I guess I can't do better. Scalar still suck, but at least it is builtin SIDM so ok I guess

This commit is contained in:
adrien 2026-04-27 22:07:39 +02:00
parent a0961e7571
commit c6f613a787
3 changed files with 63 additions and 27 deletions

View File

@ -118,6 +118,12 @@ pub fn set(self: *Self, key: Dimension, val: UnitScale) void {
self.data.set(key, val); self.data.set(key, val);
} }
pub fn eql(self: Self, other: Self) bool {
for (self.data.values, other.data.values) |l, r|
if (l != r) return false;
return true;
}
pub fn argsOpt(self: Self) ArgOpts { pub fn argsOpt(self: Self) ArgOpts {
var args: ArgOpts = undefined; var args: ArgOpts = undefined;
for (std.enums.values(Dimension)) |d| for (std.enums.values(Dimension)) |d|

View File

@ -203,7 +203,7 @@ pub fn Tensor(
if (s == 0) @compileError("Tensor shape dimensions must be strictly >= 1."); if (s == 0) @compileError("Tensor shape dimensions must be strictly >= 1.");
} }
} }
@setEvalBranchQuota(10_000_000); @setEvalBranchQuota(100_000);
const _total: usize = comptime shapeTotal(shape_); const _total: usize = comptime shapeTotal(shape_);
const _strides = comptime shapeStrides(shape_); const _strides = comptime shapeStrides(shape_);
@ -271,20 +271,20 @@ pub fn Tensor(
finerScales(Self, RhsT(@TypeOf(r))).argsOpt(), finerScales(Self, RhsT(@TypeOf(r))).argsOpt(),
shape_, shape_,
) { ) {
const rhs_q = rhs(r); const rhs_t = rhs(r);
const RhsType = @TypeOf(r); const RhsType = @TypeOf(r);
if (comptime !dims.eql(RhsType.dims)) if (comptime !dims.eql(RhsType.dims))
@compileError("Dimension mismatch in add: " ++ dims.str() ++ " vs " ++ RhsType.dims.str()); @compileError("Dimension mismatch in add: " ++ dims.str() ++ " vs " ++ RhsType.dims.str());
if (comptime RhsType.total != 1 and !shapeEql(shape_, RhsType.shape)) if (comptime RhsType.total != 1 and !shapeEql(shape_, RhsType.shape))
@compileError("Shape mismatch in add: element-wise operations require identical shapes, or a scalar RHS."); @compileError("Shape mismatch in add: element-wise operations require identical shapes, or a scalar RHS.");
// if (comptime total == 1 and RhsType.scales.eql(scales)) if (comptime total == 1 and scales.eql(RhsType.scales)) // Here rhs_t has to be {1} too
// return .{ .data = if (comptime isInt(T)) self.data +| r.data else self.data + r.data }; return .{ .data = if (comptime isInt(T)) self.data +| rhs_t.data else self.data + rhs_t.data };
const TargetType = Tensor(T, dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), shape_); const TargetType = Tensor(T, dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), shape_);
const l: Vec = if (comptime Self == TargetType) self.data else self.to(TargetType).data; const l: Vec = if (comptime Self == TargetType) self.data else self.to(TargetType).data;
const rr: Vec = blk: { const rr: Vec = blk: {
const RhsNorm = Tensor(T, RhsType.dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), RhsType.shape); const RhsNorm = Tensor(T, RhsType.dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), RhsType.shape);
const rn = if (comptime RhsType == RhsNorm) rhs_q else rhs_q.to(RhsNorm); const rn = if (comptime RhsType == RhsNorm) rhs_t else rhs_t.to(RhsNorm);
break :blk broadcastToVec(RhsNorm, rn); break :blk broadcastToVec(RhsNorm, rn);
}; };
return .{ .data = if (comptime isInt(T)) l +| rr else l + rr }; return .{ .data = if (comptime isInt(T)) l +| rr else l + rr };
@ -298,18 +298,20 @@ pub fn Tensor(
finerScales(Self, RhsT(@TypeOf(r))).argsOpt(), finerScales(Self, RhsT(@TypeOf(r))).argsOpt(),
shape_, shape_,
) { ) {
const rhs_q = rhs(r); const rhs_t = rhs(r);
const RhsType = @TypeOf(rhs_q); const RhsType = @TypeOf(rhs_t);
if (comptime !dims.eql(RhsType.dims)) if (comptime !dims.eql(RhsType.dims))
@compileError("Dimension mismatch in sub: " ++ dims.str() ++ " vs " ++ RhsType.dims.str()); @compileError("Dimension mismatch in sub: " ++ dims.str() ++ " vs " ++ RhsType.dims.str());
if (comptime RhsType.total != 1 and !shapeEql(shape_, RhsType.shape)) if (comptime RhsType.total != 1 and !shapeEql(shape_, RhsType.shape))
@compileError("Shape mismatch in sub: element-wise operations require identical shapes, or a scalar RHS."); @compileError("Shape mismatch in sub: element-wise operations require identical shapes, or a scalar RHS.");
if (comptime total == 1 and scales.eql(RhsType.scales)) // Here rhs_t has to be {1} too
return .{ .data = if (comptime isInt(T)) self.data -| rhs_t.data else self.data - rhs_t.data };
const TargetType = Tensor(T, dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), shape_); const TargetType = Tensor(T, dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), shape_);
const l: Vec = if (comptime Self == TargetType) self.data else self.to(TargetType).data; const l: Vec = if (comptime Self == TargetType) self.data else self.to(TargetType).data;
const rr: Vec = blk: { const rr: Vec = blk: {
const RhsNorm = Tensor(T, RhsType.dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), RhsType.shape); const RhsNorm = Tensor(T, RhsType.dims.argsOpt(), finerScales(Self, RhsType).argsOpt(), RhsType.shape);
const rn = if (comptime RhsType == RhsNorm) rhs_q else rhs_q.to(RhsNorm); const rn = if (comptime RhsType == RhsNorm) rhs_t else rhs_t.to(RhsNorm);
break :blk broadcastToVec(RhsNorm, rn); break :blk broadcastToVec(RhsNorm, rn);
}; };
return .{ .data = if (comptime isInt(T)) l -| rr else l - rr }; return .{ .data = if (comptime isInt(T)) l -| rr else l - rr };

