ferrotensor 0.3.0__tar.gz → 0.5.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (77) hide show
  1. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/Cargo.lock +11 -23
  2. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/Cargo.toml +1 -1
  3. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/PKG-INFO +1 -1
  4. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/Cargo.toml +1 -1
  5. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/backend/compute.rs +20 -1
  6. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/backend/kernels.rs +338 -27
  7. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/backend/mod.rs +1 -1
  8. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/device.rs +202 -19
  9. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/functional.rs +107 -12
  10. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/random.rs +0 -24
  11. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/core.rs +10 -1
  12. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/cpu.rs +96 -0
  13. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/mod.rs +1 -0
  14. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/ops.rs +23 -49
  15. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/raw.rs +274 -1
  16. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/tests.rs +230 -0
  17. ferrotensor-0.5.0/ferrotensor/tests/manual_seed.rs +24 -0
  18. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/README.md +0 -0
  19. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/README.md +0 -0
  20. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/README.md +0 -0
  21. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/char_transformer.rs +0 -0
  22. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/data/sonnets.txt +0 -0
  23. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/gpu_matmul_attention.rs +0 -0
  24. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/mlp_two_moons.rs +0 -0
  25. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/quickstart.rs +0 -0
  26. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/examples/similarity_search.rs +0 -0
  27. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/data.rs +0 -0
  28. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/lib.rs +0 -0
  29. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/activation.rs +0 -0
  30. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/attention.rs +0 -0
  31. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/container.rs +0 -0
  32. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/dropout.rs +0 -0
  33. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/embedding.rs +0 -0
  34. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/linear.rs +0 -0
  35. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/mod.rs +0 -0
  36. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/module.rs +0 -0
  37. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/normalization.rs +0 -0
  38. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/nn/tests.rs +0 -0
  39. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/adam.rs +0 -0
  40. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/mod.rs +0 -0
  41. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/optimizer.rs +0 -0
  42. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/scheduler.rs +0 -0
  43. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/sgd.rs +0 -0
  44. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/optim/tests.rs +0 -0
  45. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/serialize.rs +0 -0
  46. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/autograd.rs +0 -0
  47. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor/src/tensor/shape.rs +0 -0
  48. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/.gitignore +0 -0
  49. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/Cargo.toml +0 -0
  50. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/README.md +0 -0
  51. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/build.rs +0 -0
  52. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/examples/mlp.py +0 -0
  53. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/device.rs +0 -0
  54. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/error.rs +0 -0
  55. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/functional.rs +0 -0
  56. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/lib.rs +0 -0
  57. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/nn.rs +0 -0
  58. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/optim.rs +0 -0
  59. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/serialize.rs +0 -0
  60. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/src/tensor.rs +0 -0
  61. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/conftest.py +0 -0
  62. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_autograd.py +0 -0
  63. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_errors.py +0 -0
  64. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_functional.py +0 -0
  65. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_gpu.py +0 -0
  66. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_nn.py +0 -0
  67. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_optim.py +0 -0
  68. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_serialize.py +0 -0
  69. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_tensor.py +0 -0
  70. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/ferrotensor-py/tests/test_train.py +0 -0
  71. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/pyproject.toml +0 -0
  72. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/__init__.py +0 -0
  73. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/_ferrotensor.pyi +0 -0
  74. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/nn/__init__.py +0 -0
  75. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/nn/functional.py +0 -0
  76. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/optim.py +0 -0
  77. {ferrotensor-0.3.0 → ferrotensor-0.5.0}/python/ferrotensor/py.typed +0 -0
@@ -606,7 +606,7 @@ dependencies = [
606
606
 
607
607
  [[package]]
608
608
  name = "ferrotensor"
609
- version = "0.3.0"
609
+ version = "0.5.0"
610
610
  dependencies = [
611
611
  "bytemuck",
612
612
  "naga",
@@ -619,7 +619,7 @@ dependencies = [
619
619
 
