Execute native Windows K DLSS and validate real transformer weights

kernels/k/pwin_layer.h

@@ -259,17 +259,7 @@ __device__ void pwin_core(const PwinParams& p, const u8v* __restrict__ img, cons
 
     STAMP(1);
     // ---- L2 norm: h1 = x0 * (rsqrt(sum x0^2) * gamma1) -> hb
-    float ss = 0.0f;
-    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)
     {