Skip to content

armgit / gecko

Public Imported from GitHub
0

A GameCube/Wii emulator and debugger written in Rust.

simd pos/normal transforms via column acc

Layle52de7244277a
diff --git a/crates/gecko/src/flipper/gx/jit/attr.rs b/crates/gecko/src/flipper/gx/jit/attr.rs
index 00c0d26..9a20cd0 100644
--- a/crates/gecko/src/flipper/gx/jit/attr.rs
+++ b/crates/gecko/src/flipper/gx/jit/attr.rs
@@ -93,10 +93,13 @@ pub fn emit_vertex(ctx: &mut AttrCtx) {
}
}
- let pos_view = xf_transform_3x4(ctx, pos_mtx_idx, &pos_xyz);
+ // The matrix is loop-invariant unless the per-vertex matrix index is streamed.
+ let mtx_invariant = !vcd_lo.pos_nrm_mtx_idx();
+
+ let pos_view = xf_transform_3x4(ctx, pos_mtx_idx, &pos_xyz, mtx_invariant);
store_vec3(ctx.bd, ctx.out_ptr, offset::POS_VIEW, &pos_view);
- let nrm_view = transform_and_normalize_normal(ctx, pos_mtx_idx, &nrm_xyz);
+ let nrm_view = transform_and_normalize_normal(ctx, pos_mtx_idx, &nrm_xyz, mtx_invariant);
store_vec3(ctx.bd, ctx.out_ptr, offset::NORMAL, &nrm_view);
self::emit_texgens(ctx, &pos_xyz, &nrm_xyz, &tex_mtx_idx, &raw_st);
@@ -544,17 +547,37 @@ fn decode_texcoord(
[s, t]
}
-fn xf_transform_3x4(ctx: &mut AttrCtx, pos_mtx_idx: ir::Value, pos: &[ir::Value; 3]) -> [ir::Value; 3] {
+fn xf_transform_3x4(
+ ctx: &mut AttrCtx,
+ pos_mtx_idx: ir::Value,
+ pos: &[ir::Value; 3],
+ invariant: bool,
+) -> [ir::Value; 3] {
let base_addr = self::pos_mtx_base_addr(ctx, pos_mtx_idx);
let m = self::load_matrix12(ctx, base_addr, MEMFLAGS_RO_MOVABLE);
- let r0 = self::dot3_affine(ctx, [m[0], m[1], m[2], m[3]], *pos);
- let r1 = self::dot3_affine(ctx, [m[4], m[5], m[6], m[7]], *pos);
- let r2 = self::dot3_affine(ctx, [m[8], m[9], m[10], m[11]], *pos);
- [r0, r1, r2]
+ if !invariant {
+ // Streamed per-vertex matrix: the columns can't hoist(?), so scalar rows
+ // are cheaper than building them every vertex.
+ let r0 = self::dot3_affine(ctx, [m[0], m[1], m[2], m[3]], *pos);
+ let r1 = self::dot3_affine(ctx, [m[4], m[5], m[6], m[7]], *pos);
+ let r2 = self::dot3_affine(ctx, [m[8], m[9], m[10], m[11]], *pos);
+ return [r0, r1, r2];
+ }
+
+ let col0 = self::make_col(ctx, [m[0], m[4], m[8]]);
+ let col1 = self::make_col(ctx, [m[1], m[5], m[9]]);
+ let col2 = self::make_col(ctx, [m[2], m[6], m[10]]);
+ let col3 = self::make_col(ctx, [m[3], m[7], m[11]]);
+ self::mat_vec_simd(ctx, [col0, col1, col2], Some(col3), *pos)
}
-fn transform_and_normalize_normal(ctx: &mut AttrCtx, pos_mtx_idx: ir::Value, nrm: &[ir::Value; 3]) -> [ir::Value; 3] {
+fn transform_and_normalize_normal(
+ ctx: &mut AttrCtx,
+ pos_mtx_idx: ir::Value,
+ nrm: &[ir::Value; 3],
+ invariant: bool,
+) -> [ir::Value; 3] {
// nrm_mtx_base = XF_NRM_MTX_BASE + (pos_mtx_idx & 31) * 3 cells