620
620
  [[package]]
621
621
  name = "ferrotensor-autograd"
622
- version = "0.3.0"
622
+ version = "0.5.0"
623
623
  dependencies = [
624
624
  "ferrotensor-backend",
625
625
  "pollster",
@@ -628,11 +628,10 @@ dependencies = [
628
628
 
629
629
  [[package]]
630
630
  name = "ferrotensor-backend"
631
- version = "0.3.0"
631
+ version = "0.5.0"
632
632
  dependencies = [
633
633
  "bytemuck",
634
634
  "env_logger",
635
- "futures-intrusive",
636
635
  "log",
637
636
  "naga",
638
637
  "pollster",
@@ -641,7 +640,7 @@ dependencies = [
641
640
 
642
641
  [[package]]
643
642
  name = "ferrotensor-benchmarks"
644
- version = "0.3.0"
643
+ version = "0.5.0"
645
644
  dependencies = [
646
645
  "ferrotensor",
647
646
  "ferrotensor-backend",
@@ -652,7 +651,7 @@ dependencies = [
652
651
 
653
652
  [[package]]
654
653
  name = "ferrotensor-cli"
655
- version = "0.3.0"
654
+ version = "0.5.0"
656
655
  dependencies = [
657
656
  "clap",
658
657
  "env_logger",
@@ -670,7 +669,7 @@ dependencies = [
670
669
 
671
670
  [[package]]
672
671
  name = "ferrotensor-graph"
673
- version = "0.3.0"
672
+ version = "0.5.0"
674
673
  dependencies = [
675
674
  "bytemuck",
676
675
  "env_logger",
@@ -685,7 +684,7 @@ dependencies = [
685
684
 
686
685
  [[package]]
687
686
  name = "ferrotensor-hub"
688
- version = "0.3.0"
687
+ version = "0.5.0"
689
688
  dependencies = [
690
689
  "ferrotensor-autograd",
691
690
  "ferrotensor-backend",
@@ -698,7 +697,7 @@ dependencies = [
698
697
 
699
698
  [[package]]
700
699
  name = "ferrotensor-loader"
701
- version = "0.3.0"
700
+ version = "0.5.0"
702
701
  dependencies = [
703
702
  "bytemuck",
704
703
  "env_logger",
@@ -716,7 +715,7 @@ dependencies = [
716
715
 
717
716
  [[package]]
718
717
  name = "ferrotensor-monitor"
719
- version = "0.3.0"
718
+ version = "0.5.0"
720
719
  dependencies = [
721
720
  "anyhow",
722
721
  "chrono",
@@ -730,7 +729,7 @@ dependencies = [
730
729
 
731
730
  [[package]]
732
731
  name = "ferrotensor-py"
733
- version = "0.3.0"
732
+ version = "0.5.0"
734
733
  dependencies = [
735
734
  "ferrotensor",
736
735
  "pyo3",
@@ -739,7 +738,7 @@ dependencies = [
739
738
 
740
739
  [[package]]
741
740
  name = "ferrotensor-trainer"
742
- version = "0.3.0"
741
+ version = "0.5.0"
743
742
  dependencies = [
744
743
  "bytemuck",
745
744
  "ferrotensor-autograd",
@@ -805,17 +804,6 @@ version = "0.3.31"
805
804
  source = "registry+https://github.com/rust-lang/crates.io-index"
806
805
  checksum = "05f29059c0c2090612e8d742178b0580d2dc940c837851ad723096f87af6663e"
807
806
 
808
- [[package]]
809
- name = "futures-intrusive"
810
- version = "0.5.0"
811
- source = "registry+https://github.com/rust-lang/crates.io-index"
812
- checksum = "1d930c203dd0b6ff06e0201a4a2fe9149b43c684fd4420555b26d21b1a02956f"
813
- dependencies = [
814
- "futures-core",
815
- "lock_api",
816
- "parking_lot",
817
- ]
818
-
819
807
  [[package]]
820
808
  name = "futures-task"
821
809
  version = "0.3.31"
@@ -4,7 +4,7 @@ resolver = "2"
4
4
 
