wouter/d4r
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Execute native Windows K DLSS and validate real transformer weights

Browse Source
This commit is contained in:
Codex committed 2026-09-30 18:23:01 +03:00
1 parent a0c36a6d6f
commit 5592b7a921
20 files changed
+948 -54
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

kernels/k/pwin_pos.h
+1 -12
View File
@@ -19,18 +19,7 @@ __device__ void pos_core(const PwinParams& p, const u8v* __restrict__ img, const
{
const half_t* arow = act[16 * (wv * MT + mi) + m];
half_t* hrow = hb[16 * (wv * MT + mi) + m];
float ss = 0.0f;
#pragma unroll
for (int kt = 0; kt < KT; ++kt)
{
const u8v xv = lds_row16(arow + 16 * kt);
#pragma unroll
for (int c = 0; c < 16; ++c)
{
const half_t v = op_get(xv, c);
ss += (float)(half_t)(v * v);
}
}
const half_t ss = l2_sum<C>(arow);
const half_t r16 = (half_t)__builtin_amdgcn_rsqf((float)(half_t)ss);
for (int kt = (int)hf; kt < KT; kt += 2)
{