let masked = ctx.bd.ins().band_imm(pos_mtx_idx, 31);
let cell_off = ctx.bd.ins().imul_imm(masked, 3 * 4);
@@ -570,21 +593,18 @@ fn transform_and_normalize_normal(ctx: &mut AttrCtx, pos_mtx_idx: ir::Value, nrm
)
});
- let row = |i: usize, ctx: &mut AttrCtx| -> ir::Value {
- let m0 = nm[i * 3 + 0];
- let m1 = nm[i * 3 + 1];
- let m2 = nm[i * 3 + 2];
- let p0 = ctx.bd.ins().fmul(m0, nrm[0]);
- let p1 = ctx.bd.ins().fmul(m1, nrm[1]);
- let p2 = ctx.bd.ins().fmul(m2, nrm[2]);
- let s = ctx.bd.ins().fadd(p0, p1);
- ctx.bd.ins().fadd(s, p2)
+ let [nx, ny, nz] = if invariant {
+ let col0 = self::make_col(ctx, [nm[0], nm[3], nm[6]]);
+ let col1 = self::make_col(ctx, [nm[1], nm[4], nm[7]]);
+ let col2 = self::make_col(ctx, [nm[2], nm[5], nm[8]]);
+ self::mat_vec_simd(ctx, [col0, col1, col2], None, *nrm)
+ } else {
+ let r0 = self::dot3_linear(ctx, [nm[0], nm[1], nm[2]], *nrm);
+ let r1 = self::dot3_linear(ctx, [nm[3], nm[4], nm[5]], *nrm);
+ let r2 = self::dot3_linear(ctx, [nm[6], nm[7], nm[8]], *nrm);
+ [r0, r1, r2]
};
- let nx = row(0, ctx);
- let ny = row(1, ctx);
- let nz = row(2, ctx);
-
// length = sqrt(nx*nx + ny*ny + nz*nz)
let nx2 = ctx.bd.ins().fmul(nx, nx);
let ny2 = ctx.bd.ins().fmul(ny, ny);
@@ -848,6 +868,46 @@ fn dot3_affine(ctx: &mut AttrCtx, m: [ir::Value; 4], v: [ir::Value; 3]) -> ir::V
ctx.bd.ins().fadd(b, m[3])
}
+// (m0*v0 + m1*v1) + m2*v2, the linear 3x3 row (no translate).
+fn dot3_linear(ctx: &mut AttrCtx, m: [ir::Value; 3], v: [ir::Value; 3]) -> ir::Value {
+ let p0 = ctx.bd.ins().fmul(m[0], v[0]);
+ let p1 = ctx.bd.ins().fmul(m[1], v[1]);
+ let p2 = ctx.bd.ins().fmul(m[2], v[2]);
+ let a = ctx.bd.ins().fadd(p0, p1);
+ ctx.bd.ins().fadd(a, p2)
+}
+
+// Pack 3 scalars into the low lanes of an F32X4 (a matrix column).
+fn make_col(ctx: &mut AttrCtx, c: [ir::Value; 3]) -> ir::Value {
+ let v = ctx.bd.ins().scalar_to_vector(ir::types::F32X4, c[0]);
+ let v = ctx.bd.ins().insertlane(v, c[1], 1);
+ ctx.bd.ins().insertlane(v, c[2], 2)
+}
+
+// All three output rows at once: cols[0]*x + cols[1]*y + cols[2]*z (+ translate).
+fn mat_vec_simd(
+ ctx: &mut AttrCtx,
+ cols: [ir::Value; 3],
+ translate: Option<ir::Value>,
+ v: [ir::Value; 3],
+) -> [ir::Value; 3] {
+ let sx = ctx.bd.ins().splat(ir::types::F32X4, v[0]);
+ let sy = ctx.bd.ins().splat(ir::types::F32X4, v[1]);
+ let sz = ctx.bd.ins().splat(ir::types::F32X4, v[2]);
+
+ let acc = ctx.bd.ins().fmul(cols[0], sx);
+ let p1 = ctx.bd.ins().fmul(cols[1], sy);
+ let acc = ctx.bd.ins().fadd(acc, p1);
+ let p2 = ctx.bd.ins().fmul(cols[2], sz);
+ let acc = ctx.bd.ins().fadd(acc, p2);
+ let acc = match translate {
+ Some(c3) => ctx.bd.ins().fadd(acc, c3),
+ None => acc,
+ };
+
+ self::extract3(ctx, acc)
+}
+
fn emit_dual(
ctx: &mut AttrCtx,
tg: usize,