Skip to content

armgit / gecko

Public Imported from GitHub
0

A GameCube/Wii emulator and debugger written in Rust.

SIMD position decode + 16 byte oob slack

Layle5daabcb32713
diff --git a/crates/gecko/src/flipper/gx/fifo.rs b/crates/gecko/src/flipper/gx/fifo.rs
index 83cc20a..82ed0d3 100644
--- a/crates/gecko/src/flipper/gx/fifo.rs
+++ b/crates/gecko/src/flipper/gx/fifo.rs
@@ -10,6 +10,9 @@ impl GraphicsProcessor {
pub fn drain_fifo<const SYSTEM: SystemId>(&mut self, mmio: &mut Mmio<SYSTEM>, renderer: &mut dyn RenderSink) {
let mut fifo = std::mem::take(&mut self.fifo);
+ #[cfg(feature = "jit")]
+ fifo.reserve(crate::flipper::gx::jit::VEC_OVERREAD_BYTES);
+
let pos = self::drain_buf::<SYSTEM>(self, mmio, renderer, &fifo);
if pos > 0 {
fifo.drain(..pos);
@@ -123,6 +126,10 @@ fn drain_buf<const SYSTEM: SystemId>(
let mut buf = std::mem::take(&mut gp.dl_scratch);
buf.clear();
buf.extend_from_slice(slice);
+
+ #[cfg(feature = "jit")]
+ buf.reserve(crate::flipper::gx::jit::VEC_OVERREAD_BYTES);
+
buf
}
None => {
diff --git a/crates/gecko/src/flipper/gx/jit.rs b/crates/gecko/src/flipper/gx/jit.rs
index 38c5a47..ebba208 100644
--- a/crates/gecko/src/flipper/gx/jit.rs
+++ b/crates/gecko/src/flipper/gx/jit.rs
@@ -85,6 +85,9 @@ unsafe impl Send for ResolvedArray {}
/// 12 array slots: pos, nrm, clr0, clr1, tex0..tex7.
pub const RESOLVED_ARRAY_COUNT: usize = 12;
+/// Slack every buffer keeps past its last byte so the JITs vector loads can run off the data without faulting.
+pub const VEC_OVERREAD_BYTES: usize = 16;
+
#[repr(C)]
#[derive(Clone, Debug)]
pub struct ResolvedArrays(pub [ResolvedArray; RESOLVED_ARRAY_COUNT]);
@@ -453,7 +456,7 @@ pub fn resolve_arrays_for_draw(
};
let max_byte_off = max_idx
.saturating_mul(r.stride as u64)
- .saturating_add(attr_data_size[slot] as u64);
+ .saturating_add(attr_data_size[slot].max(VEC_OVERREAD_BYTES) as u64);
let remaining = bank_remaining_from(r.host_base) as u64;
if max_byte_off > remaining {
return false;
diff --git a/crates/gecko/src/flipper/gx/jit/attr.rs b/crates/gecko/src/flipper/gx/jit/attr.rs
index f083f7c..b19462c 100644
--- a/crates/gecko/src/flipper/gx/jit/attr.rs
+++ b/crates/gecko/src/flipper/gx/jit/attr.rs
@@ -5,7 +5,7 @@ use cranelift_frontend::FunctionBuilder;
use super::VtxKey;
use super::builder::{MEMFLAGS, MEMFLAGS_RO, array_offset, offset, xf_byte_off};
use crate::flipper::gx::constants::*;
-use crate::flipper::gx::regs::{AttributeType, ColorCount, ColorFormat, ComponentFormat, PosCount, TexCount};
+use crate::flipper::gx::regs::{AttributeType, ColorCount, ColorFormat, ComponentFormat, NrmCount, PosCount, TexCount};
pub struct AttrCtx<'a, 'f> {
pub bd: &'a mut FunctionBuilder<'f>,
@@ -205,7 +205,13 @@ fn read_tex_mtx_indices(ctx: &mut AttrCtx, vcd_lo: crate::flipper::gx::regs::Vcd
})
}
-fn attr_ptr_and_advance(ctx: &mut AttrCtx, attr: AttributeType, array_idx: usize, direct_size: usize) -> ir::Value {
+fn attr_ptr_and_advance(
+ ctx: &mut AttrCtx,
+ attr: AttributeType,
+ array_idx: usize,
+ direct_size: usize,
+ extra_index_slots: i64,
+) -> ir::Value {