View File

@ -26,11 +26,11 @@ pub fn main(init: std.process.Init) !void {
try bench_vsNative(&stdout_writer.interface); try bench_vsNative(&stdout_writer.interface);
try stdout_writer.flush(); try stdout_writer.flush();
// try bench_crossTypeVsNative(&stdout_writer.interface); // try bench_crossTypeVsNative(&stdout_writer.interface);
// try stdout_writer.flush(); try stdout_writer.flush();
// try bench_Vector(&stdout_writer.interface); try bench_Vector(&stdout_writer.interface);
// try stdout_writer.flush(); try stdout_writer.flush();
// try bench_HighDimTensor(&stdout_writer.interface); try bench_HighDimTensor(&stdout_writer.interface);
// try stdout_writer.flush(); try stdout_writer.flush();
} }
fn getTime() Io.Timestamp { fn getTime() Io.Timestamp {
@ -171,7 +171,7 @@ fn bench_Scalar(writer: *std.Io.Writer) !void {
fn bench_vsNative(writer: *std.Io.Writer) !void { fn bench_vsNative(writer: *std.Io.Writer) !void {
const ITERS: usize = 100_000; const ITERS: usize = 100_000;
const SAMPLES: usize = 5; const SAMPLES: usize = 100;
const getValT = struct { const getValT = struct {
fn f(comptime TT: type, i: usize) TT { fn f(comptime TT: type, i: usize) TT {
@ -180,8 +180,8 @@ fn bench_vsNative(writer: *std.Io.Writer) !void {
} }
}.f; }.f;
const Types = .{ i32, i64, i128, f32, f64 }; const Types = .{ f64, i64, i128, f32, f64 };
const TNames = .{ "i32", "i64", "i128", "f32", "f64" }; const TNames = .{ "f64", "i64", "i128", "f32", "f64" };
// Expanded Ops to match bench_Scalar // Expanded Ops to match bench_Scalar
const Ops = .{ "add", "sub", "mul", "div", "abs", "eq", "gt" }; const Ops = .{ "add", "sub", "mul", "div", "abs", "eq", "gt" };
@ -189,16 +189,17 @@ fn bench_vsNative(writer: *std.Io.Writer) !void {
\\ \\
\\ Scalar vs Native Overhead Analysis \\ Scalar vs Native Overhead Analysis
\\ \\
\\┌───────────┬──────┬───────────┬───────────┬─────────── \\┌───────────┬──────┬───────────┬───────────┬───────────┬───────────────────────
\\│ Operation │ Type │ Native │ Scalar │ Slowdown \\│ Operation │ Type │ Native │ @Vector │ Tensor{{1}} │ Slowdown Nat | Vec
\\├───────────┼──────┼───────────┼───────────┼─────────── \\├───────────┼──────┼───────────┼───────────┼───────────┼───────────────────────
\\ \\
, .{}); , .{});
inline for (Ops, 0..) |op_name, j| { inline for (Ops, 0..) |op_name, j| {
inline for (Types, 0..) |T, tidx| { inline for (Types, 0..) |T, tidx| {
var native_total_ns: f64 = 0; var native_total_ns: f64 = 0;
var quantity_total_ns: f64 = 0; var vector_total_ns: f64 = 0;
var tensor_total_ns: f64 = 0;
const M = Tensor(T, .{}, .{}, &.{1}); const M = Tensor(T, .{}, .{}, &.{1});
@ -230,6 +231,31 @@ fn bench_vsNative(writer: *std.Io.Writer) !void {
const n_end = getTime(); const n_end = getTime();
native_total_ns += @as(f64, @floatFromInt(n_start.durationTo(n_end).toNanoseconds())); native_total_ns += @as(f64, @floatFromInt(n_start.durationTo(n_end).toNanoseconds()));