5
5
  [workspace.package]
6
6
  edition = "2021"
7
- version = "0.3.0"
7
+ version = "0.5.0"
8
8
  license = "MIT OR Apache-2.0"
9
9
  repository = "https://github.com/puranikyashaswin/ferrotensor"
10
10
  authors = ["Yashaswin Puranik <puranikyashaswinsharma@gmail.com>"]
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: ferrotensor
3
- Version: 0.3.0
3
+ Version: 0.5.0
4
4
  Classifier: Programming Language :: Rust
5
5
  Classifier: Programming Language :: Python :: 3
6
6
  Classifier: Programming Language :: Python :: Implementation :: CPython
@@ -1,6 +1,6 @@
1
1
  [package]
2
2
  name = "ferrotensor"
3
- version = "0.3.0"
3
+ version = "0.5.0"
4
4
  edition = "2021"
5
5
  authors.workspace = true
6
6
  license.workspace = true
@@ -9,7 +9,7 @@ use wgpu::{BindGroup, BindGroupLayout, BindingResource, CommandEncoder, ComputeP
9
9
  use super::KERNELS_WGSL;
10
10
 
11
11
  /// Entry points compiled from [`KERNELS_WGSL`].
12
- pub(crate) const KERNEL_NAMES: [&str; 9] = [
12
+ pub(crate) const KERNEL_NAMES: [&str; 15] = [
13
13
  "unary",
14
14
  "binary",
15
15
  "reduce",
@@ -19,8 +19,19 @@ pub(crate) const KERNEL_NAMES: [&str; 9] = [
19
19
  "matmul",
20
20
  "permute",
21
21
  "gather_rows",
22
+ "sort_ids",
23
+ "scatter_rows",
24
+ "copy_block",
25
+ "cross_entropy_rows",
26
+ "cross_entropy_grad",
27
+ "fused",
22
28
  ];
23
29
 
30
+ /// Storage buffers the `fused` kernel binds: seven tensor slots and its
31
+ /// program. Devices with fewer per-stage bindings do not compile it, and
32
+ /// element-wise ops then run unfused.
33
+ pub(crate) const FUSED_STORAGE_BINDINGS: u32 = 8;
34
+
24
35
  /// Threads per workgroup of the 1-D kernels.
25
36
  const WORKGROUP_1D: u32 = 256;
26
37
  /// Largest workgroup count allowed in one grid dimension.
@@ -47,8 +58,11 @@ impl ComputeEngine {
47
58
  source: wgpu::ShaderSource::Wgsl(Cow::Borrowed(KERNELS_WGSL)),
48
59
  });
49
60
 
61
+ let fused_fits =
62
+ device.limits().max_storage_buffers_per_shader_stage >= FUSED_STORAGE_BINDINGS;
50
63
  let pipelines = KERNEL_NAMES
51
64
  .iter()
65
+ .filter(|&&name| name != "fused" || fused_fits)
52
66
  .map(|&name| {
53
67
  let pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
54
68
  label: Some(name),
@@ -64,6 +78,11 @@ impl ComputeEngine {
64
78
  Self { pipelines }
65
79
  }
66
80
 
81
+ /// Whether `name` was compiled for this device.
82
+ pub(crate) fn has(&self, name: &str) -> bool {
83
+ self.pipelines.contains_key(name)
84
+ }
85
+
67
86
  /// Get a compiled pipeline by entry point name.
68
87
  pub fn get_pipeline(&self, name: &str) -> Option<&ComputePipeline> {
69
88
  self.pipelines.get(name).map(|(pipeline, _)| pipeline)
@@ -100,16 +100,9 @@ struct UnaryParams {
100
100
  @group(0) @binding(1) var<storage, read_write> un_out: array<f32>;
101
101
  @group(0) @binding(2) var<uniform> un_p: UnaryParams;
102
102
 
