diff --git a/clean/dx12/kernels/l_post.hlsl b/clean/dx12/kernels/l_post.hlsl index e12a8e07..652304fa 100644 --- a/clean/dx12/kernels/l_post.hlsl +++ b/clean/dx12/kernels/l_post.hlsl @@ -4,7 +4,8 @@ // ceil(dw / 32) * ceil(ceil(dh / 2) / 8) workgroups of 256 lanes; lane (l, k) of wave k handles D column x0 + l and the // row pair k0 + k. w = the post blob (w[0] = s). Model L's changes against post_recon: the 4 x 4 colour gather in the // model's precision (FP16 features, offsets, exponent and accumulators), linear-colour history candidates -// L = D(h) / e_prev combined as fma(w1, L1, w2 L2) e, and the render pixel's colour as the fallback. +// L = D(h) r_prev (r_prev = 1 / e_prev) combined as fma(w1, L1, w2 L2) e, and the render pixel's colour as the +// fallback. // // The operations follow HIP's ISA of l_post_recon (llvm-objdump of l_post.co), not only its source: // - E_in(e c) and the B / A codes k / 255 (an IEEE division: wg_div) are rounded to FP16 by v_cvt_f16_f32 (wg_f2h); @@ -17,7 +18,8 @@ // - the tap weight w = 2^eps (v_exp_f32) is flushed to +0 below 2^-126 and rounded to FP16 by v_cvt_f16_f32 (wg_f2h); // - the full history's texel is (_Float16)(kEncK log2(..)) in one rounding (v_fma_mixlo_f16: wg_mul_f16); the display // value is clamp(E(o), 0, 2.45) rounded by v_cvt_f16_f32 (wg_f2h); -// - D(h) / e_prev is an IEEE division (wg_div). +// - r_prev = 1 / e_prev is one IEEE division per launch (wg_div), then L = D(h) r_prev (HIP: the same product; the +// spec's division as a product with its reciprocal is free within the post tolerance, spec answer Q5). // Model M's DXILs are untouched: this file only adds the kernel (bound to l_post.co). // l_post_recon_p12 (LP_P12 = 1): the same kernel on the pack12 coefficients of l_coef_pack12: A = image A8 (8 bytes per // render pixel: codes A0 A1 A2 B0 | B1 B2, C2 FP16), C = image C4 (4 bytes per render pixel: C0, C1 FP16), B unused. @@ -65,7 +67,7 @@ groupshared uint sh_tex[kStageCap * kTexDw]; static uint g_z; // 0, opaque to the compiler (from the grid size) static float g_e, g_ep; // exposures (0 replaced by 1) -static float g_epy, g_eplo, g_ephi; // the division window of e_prev (ep_window) +static float g_rp; // 1 / e_prev (IEEE division, round to nearest even) float opq(float x) { return asfloat(asuint(x) ^ g_z); } half opqh(half x) { return asfloat16((uint16_t)(asuint16(x) ^ (uint16_t)g_z)); } half h_bits(uint b) { return asfloat16((uint16_t)b); } @@ -213,36 +215,6 @@ void gather_acc(inout LGather g, uint t[kTexDw], float w, uint hwb) } } -// a / e_prev as wg_div (IEEE, round to nearest even) for the three channels. e_prev is one value per launch: with -// y = RN(1 / e_prev) the lanes whose |a| lies in [lo, hi) take q0 = a y, rn = fma(e_prev, q0, -a), q = fma(-rn, y, q0), -// which is the correctly rounded quotient there (clean/runtime/native_direct_pi.h: every intermediate normal, the -// [1, 2) significand form checked for all significand pairs; the direct route's colour division, -// native_direct.hlsli native_direct_color_div3); every other lane (NaN, infinity, zero, a quotient near the FP32 range -// ends, an e_prev outside the window's domain: hi = 0) runs wg_div itself. -void ep_window() -{ - const uint ub = asuint(g_ep); - const int eb = (int)((ub >> 23) & 255u), B = eb - 127; - const bool ok = (ub >> 31) == 0u && eb >= 1 && eb <= 252; - g_epy = wg_div(1.0f, g_ep); - g_eplo = asfloat((uint)(max(B - 125, -79) + 127) << 23); - g_ephi = !ok ? 