feat(fft): add interleaved C32 FFT APIs to t4a branch

.superpowers/sdd/task-1-fix-report.md

@@ -0,0 +1,23 @@
+# Task 1 review-fix report
+
+Status: DONE
+
+Review finding addressed:
+
+- Restored `sync_cube()` immediately after the final global-write loops in the
+  small/shared CFFT, RFFT, and IRFFT interleaved kernels.
+- The three insertion points match upstream commit
+  `e66fe2e60f9dfb4c77be5c68cc9979bbdbabc354` (CubeK PR #8); no algorithm,
+  threshold, layout, or launch-interface changes were made.
+
+Verification:
+
+- Targeted small CFFT round-trip test: passed.
+- Targeted RFFT reference test: passed.
+- Targeted IRFFT reference test: passed.
+- `cargo test -p cubek-fft --all-features`: passed (2 unit + 77 integration + doctests).
+- `cargo clippy -p cubek-fft --all-targets --all-features -- -D warnings`: passed.
+- `cargo fmt --all -- --check`: passed.
+- `git diff --check`: passed.
+
+Concerns: none.

Cargo.toml

@@ -5,6 +5,7 @@
 resolver = "2"
 
 members = [
+    "crates/cubek-fft",
     "crates/cubek-matmul",
     "crates/cubek-quant",
     "crates/cubek-random",
@@ -18,7 +19,6 @@ exclude = [
     "crates/cubek",
     "crates/cubek-attention",
     "crates/cubek-convolution",
-    "crates/cubek-fft",
     "crates/cubek-interpolate",
     "crates/cubek-reduce",
 ]
@@ -35,8 +35,8 @@ version = "0.2.0"
 # PR/CI builds use the tensor4all CubeCL commit that adds the `t4a-cubecl*`
 # package manifests. Published packages use the registry packages at the same
 # exact versions.
-cubecl = { package = "t4a-cubecl", git = "https://github.com/tensor4all/cubecl.git", rev = "6424d9da407d4bc10034aee3bae52705eb5d2da0", version = "=0.10.0", default-features = false }
-cubecl-common = { package = "t4a-cubecl-common", git = "https://github.com/tensor4all/cubecl.git", rev = "6424d9da407d4bc10034aee3bae52705eb5d2da0", version = "=0.10.0", default-features = false }
+cubecl = { package = "t4a-cubecl", git = "https://github.com/tensor4all/cubecl.git", rev = "11b52669f13e27bbe188f988fd696df6d989a562", version = "=0.10.0", default-features = false }
+cubecl-common = { package = "t4a-cubecl-common", git = "https://github.com/tensor4all/cubecl.git", rev = "11b52669f13e27bbe188f988fd696df6d989a562", version = "=0.10.0", default-features = false }
 
 derive-new = { version = "0.7.0", default-features = false }
 log = { default-features = false, version = "0.4.22" }

crates/cubek-attention/src/components/stage/unit/setup.rs

@@ -58,7 +58,7 @@ impl<SK: StageFamily, SV: StageFamily, SO: StageFamily<ReadWrite>> StageAttentio
                 CubeDimResource::Units(units * blueprint.tiling_scheme.stage_size.seq_q)
             }
             _ => {
-                return Err(AttentionSetupError::InvalidConfig(Box::new(
+                return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
                     "Error: Expected unit tile attention, got a plane tile attention".to_string(),
                 )));
             }

crates/cubek-attention/src/components/tile/attention.rs

@@ -184,7 +184,7 @@ fn validate_unit(
     let check_divisible =
         |dim: u32, vec_size: u32, name: &str, vec_name: &str| -> Result<(), AttentionSetupError> {
             if !dim.is_multiple_of(vec_size) {
-                return Err(AttentionSetupError::InvalidConfig(Box::new(format!(
+                return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(format!(
                     "Tile's {} ({:?}) must be divisible by {} vector size ({:?})",
                     name, dim, vec_name, vec_size
                 ))));
@@ -223,7 +223,7 @@ fn validate_blackbox(
     dtypes: &AttentionElems,
 ) -> Result<(), AttentionSetupError> {
     if dtypes.query_global != dtypes.query_tile {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Query global and tile types must be the same because no stage to cast in between",
         )));
     }
@@ -268,7 +268,7 @@ fn validate_blackbox(
     }
 
     if line_sizes_mask > 1 {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Line size mask > 1 not supported yet on accelerated tile attention",
         )));
     }
@@ -278,27 +278,27 @@ fn validate_blackbox(
     let softmax_total = softmax_num_rows * softmax_num_cols;
 
     if !softmax_total.is_multiple_of(cfg.plane_dim) {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Softmax size should be divisible by plane dim",
         )));
     }
 
     if cfg.inner_layout == InnerLayout::Contiguous && softmax_num_rows > cfg.plane_dim {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "More than one row per unit not supported with this inner layout",
         )));
     }
 
     if cfg.inner_layout == InnerLayout::SplitRows
         && !softmax_total.is_multiple_of(2 * cfg.plane_dim)
     {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "With split rows, units must have two elements each",
         )));
     }
 
     if cfg.tile_size.head_dim < cfg.tile_size.val_dim {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Can't have tile head_dim < tile val dim (not sure why)",
         )));
     }