103
- @compute @workgroup_size(256)
104
- fn unary(@builtin(global_invocation_id) gid: vec3<u32>,
105
- @builtin(num_workgroups) nwg: vec3<u32>) {
106
- let i = flat_index(gid, nwg);
107
- if (i >= un_p.n) { return; }
108
- let x = un_in[i];
109
- let a = un_p.a;
110
- let b = un_p.b;
103
+ fn unary_apply(op: u32, x: f32, a: f32, b: f32) -> f32 {
111
104
  var y = x;
112
- switch un_p.op {
105
+ switch op {
113
106
  case 0u: { y = max(x, 0.0); }
114
107
  case 1u: { y = x * sigmoid_f(x); }
115
108
  case 2u: { y = gelu_tanh(x); }
@@ -152,7 +145,15 @@ fn unary(@builtin(global_invocation_id) gid: vec3<u32>,
152
145
  case 25u: { y = select(a, 1.0, x > 0.0); }
153
146
  default: { y = x; }
154
147
  }
155
- un_out[i] = y;
148
+ return y;
149
+ }
150
+
151
+ @compute @workgroup_size(256)
152
+ fn unary(@builtin(global_invocation_id) gid: vec3<u32>,
153
+ @builtin(num_workgroups) nwg: vec3<u32>) {
154
+ let i = flat_index(gid, nwg);
155
+ if (i >= un_p.n) { return; }
156
+ un_out[i] = unary_apply(un_p.op, un_in[i], un_p.a, un_p.b);
156
157
  }
157
158
 
158
159
  // --- Broadcasting binary ops ------------------------------------------------
@@ -172,6 +173,25 @@ struct BinaryParams {
172
173
  @group(0) @binding(3) var<uniform> bin_p: BinaryParams;
173
174
  @group(0) @binding(4) var<storage, read> bin_meta: array<u32>;
174
175
 