match attr {
AttributeType::Direct => {
let p = ctx.data_ptr;
@@ -215,14 +221,14 @@ fn attr_ptr_and_advance(ctx: &mut AttrCtx, attr: AttributeType, array_idx: usize
AttributeType::Index8 => {
let raw = ctx.bd.ins().load(ir::types::I8, MEMFLAGS_RO, ctx.data_ptr, 0);
let idx = ctx.bd.ins().uextend(ir::types::I32, raw);
- ctx.data_ptr = ctx.bd.ins().iadd_imm(ctx.data_ptr, 1);
+ ctx.data_ptr = ctx.bd.ins().iadd_imm(ctx.data_ptr, 1 + extra_index_slots);
indexed_addr(ctx, array_idx, idx)
}
AttributeType::Index16 => {
let raw = ctx.bd.ins().load(ir::types::I16, MEMFLAGS_RO, ctx.data_ptr, 0);
let raw = ctx.bd.ins().bswap(raw);
let idx = ctx.bd.ins().uextend(ir::types::I32, raw);
- ctx.data_ptr = ctx.bd.ins().iadd_imm(ctx.data_ptr, 2);
+ ctx.data_ptr = ctx.bd.ins().iadd_imm(ctx.data_ptr, 2 * (1 + extra_index_slots));
indexed_addr(ctx, array_idx, idx)
}
AttributeType::None => ctx.data_ptr,
@@ -299,29 +305,123 @@ fn decode_component_at(
}
}
+const LITTLE: ir::MemFlagsData = ir::MemFlagsData::new().with_endianness(ir::Endianness::Little);
+
+const ZERO_LANE: u8 = 0xFF;
+
+const SWZ_S16_VEC2: [u8; 16] = [
+ ZERO_LANE, ZERO_LANE, 1, 0, ZERO_LANE, ZERO_LANE, 3, 2, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_S16_VEC3: [u8; 16] = [
+ ZERO_LANE, ZERO_LANE, 1, 0, ZERO_LANE, ZERO_LANE, 3, 2, ZERO_LANE, ZERO_LANE, 5, 4, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE,
+];
+const SWZ_U16_VEC2: [u8; 16] = [
+ 1, 0, ZERO_LANE, ZERO_LANE, 3, 2, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_U16_VEC3: [u8; 16] = [
+ 1, 0, ZERO_LANE, ZERO_LANE, 3, 2, ZERO_LANE, ZERO_LANE, 5, 4, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE,
+];
+const SWZ_S8_VEC2: [u8; 16] = [
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, 0, ZERO_LANE, ZERO_LANE, ZERO_LANE, 1, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_S8_VEC3: [u8; 16] = [
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, 0, ZERO_LANE, ZERO_LANE, ZERO_LANE, 1, ZERO_LANE, ZERO_LANE, ZERO_LANE, 2,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_U8_VEC2: [u8; 16] = [
+ 0, ZERO_LANE, ZERO_LANE, ZERO_LANE, 1, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_U8_VEC3: [u8; 16] = [
+ 0, ZERO_LANE, ZERO_LANE, ZERO_LANE, 1, ZERO_LANE, ZERO_LANE, ZERO_LANE, 2, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+ ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_F32_VEC2: [u8; 16] = [
+ 3, 2, 1, 0, 7, 6, 5, 4, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+const SWZ_F32_VEC3: [u8; 16] = [
+ 3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, ZERO_LANE, ZERO_LANE, ZERO_LANE, ZERO_LANE,
+];
+
+fn swizzle_const(ctx: &mut AttrCtx, mask: &[u8; 16]) -> ir::Value {
+ let handle = ctx.bd.func.dfg.constants.insert(ir::ConstantData::from(&mask[..]));
+ ctx.bd.ins().vconst(ir::types::I8X16, handle)
+}
+
+fn extract3(ctx: &mut AttrCtx, v: ir::Value) -> [ir::Value; 3] {
+ let x = ctx.bd.ins().extractlane(v, 0);
+ let y = ctx.bd.ins().extractlane(v, 1);
+ let z = ctx.bd.ins().extractlane(v, 2);
+
+ [x, y, z]
+}
+
