Round 20 Track C Audit: gdn_pre Key-Head t >= 3 Vectorized conv_w Fast Path + gdn_scan_r 32-Bit Base Offset Hoisting & Final-Token Dead-Lookahead Elimination #

1. Summary of Changes (metal/kernels/gdn_prompt.metal) #

1. gp_conv & gdn_pre Key-Head t >= 3 Fast Path (lines 26–105):

   • In gp_conv (t < 3 fallback), replaced 4 scalar cw[ch * 4 + k] loads with a single 128-bit float4 w = *(device const float4*)(cw + ch * 4); load and unrolled the 4-tap FMA loop.
   • In gdn_pre's key-head branch (if (x < k_tgs)), added a t >= 3 branchless fast path preserving the exact lane-to-channel mapping d = int(lane) + m * 32 (m = 0..3).
   • Hoisted pq = qkv + (t - 3) * C + kh * S + int(lane) and pk = qkv + (t - 3) * C + (hk_n + kh) * S + int(lane), loaded 128-bit float4 wq and wk from conv_w, and evaluated the 4-tap FMA chain at offsets 0, C, 2*C, 3*C without src >= 0 ternary branches.

2. gdn_scan_r State Offset Hoisting, 32-Bit Base Offsets, and Final-Token Dead-Lookahead Elimination (lines 261–351, 421–423):

   • Hardware Register Discovery (64-bit VGPR vs 32-bit base offset): Holding 64-bit device pointers (sp, kn_p, qn_p, gb_p, vc_p, kq_p, o_p) live across the threadgroup_barrier inside the token scan loop consumes 2 VGPRs per pointer per lane on Apple Silicon, crossing an occupancy cliff (192.2 ms -> 197.8 ms).
   • Instead, hoisted 32-bit integer strides and base offsets (const int rs = hv * S, s0 = (i0 * hv + h) * S + j;, const int ks = hkn * S, k0 = hk * S + i0 + int(lr), v0 = h * S + j;), keeping device base pointers in uniform argument registers and indexing (state + s0)[i * rs], kn[tn * ks + k0 + c * 32], qn[tn * ks + k0 + c * 32], vc[tn * rs + v0], and o[t * rs + v0].
   • Split the TB == 0 loop into for (int t = 0; t < T - 1; ++t) (with tn = t + 1, eliminating min(t + 1, T - 1)) followed by a if (T > 0) final-token epilogue for t = T - 1 that skips the dead kn_l, qn_l, g_b_n, v_n global loads and skips the 64 dead simd_broadcasts (knr, qnr) and 64 dead ka/qa FMAs.

2. Bitwise Parity Verification (memcmp == 0) #

Verified via /tmp/argon_r20_gdn/strata/build-metal/gdn_prompt_parity (standalone suite + --ab bitwise comparison against baseline gdn_prompt.metal):

• All gdn_pre outputs (qn, kn, vc, gb, kq, history), gdn_scanr_4 outputs (o, state), and gdn_post outputs (yh, y) pass with 100% bitwise memcmp == 0 across T = 1, 2, 3, 4, 64, 512, 1024, 2048.

3. Paired A/B Timings (Apple M5, 36 GDN Layers, Quiet GPU) #

4. Artifacts #

• Clean unified diff: /tmp/argon_r20_gdn/argon_r20_gdn.patch (verified with git -C vendor/strata apply --check).
• Isolated build & test binaries: /tmp/argon_r20_gdn/strata/build-metal/gdn_prompt_parity and /tmp/argon_r20_gdn/strata/build-metal/strata_window_cli.