176
+ fn binary_apply(op: u32, x: f32, y: f32) -> f32 {
177
+ var r = 0.0;
178
+ switch op {
179
+ case 0u: { r = x + y; }
180
+ case 1u: { r = x - y; }
181
+ case 2u: { r = x * y; }
182
+ case 3u: { r = x / y; }
183
+ case 4u: { r = max(x, y); }
184
+ case 5u: { r = min(x, y); }
185
+ case 6u: { r = select(0.0, 1.0, x > y); }
186
+ case 7u: { r = select(0.0, 1.0, x < y); }
187
+ case 8u: { r = select(0.0, 1.0, x == y); }
188
+ case 9u: { r = select(0.0, 1.0, x >= y); }
189
+ case 10u: { r = select(0.0, 1.0, x <= y); }
190
+ default: { r = x; }
191
+ }
192
+ return r;
193
+ }
194
+
175
195
  @compute @workgroup_size(256)
176
196
  fn binary(@builtin(global_invocation_id) gid: vec3<u32>,
177
197
  @builtin(num_workgroups) nwg: vec3<u32>) {
@@ -189,24 +209,78 @@ fn binary(@builtin(global_invocation_id) gid: vec3<u32>,
189
209
  ia = ia + idx * bin_meta[nd + axis];
190
210
  ib = ib + idx * bin_meta[2u * nd + axis];
191
211
  }
192
- let x = bin_a[ia];
193
- let y = bin_b[ib];
194
- var r = 0.0;
195
- switch bin_p.op {
196
- case 0u: { r = x + y; }
197
- case 1u: { r = x - y; }
198
- case 2u: { r = x * y; }
199
- case 3u: { r = x / y; }
200
- case 4u: { r = max(x, y); }
201
- case 5u: { r = min(x, y); }
202
- case 6u: { r = select(0.0, 1.0, x > y); }
203
- case 7u: { r = select(0.0, 1.0, x < y); }
204
- case 8u: { r = select(0.0, 1.0, x == y); }
205
- case 9u: { r = select(0.0, 1.0, x >= y); }
206
- case 10u: { r = select(0.0, 1.0, x <= y); }
207
- default: { r = x; }
212
+ bin_out[i] = binary_apply(bin_p.op, bin_a[ia], bin_b[ib]);
213
+ }
214
+
215
+ // --- Fused element-wise chains ---------------------------------------------
216
+ // A chain of same-size unary and non-broadcasting binary ops in one dispatch
217
+ // (FERROTENSOR_FUSION=1, see docs/design/op-fusion.md). Each step is 8 words:
218
+ // [kind, op, src0, src1, dst, a bits, b bits, 0]; kind 0 applies a unary op
219
+ // to src0, kind 1 a binary op to src0 and src1. Sources and destinations are
220
+ // binding slots 0..6. Every step stores its result and later steps load it
221
+ // back, as separate dispatches would.
222
+
223
+ struct FusedParams {
224
+ n: u32,
225
+ steps: u32,
226
+ _pad0: u32,
227
+ _pad1: u32,
228
+ };
229
+
230
+ @group(0) @binding(0) var<storage, read_write> fb0: array<f32>;
231
+ @group(0) @binding(1) var<storage, read_write> fb1: array<f32>;
232
+ @group(0) @binding(2) var<storage, read_write> fb2: array<f32>;
233
+ @group(0) @binding(3) var<storage, read_write> fb3: array<f32>;
234
+ @group(0) @binding(4) var<storage, read_write> fb4: array<f32>;
235
+ @group(0) @binding(5) var<storage, read_write> fb5: array<f32>;
236
+ @group(0) @binding(6) var<storage, read_write> fb6: array<f32>;
237
+ @group(0) @binding(7) var<storage, read> fused_prog: array<u32>;
238
+ @group(0) @binding(8) var<uniform> fused_p: FusedParams;
239
+
240
+ fn fused_load(slot: u32, i: u32) -> f32 {
241
+ var v = 0.0;
242
+ switch slot {
243
+ case 0u: { v = fb0[i]; }
244
+ case 1u: { v = fb1[i]; }
245
+ case 2u: { v = fb2[i]; }
246
+ case 3u: { v = fb3[i]; }
247
+ case 4u: { v = fb4[i]; }
248
+ case 5u: { v = fb5[i]; }
249
+ default: { v = fb6[i]; }
250
+ }
251
+ return v;
252
+ }
253
+
254
+ fn fused_store(slot: u32, i: u32, v: f32) {
255
+ switch slot {
256
+ case 0u: { fb0[i] = v; }
257
+ case 1u: { fb1[i] = v; }
258
+ case 2u: { fb2[i] = v; }
259
+ case 3u: { fb3[i] = v; }
260
+ case 4u: { fb4[i] = v; }
261
+ case 5u: { fb5[i] = v; }
262
+ default: { fb6[i] = v; }
263
+ }
264
+ }
265
+
266
+ @compute @workgroup_size(256)
267
+ fn fused(@builtin(global_invocation_id) gid: vec3<u32>,
268
+ @builtin(num_workgroups) nwg: vec3<u32>) {
269
+ let i = flat_index(gid, nwg);
270
+ if (i >= fused_p.n) { return; }
271
+ for (var k = 0u; k < fused_p.steps; k = k + 1u) {
272
+ let w = k * 8u;
273
+ let x = fused_load(fused_prog[w + 2u], i);
274
+ var r = 0.0;
275
+ if (fused_prog[w] == 0u) {
276
+ let a = bitcast<f32>(fused_prog[w + 5u]);
277
+ let b = bitcast<f32>(fused_prog[w + 6u]);
278
+ r = unary_apply(fused_prog[w + 1u], x, a, b);
279
+ } else {
280
+ r = binary_apply(fused_prog[w + 1u], x, fused_load(fused_prog[w + 3u], i));
281
+ }
282
+ fused_store(fused_prog[w + 4u], i, r);
208
283
  }
209
- bin_out[i] = r;
210
284
  }
211
285
 