+fn decode_vec_simd(ctx: &mut AttrCtx, ptr: ir::Value, fmt: ComponentFormat, triplet: bool, recip: f32) -> ir::Value {
+ if matches!(fmt, ComponentFormat::F32) {
+ // swappedy swap swap swap
+ let raw = ctx.bd.ins().load(ir::types::I32X4, MEMFLAGS_RO, ptr, 0);
+ let bytes = ctx.bd.ins().bitcast(ir::types::I8X16, LITTLE, raw);
+
+ let mask = if triplet { &SWZ_F32_VEC3 } else { &SWZ_F32_VEC2 };
+ let mask = self::swizzle_const(ctx, mask);
+ let shuffled = ctx.bd.ins().swizzle(bytes, mask);
+
+ return ctx.bd.ins().bitcast(ir::types::F32X4, LITTLE, shuffled);
+ }
+
+ let signed = matches!(fmt, ComponentFormat::S8 | ComponentFormat::S16);
+ let wide16 = matches!(fmt, ComponentFormat::U16 | ComponentFormat::S16);
+
+ let mask = match (wide16, signed, triplet) {
+ (true, true, true) => &SWZ_S16_VEC3,
+ (true, true, false) => &SWZ_S16_VEC2,
+ (true, false, true) => &SWZ_U16_VEC3,
+ (true, false, false) => &SWZ_U16_VEC2,
+ (false, true, true) => &SWZ_S8_VEC3,
+ (false, true, false) => &SWZ_S8_VEC2,
+ (false, false, true) => &SWZ_U8_VEC3,
+ (false, false, false) => &SWZ_U8_VEC2,
+ };
+
+ let bytes = ctx.bd.ins().load(ir::types::I8X16, MEMFLAGS_RO, ptr, 0);
+ let mask = self::swizzle_const(ctx, mask);
+ let shuffled = ctx.bd.ins().swizzle(bytes, mask);
+ let lanes = ctx.bd.ins().bitcast(ir::types::I32X4, LITTLE, shuffled);
+
+ let floats = if signed {
+ let bits = if wide16 { 16 } else { 8 };
+ let ext = ctx.bd.ins().sshr_imm(lanes, 32 - bits);
+ ctx.bd.ins().fcvt_from_sint(ir::types::F32X4, ext)
+ } else {
+ ctx.bd.ins().fcvt_from_uint(ir::types::F32X4, lanes)
+ };
+
+ let scale = ctx.bd.ins().f32const(recip);
+ let scale = ctx.bd.ins().splat(ir::types::F32X4, scale);
+ ctx.bd.ins().fmul(floats, scale)
+}
+
fn decode_position(ctx: &mut AttrCtx, attr: AttributeType, vat_a: crate::flipper::gx::regs::VatA) -> [ir::Value; 3] {
if matches!(attr, AttributeType::None) {
let z = ctx.bd.ins().f32const(0.0);
return [z, z, z];
}
- let count = vat_a.pos_cnt().components();
+ let triplet = matches!(vat_a.pos_cnt(), PosCount::Xyz);
let fmt = vat_a.pos_fmt();
let direct = vat_a.pos_data_size();
let recip = 1.0f32 / ((1u32 << vat_a.pos_shift()) as f32);
- let ptr = attr_ptr_and_advance(ctx, attr, ARRAY_POS, direct);
- let comp_size = fmt.size() as i32;
- let mut out = [ctx.bd.ins().f32const(0.0); 3];
- for i in 0..count {
- out[i] = decode_component_at(ctx, ptr, (i as i32) * comp_size, fmt, recip);
- }
+ let ptr = attr_ptr_and_advance(ctx, attr, ARRAY_POS, direct, 0);
- if matches!(vat_a.pos_cnt(), PosCount::Xy) {
- out[2] = ctx.bd.ins().f32const(0.0);
- }
+ let vector = self::decode_vec_simd(ctx, ptr, fmt, triplet, recip);
- out
+ self::extract3(ctx, vector)
}
fn decode_normal(ctx: &mut AttrCtx, attr: AttributeType, vat_a: crate::flipper::gx::regs::VatA) -> [ir::Value; 3] {
@@ -330,7 +430,6 @@ fn decode_normal(ctx: &mut AttrCtx, attr: AttributeType, vat_a: crate::flipper::
let one = ctx.bd.ins().f32const(1.0);
return [z, z, one];
}
- let cnt = vat_a.nrm_cnt().components().min(3);
let fmt = vat_a.nrm_fmt();
let direct = vat_a.nrm_data_size();