crates/cubek-attention/src/routines/blackbox_accelerated.rs

@@ -151,13 +151,13 @@ fn blueprint<R: Runtime>(
             .map_err(map_err)?;
 
             if tile_size_score_matmul.m != values_matmul.m {
-                return Err(AttentionSetupError::InvalidConfig(Box::new(
+                return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
                     "Seq_q mismatch: `m` of score_matmul does not match `m` of values_matmul. ",
                 )));
             }
 
             if tile_size_score_matmul.n != values_matmul.k {
-                return Err(AttentionSetupError::InvalidConfig(Box::new(
+                return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
                     "Seq_kv mismatch: `n` of score_matmul does not match `k` of values_matmul. ",
                 )));
             }
@@ -208,21 +208,21 @@ fn validate(
     if !(problem.dims.seq_q as u32)
         .is_multiple_of(blueprint.tiling_scheme.elements_in_stage_seq_q())
     {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Stage seq_q must divide problem seq_q".to_string(),
         )));
     }
 
     if !(problem.dims.head_dim as u32).is_multiple_of(blueprint.tiling_scheme.tile_size.head_dim) {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Tile size head dim must divide problem head dim".to_string(),
         )));
     }
 
     if blueprint.tiling_scheme.partition_size.head_dim * blueprint.tiling_scheme.tile_size.head_dim
         != problem.dims.head_dim as u32
     {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(format!(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(format!(
             "Tiling scheme's total head dim ({}) does not match problem's head dim ({})",
             blueprint.tiling_scheme.partition_size.head_dim
                 * blueprint.tiling_scheme.tile_size.head_dim,

crates/cubek-attention/src/routines/unit.rs

@@ -68,7 +68,7 @@ impl Routine for UnitRoutine {
                 CubeDimResource::Units(units * blueprint.tiling_scheme.stage_size.seq_q)
             }
             _ => {
-                return Err(AttentionSetupError::InvalidConfig(Box::new(
+                return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
                     "Error: Expected unit tile attention, got a plane tile attention".to_string(),
                 )));
             }
@@ -136,15 +136,15 @@ fn validate(
     blueprint: AttentionBlueprint,
 ) -> Result<AttentionBlueprint, AttentionSetupError> {
     if !(problem.dims.head_dim as u32).is_multiple_of(blueprint.tiling_scheme.tile_size.head_dim) {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
             "Tile size head dim must divide problem head dim".to_string(),
         )));
     }
 
     if blueprint.tiling_scheme.partition_size.head_dim * blueprint.tiling_scheme.tile_size.head_dim
         != problem.dims.head_dim as u32
     {
-        return Err(AttentionSetupError::InvalidConfig(Box::new(format!(
+        return Err(AttentionSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(format!(
             "Tiling scheme's total head dim ({}) does not match problem's head dim ({})",
             blueprint.tiling_scheme.partition_size.head_dim
                 * blueprint.tiling_scheme.tile_size.head_dim,

crates/cubek-convolution/src/components/stage/reader.rs

@@ -47,7 +47,7 @@ impl TilingValidation for BiasTilingLayout {
     fn check(config: StageMemoryConfig) -> Result<(), InvalidConfigError> {
         let stage_width = config.elements_per_stage_along_col();
         if config.vector_size > stage_width {
-            return Err(Box::new(format!(
+            return Err(cubek_std::InvalidConfigError::new(format!(
                 "Invalid vector size. Got {:?} which should not be >{:?}",
                 config.vector_size, stage_width,
             )));

crates/cubek-convolution/src/kernels/backward_data/launch.rs

@@ -195,7 +195,7 @@ where
 /// rejected.
 #[allow(dead_code)]
 pub(crate) fn unsupported_tma_error() -> ConvSetupError {
-    ConvSetupError::Matmul(MatmulSetupError::InvalidConfig(Box::new(
+    ConvSetupError::Matmul(MatmulSetupError::InvalidConfig(cubek_std::InvalidConfigError::new(
         "Data backprop doesn't yet work with current TMA tiling strategy",
     )))
 }

crates/cubek-fft/Cargo.toml

@@ -23,12 +23,13 @@ cpu-reference = ["dep:num-complex", "dep:cubek-test-utils"]
 
 [dependencies]
 cubecl = { workspace = true }
-cubek-test-utils = { package = "t4a-cubek-test-utils", path = "./../cubek-test-utils/", version = "0.2.0", default-features = false, optional = true }
-num-complex = { workspace = true, optional = true }
+thiserror = { workspace = true }
+cubek-test-utils = { package = "t4a-cubek-test-utils", path = "./../cubek-test-utils/", version = "=0.2.0", default-features = false, optional = true }
+num-complex = { version = "0.4.6", optional = true }
 
 [dev-dependencies]
-num-complex = { workspace = true }
+num-complex = "0.4.6"
 cubecl = { workspace = true, features = ["test-runtime"] }
 cubecl-common = { workspace = true }
 cubek-fft = { path = ".", features = ["cpu-reference"] }
-cubek-test-utils = { package = "t4a-cubek-test-utils", path = "./../cubek-test-utils/", version = "0.2.0", default-features = false }
+cubek-test-utils = { package = "t4a-cubek-test-utils", path = "./../cubek-test-utils/", version = "=0.2.0", default-features = false }