212
286
  // --- Axis reductions --------------------------------------------------------
@@ -547,4 +621,241 @@ fn gather_rows(@builtin(global_invocation_id) gid: vec3<u32>,
547
621
  }
548
622
  gat_out[i] = gat_table[id * gat_p.row + c];
549
623
  }
624
+
625
+ // --- Row scatter-add (index_select and embedding backward) -------------------
626
+ // WGSL has no float atomics, so the gradient rows are not added with atomics.
627
+ // Instead the (row, position) pairs of the ids are sorted by row, then by
628
+ // position; head[row] is the index of the row's first pair, or SORT_NONE when
629
+ // no id picks it. One thread per output element adds its row's gradients in
630
+ // position order, as the CPU loop does, so the result is deterministic and
631
+ // matches the CPU bit for bit.
632
+
633
+ const SORT_NONE: u32 = 0xffffffffu;
634
+
635
+ // sort_ids builds the pairs and the head table on the GPU from f32 ids.
636
+ // op 0: pair i = (row of id i, i) for i < n, padded with (SORT_NONE, i) up to
637
+ // a = the power-of-two pair count; b = rows. An id outside [0, rows)
638
+ // gets row SORT_NONE: it sorts last and adds nothing.
639
+ // op 1: one compare-and-swap step of a bitonic sort of n pairs (a = k, b = j).
640
+ // op 2: head[r] = SORT_NONE for r < n.
641
+ // op 3: head[row] = i for each pair i < n that is the first of its row.
642
+
643
+ struct SortParams {
644
+ op: u32,
645
+ n: u32,
646
+ a: u32,
647
+ b: u32,
648
+ };
649
+
650
+ @group(0) @binding(0) var<storage, read> srt_ids: array<f32>;
651
+ @group(0) @binding(1) var<storage, read_write> srt_keys: array<u32>;
652
+ @group(0) @binding(2) var<storage, read_write> srt_head: array<u32>;
653
+ @group(0) @binding(3) var<uniform> srt_p: SortParams;
654
+
655
+ @compute @workgroup_size(256)
656
+ fn sort_ids(@builtin(global_invocation_id) gid: vec3<u32>,
657
+ @builtin(num_workgroups) nwg: vec3<u32>) {
658
+ let i = flat_index(gid, nwg);
659
+ switch srt_p.op {
660
+ case 0u: {
661
+ if (i >= srt_p.a) { return; }
662
+ var row = SORT_NONE;
663
+ if (i < srt_p.n) {
664
+ let f = srt_ids[i];
665
+ if (f >= 0.0 && f < f32(srt_p.b)) { row = u32(f); }
666
+ }
667
+ srt_keys[2u * i] = row;
668
+ srt_keys[2u * i + 1u] = i;
669
+ }
670
+ case 1u: {
671
+ let l = i ^ srt_p.b;
672
+ if (i >= srt_p.n || l <= i) { return; }
673
+ let ri = srt_keys[2u * i];
674
+ let pi = srt_keys[2u * i + 1u];
675
+ let rl = srt_keys[2u * l];
676
+ let pl = srt_keys[2u * l + 1u];
677
+ let greater = ri > rl || (ri == rl && pi > pl);
678
+ if (greater == ((i & srt_p.a) == 0u)) {
679
+ srt_keys[2u * i] = rl;
680
+ srt_keys[2u * i + 1u] = pl;
681
+ srt_keys[2u * l] = ri;
682
+ srt_keys[2u * l + 1u] = pi;
683
+ }
684
+ }
685
+ case 2u: {
686
+ if (i >= srt_p.n) { return; }
687
+ srt_head[i] = SORT_NONE;
688
+ }
689
+ default: {
690
+ if (i >= srt_p.n) { return; }
691
+ let row = srt_keys[2u * i];
692
+ if (row == SORT_NONE) { return; }
693
+ if (i == 0u || srt_keys[2u * i - 2u] != row) {
694
+ srt_head[row] = i;
695
+ }
696
+ }
697
+ }
698
+ }
699
+
700
+ // The gradient is viewed as [outer, n, row] and the output as
701
+ // [outer, rows, row].