@@ -340,15 +439,16 @@ fn decode_normal(ctx: &mut AttrCtx, attr: AttributeType, vat_a: crate::flipper::
ComponentFormat::F32 => 1.0f32,
};
- let ptr = attr_ptr_and_advance(ctx, attr, ARRAY_NRM, direct);
- let comp_size = fmt.size() as i32;
- let mut out = [ctx.bd.ins().f32const(0.0); 3];
+ let extra_index_slots = if vat_a.nrm_index3() && matches!(vat_a.nrm_cnt(), NrmCount::Nbt) {
+ 2
+ } else {
+ 0
+ };
+ let ptr = attr_ptr_and_advance(ctx, attr, ARRAY_NRM, direct, extra_index_slots);
- for i in 0..cnt {
- out[i] = decode_component_at(ctx, ptr, (i as i32) * comp_size, fmt, recip);
- }
+ let vector = self::decode_vec_simd(ctx, ptr, fmt, true, recip);
- out
+ self::extract3(ctx, vector)
}
fn decode_color(
@@ -364,7 +464,7 @@ fn decode_color(
}
let direct = fmt.data_size(cnt);
- let ptr = attr_ptr_and_advance(ctx, attr, array_idx, direct);
+ let ptr = attr_ptr_and_advance(ctx, attr, array_idx, direct, 0);
decode_color_bytes(ctx, ptr, fmt, cnt)
}
@@ -464,16 +564,21 @@ fn decode_texcoord(
array_idx: usize,
) -> [ir::Value; 2] {
let count = cnt.components();
- let comp_size = fmt.size() as i32;
let direct = count * fmt.size();
let recip = 1.0f32 / ((1u32 << shift) as f32);
- let ptr = attr_ptr_and_advance(ctx, attr, array_idx, direct);
- let zero = ctx.bd.ins().f32const(0.0);
- let mut out = [zero, zero];
- for i in 0..count {
- out[i] = decode_component_at(ctx, ptr, (i as i32) * comp_size, fmt, recip);
+ let ptr = attr_ptr_and_advance(ctx, attr, array_idx, direct, 0);
+
+ if count == 2 {
+ let vector = self::decode_vec_simd(ctx, ptr, fmt, false, recip);
+ let s = ctx.bd.ins().extractlane(vector, 0);
+ let t = ctx.bd.ins().extractlane(vector, 1);
+ return [s, t];
}
- out
+
+ let s = decode_component_at(ctx, ptr, 0, fmt, recip);
+ let t = ctx.bd.ins().f32const(0.0);
+
+ [s, t]
}
fn xf_transform_3x4(ctx: &mut AttrCtx, pos_mtx_idx: ir::Value, pos: &[ir::Value; 3]) -> [ir::Value; 3] {
diff --git a/crates/gecko/src/flipper/gx/jit/builder.rs b/crates/gecko/src/flipper/gx/jit/builder.rs
index 56528cc..f575cd7 100644
--- a/crates/gecko/src/flipper/gx/jit/builder.rs
+++ b/crates/gecko/src/flipper/gx/jit/builder.rs
@@ -7,7 +7,6 @@ use std::mem::offset_of;
use super::attr::{self, AttrCtx};
use super::{ResolvedArray, VtxKey};
-use crate::flipper::gx::regs::{AttributeType, NrmCount};
use crate::host::DrawVertex;
pub(crate) const MEMFLAGS: ir::MemFlagsData = ir::MemFlagsData::new().with_notrap();
@@ -20,7 +19,7 @@ pub fn build_parser(
pointer_ty: ir::Type,
key: VtxKey,
) -> bool {
- if !is_supported(key) {
+ if !self::is_supported() {
return false;
}
@@ -110,22 +109,8 @@ pub fn build_parser(
true
}
-fn is_supported(key: VtxKey) -> bool {
- if std::env::var("GECKO_VTX_JIT_OFF").is_ok() {
- return false;
- }
-
- let vcd_lo = key.vcd_lo();
- let vat_a = key.vat_a();
-
- if matches!(vcd_lo.normal(), AttributeType::Index8 | AttributeType::Index16)
- && vat_a.nrm_index3()
- && vat_a.nrm_cnt() == NrmCount::Nbt
- {
- return false;
- }
-
- true
+fn is_supported() -> bool {
+ std::env::var("GECKO_VTX_JIT_OFF").is_err()
}
pub mod offset {