const v_start = getTime();
const va = getValT(T, 10);
const vb = getValT(T, 2);
for (0..ITERS) |_| {
// Native logic branch
_ = if (comptime std.mem.eql(u8, op_name, "add"))
if (comptime @typeInfo(T) == .int) va +| vb else va + vb
else if (comptime std.mem.eql(u8, op_name, "sub"))
if (comptime @typeInfo(T) == .int) va -| vb else va - vb
else if (comptime std.mem.eql(u8, op_name, "mul"))
if (comptime @typeInfo(T) == .int) va *| vb else va * vb
else if (comptime std.mem.eql(u8, op_name, "div"))
if (comptime @typeInfo(T) == .int) @divTrunc(va, vb) else va / vb
else if (comptime std.mem.eql(u8, op_name, "abs"))
if (comptime @typeInfo(T) == .int) @abs(va) else @as(T, @abs(va))
else if (comptime std.mem.eql(u8, op_name, "eq"))
va == vb
else if (comptime std.mem.eql(u8, op_name, "gt"))
va > vb
else
unreachable;
}
const v_end = getTime();
vector_total_ns += @as(f64, @floatFromInt(v_start.durationTo(v_end).toNanoseconds()));
// --- 2. Benchmark Scalar --- // --- 2. Benchmark Scalar ---
const q_start = getTime(); const q_start = getTime();
const qa = M.splat(getValT(T, 10)); const qa = M.splat(getValT(T, 10));
@ -254,22 +280,24 @@ fn bench_vsNative(writer: *std.Io.Writer) !void {
unreachable; unreachable;
} }
const q_end = getTime(); const q_end = getTime();
quantity_total_ns += @as(f64, @floatFromInt(q_start.durationTo(q_end).toNanoseconds())); tensor_total_ns += @as(f64, @floatFromInt(q_start.durationTo(q_end).toNanoseconds()));
} }
}); });
const avg_n = (native_total_ns / SAMPLES) / @as(f64, @floatFromInt(ITERS)); const avg_n = (native_total_ns / SAMPLES) / @as(f64, @floatFromInt(ITERS));
const avg_q = (quantity_total_ns / SAMPLES) / @as(f64, @floatFromInt(ITERS)); const avg_v = (vector_total_ns / SAMPLES) / @as(f64, @floatFromInt(ITERS));
const slowdown = avg_q / avg_n; const avg_t = (tensor_total_ns / SAMPLES) / @as(f64, @floatFromInt(ITERS));
const slowdown_nt = avg_t / avg_n;
const slowdown_vt = avg_t / avg_v;
try writer.print("│ {s:<9} │ {s:<4} │ {d:>7.2}ns │ {d:>7.2}ns │ {d:>8.2}x │\n", .{ try writer.print("│ {s:<9} │ {s:<4} │ {d:>7.2}ns │ {d:>7.2}ns │ {d:>7.2}ns │ {d:>8.2}x {d:>8.2}x │\n", .{
op_name, TNames[tidx], avg_n, avg_q, slowdown, op_name, TNames[tidx], avg_n, avg_v, avg_t, slowdown_nt, slowdown_vt,
}); });
} }
if (j != Ops.len - 1) try writer.print("├───────────┼──────┼───────────┼───────────┼───────────\n", .{}); if (j != Ops.len - 1) try writer.print("├───────────┼──────┼───────────┼───────────┼───────────┼───────────────────────\n", .{});
} }
try writer.print("└───────────┴──────┴───────────┴───────────┴───────────\n", .{}); try writer.print("└───────────┴──────┴───────────┴───────────┴───────────┴───────────────────────\n", .{});
} }
fn bench_crossTypeVsNative(writer: *std.Io.Writer) !void { fn bench_crossTypeVsNative(writer: *std.Io.Writer) !void {