702
+ struct ScatterParams {
703
+ rows: u32,
704
+ row: u32,
705
+ n: u32,
706
+ outer: u32,
707
+ };
708
+
709
+ @group(0) @binding(0) var<storage, read> sct_grad: array<f32>;
710
+ @group(0) @binding(1) var<storage, read> sct_keys: array<u32>;
711
+ @group(0) @binding(2) var<storage, read> sct_head: array<u32>;
712
+ @group(0) @binding(3) var<storage, read_write> sct_out: array<f32>;
713
+ @group(0) @binding(4) var<uniform> sct_p: ScatterParams;
714
+
715
+ @compute @workgroup_size(256)
716
+ fn scatter_rows(@builtin(global_invocation_id) gid: vec3<u32>,
717
+ @builtin(num_workgroups) nwg: vec3<u32>) {
718
+ let i = flat_index(gid, nwg);
719
+ if (i >= sct_p.outer * sct_p.rows * sct_p.row) { return; }
720
+ let o = i / (sct_p.rows * sct_p.row);
721
+ let r = (i / sct_p.row) % sct_p.rows;
722
+ let c = i % sct_p.row;
723
+ let base = o * sct_p.n * sct_p.row + c;
724
+ var acc = 0.0;
725
+ var t = sct_head[r];
726
+ if (t != SORT_NONE) {
727
+ loop {
728
+ if (t >= sct_p.n || sct_keys[2u * t] != r) { break; }
729
+ acc = acc + sct_grad[base + sct_keys[2u * t + 1u] * sct_p.row];
730
+ t = t + 1u;
731
+ }
732
+ }
733
+ sct_out[i] = acc;
734
+ }
735
+
736
+ // --- Block copy (cat) --------------------------------------------------------
737
+ // Copies an input viewed as [outer, block_in] into columns
738
+ // [offset, offset + block_in) of an output viewed as [outer, block_out].
739
+
740
+ struct CopyParams {
741
+ n: u32,
742
+ block_in: u32,
743
+ block_out: u32,
744
+ offset: u32,
745
+ };
746
+
747
+ @group(0) @binding(0) var<storage, read> cpy_in: array<f32>;
748
+ @group(0) @binding(1) var<storage, read_write> cpy_out: array<f32>;
749
+ @group(0) @binding(2) var<uniform> cpy_p: CopyParams;
750
+
751
+ @compute @workgroup_size(256)
752
+ fn copy_block(@builtin(global_invocation_id) gid: vec3<u32>,
753
+ @builtin(num_workgroups) nwg: vec3<u32>) {
754
+ let i = flat_index(gid, nwg);
755
+ if (i >= cpy_p.n) { return; }
756
+ let o = i / cpy_p.block_in;
757
+ cpy_out[o * cpy_p.block_out + cpy_p.offset + i % cpy_p.block_in] = cpy_in[i];
758
+ }
759
+
760
+ // --- Cross-entropy with class-id targets ---------------------------------------
761
+ // Logits are viewed as [rows, classes]; targets hold one f32 class id per row.
762
+ // A target <= -1 (an id that is negative as an integer, such as -100) skips
763
+ // its row: loss 0, valid 0, gradient 0. Otherwise the id is truncated to an
764
+ // integer; one that is not below `classes` (or NaN) gives a NaN loss.
765
+ // cross_entropy_rows: one workgroup per row writes the row's loss
766
+ // (lse - logit[target]), its log-sum-exp and whether it counts.
767
+ // cross_entropy_grad: d loss[r] / d logit[r, c] = softmax[r, c] - [c == target],
768
+ // times the incoming gradient of the row's loss.
769
+
770
+ struct CeParams {
771
+ rows: u32,
772
+ classes: u32,
773
+ _pad0: u32,
774
+ _pad1: u32,
775
+ };
776
+
777
+ @group(0) @binding(0) var<storage, read> ce_logits: array<f32>;
778
+ @group(0) @binding(1) var<storage, read> ce_targets: array<f32>;
779
+ @group(0) @binding(2) var<storage, read_write> ce_loss: array<f32>;
780