0.0f : (B > 0 ? asfloat(0x7f800000u) : asfloat((uint)(B + 254) << 23)); -} -float3 ep_div3(float3 a) -{ - const float3 q0 = a * g_epy; - const float3 rn = mad(g_ep, q0, -a); - float3 q = mad(-rn, g_epy, q0); - const float3 m = abs(a); - const bool3 fast = and(m < g_ephi, m >= g_eplo); - [branch] if (!all(fast)) { - q.x = fast.x ? q.x : wg_div(a.x, g_ep); - q.y = fast.y ? q.y : wg_div(a.y, g_ep); - q.z = fast.z ? q.z : wg_div(a.z, g_ep); - } - return q; -} - // ---- history: 12-tap Catmull-Rom of the previous full history (history_cr, 8-byte RGBA16F texels) ------------ float4 cr_weights(float f) { @@ -319,7 +291,7 @@ void l_post_core(LGather g, int x, int y, int fx, int fy, float mx[3], float my[ const bool valid = P_reset == 0 && qx >= 0.0f && qx <= (float)P_dw && qy >= 0.0f && qy <= (float)P_dh; if (valid) { const float3 h = history_cr(qx, qy); - const float3 lq = ep_div3(float3(l_dec(h.x), l_dec(h.y), l_dec(h.z))); + const float3 lq = float3(l_dec(h.x), l_dec(h.y), l_dec(h.z)) * g_rp; L[n][0] = lq.x; L[n][1] = lq.y; L[n][2] = lq.z; @@ -397,7 +369,7 @@ WG_KERNEL(l_post_recon) g_e = e0 == 0.0f ? 1.0f : e0; g_ep = e1 == 0.0f ? 1.0f : e1; } - ep_window(); + g_rp = wg_div(1.0f, g_ep); const int lane = (int)wg_lane_id(); const int tiles_x = (P_dw + kTileX - 1) / kTileX; const int tile = (int)wg_block_id(); diff --git a/clean/l/kernels/l_post.hip b/clean/l/kernels/l_post.hip index 168cfe13..97fb4d60 100644 --- a/clean/l/kernels/l_post.hip +++ b/clean/l/kernels/l_post.hip @@ -4,7 +4,8 @@ // (clean/l/reference/recon.py, the oracle): // 1. a 4 x 4 colour gather: columns ax - 1 ... ax + 2 and rows b - 1 ... b + 2 (b from the even row of the pair); // 2. each history candidate decoded and divided by the previous exposure before the two are combined: -// H_lin = (w1 L1 + w2 L2) e, with L_i = D(H_i) / e_prev (L2 only when w2 > 0); +// H_lin = (w1 L1 + w2 L2) e, with L_i = D(H_i) r_prev, r_prev = 1 / e_prev (L2 only when w2 > 0; the spec's +// division by e_prev as a product with its reciprocal: free within the post tolerance, spec answer Q5); // 3. an invalid or reset sample falls back to the render pixel's linear colour (no exposure, no encode). // The colour gather runs in the model's own precision as confirmed by the spec side (oracle: recon.gather_half): // fused target positions, FP16 offsets and exponent with the fused cross term, float32 weights flushed to +0 below @@ -149,6 +150,7 @@ WG_FN void l_post_core(const LGather& g, int x, int y, int fx, int fy, const flo const LSel sl = l_select(g); const float X = (float)x + 0.5f, Y = (float)y + 0.5f; const int ncand = sl.w2 != 0.0f ? 2 : 1; + const float rp = 1.0f / p.e_prev; // IEEE division float L[2][3]; for (int n = 0; n < ncand; ++n) { const int k = n == 0 ? sl.first : sl.second; @@ -158,7 +160,7 @@ WG_FN void l_post_core(const LGather& g, int x, int y, int fx, int fy, const flo float h[3]; history_cr(hist_prev, p.dw, p.dh, qx, qy, h); #pragma unroll - for (int i = 0; i < 3; ++i) L[n][i] = l_dec(h[i]) / p.e_prev; + for (int i = 0; i < 3; ++i) L[n][i] = l_dec(h[i]) * rp; } else { const v4f c = color[fy * p.rw + fx]; #pragma unroll