||||
rhs(rhs_),
|
||||
sum(0.0),
|
||||
count(0) { }
|
||||
|
||||
/// Visits a coordinate
|
||||
void operator()(Coord<Layout::kRank> const &coord) {
|
||||
|
||||
Element lhs_ = lhs.at(coord);
|
||||
Element rhs_ = rhs.at(coord);
|
||||
|
||||
sum += std::pow((double(lhs_) - double(rhs_)), 2);
|
||||
++count;
|
||||
}
|
||||
|
||||
/// Returns true if equal
|
||||
operator double() const {
|
||||
return sum / double(count);
|
||||
}
|
||||
};
|
||||
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
@@ -144,6 +262,81 @@ struct TensorRelativelyEqualsFunc {
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Returns the Mean Squared Error between two tensors.
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
double TensorMSE(
|
||||
TensorView<Element, Layout> const &lhs,
|
||||
TensorView<Element, Layout> const &rhs) {
|
||||
|
||||
// Extents must be identical
|
||||
if (lhs.extent() != rhs.extent()) {
|
||||
return -1;
|
||||
}
|
||||
|
||||
detail::TensorMSEFunc<Element, Layout> func(lhs, rhs);
|
||||
TensorForEach(
|
||||
lhs.extent(),
|
||||
func
|
||||
);
|
||||
|
||||
return double(func);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Returns the Mean Relative Error between two tensors.
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
double TensorMRE(
|
||||
TensorView<Element, Layout> const &lhs,
|
||||
TensorView<Element, Layout> const &rhs) {
|
||||
|
||||
// Extents must be identical
|
||||
if (lhs.extent() != rhs.extent()) {
|
||||
return -1;
|
||||
}
|
||||
|
||||
detail::TensorMREFunc<Element, Layout> func(lhs, rhs);
|
||||
TensorForEach(
|
||||
lhs.extent(),
|
||||
func
|
||||
);
|
||||
|
||||
return double(func);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Returns the greatest error between two tensors.
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
typename Layout> ///< Layout function
|
||||
double TensorGreatestError(
|
||||
TensorView<Element, Layout> const &lhs,
|
||||
TensorView<Element, Layout> const &rhs) {
|
||||
|
||||
// Extents must be identical
|
||||
if (lhs.extent() != rhs.extent()) {
|
||||
return -1;
|
||||
}
|
||||
|
||||
detail::TensorGreatestErrorFunc<Element, Layout> func(lhs, rhs);
|
||||
TensorForEach(
|
||||
lhs.extent(),
|
||||
func
|
||||
);
|
||||
|
||||
return double(func);
|
||||
}
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
///////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Returns true if two tensor views are equal.
|
||||
template <
|
||||
typename Element, ///< Element type
|
||||
|
||||
Reference in New Issue
Block a user