Dusk/WeirdUtils
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

silicon_sse: naked FMA asm for vec3Dot (0.4x->0.8x) and distanceToPlane (0.7x->1.0x)

Browse Source
This commit is contained in:
MarcelineVQ
2026-03-17 01:01:07 -07:00
parent 9488fb568c
commit cb0888e0fa
2 changed files with 84 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/silicon/silicon_sse.zig
+53 -17
View File
@@ -108,18 +108,40 @@ export fn si_createRotMat3x4(out: u32, axis_ptr: u32, angle_bits: u32, is_normal
}
// --- 0x6329E0: distanceToPlane (525K/7.5s) ---
// (dot(point,normal)+d) / dot(direction,normal)
// Fix: @mulAdd chains, keep f32 until return
export fn si_distanceToPlane(point: u32, plane: u32, direction: u32) f64 {
const p: [*]const f32 = @ptrFromInt(point);
const pl: [*]const f32 = @ptrFromInt(plane);
const dir: [*]const f32 = @ptrFromInt(direction);
const dot1 = @mulAdd(f32, p[2], pl[2], @mulAdd(f32, p[1], pl[1], @mulAdd(f32, p[0], pl[0], pl[3])));
const dot2 = @mulAdd(f32, dir[2], pl[2], @mulAdd(f32, dir[1], pl[1], dir[0] * pl[0]));
if (@abs(dot2) < 1.0e-20) return 0.0;
return @as(f64, dot1) / @as(f64, dot2);
// Original: __fastcall(ECX=point, EDX=plane, stack[0]=direction), returns ST(0), RET 0x4.
// (dot(point,normal)+d) / dot(direction,normal). 70 bytes, 14cy.
// Naked FMA version: 2 dot products via vfmadd + vdivss, transfer to ST(0).
export fn si_distanceToPlane() callconv(.naked) void {
// ECX=point, EDX=plane, [ESP+4]=direction. Return ST(0), RET 0x4.
asm volatile (
// dot1 = p[0]*pl[0] + p[1]*pl[1] + p[2]*pl[2] + pl[3]
\\vmovss (%%ecx), %%xmm0
\\vmulss (%%edx), %%xmm0, %%xmm0
\\vmovss 4(%%ecx), %%xmm1
\\vfmadd231ss 4(%%edx), %%xmm1, %%xmm0
\\vmovss 8(%%ecx), %%xmm1
\\vfmadd231ss 8(%%edx), %%xmm1, %%xmm0
\\vaddss 12(%%edx), %%xmm0, %%xmm0
// dot2 = dir[0]*pl[0] + dir[1]*pl[1] + dir[2]*pl[2]
\\mov 4(%%esp), %%eax
\\vmovss (%%eax), %%xmm2
\\vmulss (%%edx), %%xmm2, %%xmm2
\\vmovss 4(%%eax), %%xmm1
\\vfmadd231ss 4(%%edx), %%xmm1, %%xmm2
\\vmovss 8(%%eax), %%xmm1
\\vfmadd231ss 8(%%edx), %%xmm1, %%xmm2
// result = dot1 / dot2 (skip epsilon check for speed — game tolerates it)
\\vdivss %%xmm2, %%xmm0, %%xmm0
// Transfer to ST(0)
\\sub $4, %%esp
\\vmovss %%xmm0, (%%esp)
\\flds (%%esp)
\\add $4, %%esp
\\ret $4
);
}
// --- 0x686C20: classifyPointFrustum (3.2M/7.5s) ---
// Tests point against 6 frustum planes, produces 6-bit bitmask.
// Uses V4 dot product: load plane as V4, point as {x,y,z,1}, dot4.
@@ -384,13 +406,27 @@ export fn si_ftol() callconv(.naked) void {
);
}
// --- 0x602630: vec3Dot ---
// Uses @mulAdd FMA chain. Returns f64 (x87 ABI contract).
export fn si_vec3Dot(a: u32, b: u32) f64 {
const va: [*]const f32 = @ptrFromInt(a);
const vb: [*]const f32 = @ptrFromInt(b);
const result = @mulAdd(f32, va[2], vb[2], @mulAdd(f32, va[1], vb[1], va[0] * vb[0]));
return @floatCast(result);
// --- 0x602630: vec3Dot (31K/7.5s, 0.05ms total) ---
// Original: __fastcall(ECX=a, EDX=b), returns ST(0). 21 bytes, 5cy.
// Achieved 0.8x (6cy) via naked FMA — the 1cy gap is the SSE->x87 transfer
// for the ST(0) return that callers expect. DPPS (SSE4.1) is 7-11cy, no better.
// x87 is inherently optimal here: 21 bytes, no domain crossing, pipelined.
// Not worth further optimization at 31K calls — 0.05ms total frame cost.
export fn si_vec3Dot() callconv(.naked) void {
// ECX = a ptr, EDX = b ptr (fastcall), return in ST(0)
asm volatile (
\\vmovss (%%ecx), %%xmm0
\\vmulss (%%edx), %%xmm0, %%xmm0
\\vmovss 4(%%ecx), %%xmm1
\\vfmadd231ss 4(%%edx), %%xmm1, %%xmm0
\\vmovss 8(%%ecx), %%xmm1
\\vfmadd231ss 8(%%edx), %%xmm1, %%xmm0
\\sub $4, %%esp
\\vmovss %%xmm0, (%%esp)
\\flds (%%esp)
\\add $4, %%esp
\\ret
);
}
// --- 0x686820: translateBoundingVol ---