+ @group(0) @binding(3) var<storage, read_write> ce_lse: array<f32>;
781
+ @group(0) @binding(4) var<storage, read_write> ce_valid: array<f32>;
782
+ @group(0) @binding(5) var<uniform> ce_p: CeParams;
783
+
784
+ var<workgroup> ce_part: array<f32, 256>;
785
+
786
+ @compute @workgroup_size(256)
787
+ fn cross_entropy_rows(@builtin(local_invocation_index) lid: u32,
788
+ @builtin(workgroup_id) wid: vec3<u32>,
789
+ @builtin(num_workgroups) nwg: vec3<u32>) {
790
+ let r = wid.x + wid.y * nwg.x;
791
+ // Uniform per workgroup, so every thread reaches the barriers or none.
792
+ if (r >= ce_p.rows) { return; }
793
+ let c = ce_p.classes;
794
+ let base = r * c;
795
+ var m = bitcast<f32>(0xff800000u);
796
+ for (var j = lid; j < c; j = j + 256u) {
797
+ m = max(m, ce_logits[base + j]);
798
+ }
799
+ ce_part[lid] = m;
800
+ workgroupBarrier();
801
+ for (var half = 128u; half > 0u; half = half >> 1u) {
802
+ if (lid < half) {
803
+ ce_part[lid] = max(ce_part[lid], ce_part[lid + half]);
804
+ }
805
+ workgroupBarrier();
806
+ }
807
+ m = ce_part[0];
808
+ workgroupBarrier();
809
+ var s = 0.0;
810
+ for (var j = lid; j < c; j = j + 256u) {
811
+ s = s + exp(ce_logits[base + j] - m);
812
+ }
813
+ ce_part[lid] = s;
814
+ workgroupBarrier();
815
+ for (var half = 128u; half > 0u; half = half >> 1u) {
816
+ if (lid < half) {
817
+ ce_part[lid] = ce_part[lid] + ce_part[lid + half];
818
+ }
819
+ workgroupBarrier();
820
+ }
821
+ if (lid == 0u) {
822
+ let lse = m + log(ce_part[0]);
823
+ let f = ce_targets[r];
824
+ ce_lse[r] = lse;
825
+ if (f <= -1.0) {
826
+ ce_loss[r] = 0.0;
827
+ ce_valid[r] = 0.0;
828
+ } else {
829
+ ce_valid[r] = 1.0;
830
+ if (f < f32(c)) {
831
+ ce_loss[r] = lse - ce_logits[base + u32(max(f, 0.0))];
832
+ } else {
833
+ ce_loss[r] = bitcast<f32>(0x7fc00000u);
834
+ }
835
+ }
836
+ }
837
+ }
838
+
839
+ @group(0) @binding(0) var<storage, read> ceg_logits: array<f32>;
840
+ @group(0) @binding(1) var<storage, read> ceg_targets: array<f32>;
841
+ @group(0) @binding(2) var<storage, read> ceg_lse: array<f32>;
842
+ @group(0) @binding(3) var<storage, read> ceg_grad: array<f32>;
843
+ @group(0) @binding(4) var<storage, read_write> ceg_out: array<f32>;
844
+ @group(0) @binding(5) var<uniform> ceg_p: CeParams;
845
+
846
+ @compute @workgroup_size(256)
847
+ fn cross_entropy_grad(@builtin(global_invocation_id) gid: vec3<u32>,
848
+ @builtin(num_workgroups) nwg: vec3<u32>) {
849
+ let i = flat_index(gid, nwg);
850
+ if (i >= ceg_p.rows * ceg_p.classes) { return; }
851
+ let r = i / ceg_p.classes;
852
+ let f = ceg_targets[r];
853
+ if (f <= -1.0) {
854
+ ceg_out[i] = 0.0;
855
+ return;
856
+ }
857
+ let p = exp(ceg_logits[i] - ceg_lse[r]);
858
+ let hit = f < f32(ceg_p.classes) && i % ceg_p.classes == u32(max(f, 0.0));
859
+ ceg_out[i] = (p - select(0.0, 1.0, hit)) * ceg_grad[r];
860
+ }
550
861
  "#;
@@ -7,5 +7,5 @@ mod compute;
7
7
  mod kernels;
8
8
 
9
9
  pub use compute::ComputeEngine;
10
- pub(crate) use compute::Dispatch;
10
+ pub(crate) use compute::{Dispatch, FUSED_STORAGE_BINDINGS};
11
11
  pub(crate) use kernels::{binary_op, reduce_op, unary_op, KERNELS_WGSL};