30 lines
705 B
C++
30 lines
705 B
C++
// TODO: merge this file with `math.cuh` (the device part)
|
|
#pragma once
|
|
|
|
#include <torch/python.h>
|
|
|
|
#include "exception.hpp"
|
|
|
|
namespace deep_gemm {
|
|
|
|
// TODO: use `torch::kFloat4_e2m1fn_x2`
|
|
constexpr auto kPackedFP4 = torch::kInt8;
|
|
|
|
template <typename T>
|
|
static T ceil_div(const T& a, const T& b) {
|
|
return (a + b - 1) / b;
|
|
}
|
|
|
|
template <typename T>
|
|
static constexpr T align(const T& a, const T& b) {
|
|
return ceil_div(a, b) * b;
|
|
}
|
|
|
|
static int get_tma_aligned_size(const int& x, const int& element_size) {
|
|
constexpr int kNumTMAAlignmentBytes = 16;
|
|
DG_HOST_ASSERT(kNumTMAAlignmentBytes % element_size == 0);
|
|
return align(x, kNumTMAAlignmentBytes / element_size);
|
|
}
|
|
|
|
} // namespace deep_gemm
|