2222from lighthouse .schedule import schedule_boilerplate
2323from lighthouse .dialects .transform .transform_ext import (
2424 replace_with_fused_attention ,
25- update_address_space ,
2625)
2726
2827
@@ -31,14 +30,15 @@ def fused_attention_schedule(
3130 parameters : Optional [dict ] = None ,
3231) -> ir .Module :
3332 """
34- Generate transform schedule for fused attention operation .
33+ Generate transform schedule for attention kernel .
3534
3635 The schedule performs the following transformations:
37- 1. Tile the fused attention operation
36+ 1. Tile the fuse the strandard attention computation along parallel dims
3837 2. Vectorize operations
3938 3. Bufferize tensors
40- 4. Convert to GPU dialect
41- 5. Lower to XeGPU operations
39+ 4. Perform the fused attention optimization for the innermost computation
40+ 5. Convert to GPU dialect
41+ 6. Lower to XeGPU operations
4242
4343 Args:
4444 stop_at_stage: Optional stage name to stop early (for debugging)
@@ -86,7 +86,7 @@ def bundle_xegpu_fused_attention_schedule(
8686 parameters : dict ,
8787 stop_at_stage : str = "" ,
8888) -> ir .Value [transform .AnyOpType ]:
89- """Schedule for lowering fused attention payload to xegpu wg level."""
89+ """Schedule for lowering attention payload to xegpu wg level."""
9090
9191 if stop_at_stage == "initial" :
9292 raise PipelineInterrupt ()
@@ -165,6 +165,7 @@ def bundle_xegpu_fused_attention_schedule(
165165 transform .apply_cse (func )
166166 canonicalize (func )
167167
168+ # Fuse the remaining operations into the scf.forall loop.
168169 linalg_mul_op = match_and_split (func , ops = {"linalg.mul" }, nhandles = 1 )[0 ]
169170 first_matmul = transform .get_producer_of_operand (
170171 anytype , linalg_mul_op , operand_number = 0
@@ -197,7 +198,7 @@ def bundle_xegpu_fused_attention_schedule(
197198 if stop_at_stage == "outer-tiled" :
198199 raise PipelineInterrupt ()
199200
200- # vectorize
201+ # Vectorize
201202 func = structured .VectorizeChildrenAndApplyPatternsOp (
202203 func ,
203204 fold_type_extensions_into_contract = True ,
@@ -212,7 +213,7 @@ def bundle_xegpu_fused_attention_schedule(
212213 if stop_at_stage == "vectorized" :
213214 raise PipelineInterrupt ()
214215
215- # bufferize
216+ # Bufferize
216217 mod = apply_registered_pass (mod , "eliminate-empty-tensors" )
217218 identity_layout = LayoutMapOption .IdentityLayoutMap
218219 mod = transform_bufferization .OneShotBufferizeOp (
@@ -221,23 +222,13 @@ def bundle_xegpu_fused_attention_schedule(
221222 bufferize_function_boundaries = True ,
222223 function_boundary_type_conversion = identity_layout ,
223224 ).result
224- transform .apply_cse (mod )
225- canonicalize (mod )
226225 # fold memref.subviews into vector.transfer_read/write ops
227226 mod = apply_registered_pass (mod , "fold-memref-alias-ops" )
228227 transform .apply_cse (mod )
229228 canonicalize (mod )
230229
231- # promote memref.alloc to memref.alloca in payload function
232- func = match (mod , ops = {"func.func" })
233- func = apply_registered_pass (
234- func ,
235- "promote-buffers-to-stack" ,
236- options = {
237- "max-alloc-size-in-bytes" : "8192" ,
238- "max-rank-of-allocated-memref" : "2" ,
239- },
240- )
230+ if stop_at_stage == "bufferized" :
231+ raise PipelineInterrupt ()
241232
242233 # Extract q, k, v memrefs from the bufferized IR
243234 # Match vector.contract ops to find the q, k, v loads
@@ -256,8 +247,8 @@ def bundle_xegpu_fused_attention_schedule(
256247 anytype , first_contract , operand_number = 1
257248 )
258249
259- # # Second vector.contract is attention_weights @ V
260- # # Its second operand is the v load (vector.transfer_read)
250+ # Second vector.contract is attention_weights @ V
251+ # Its second operand is the v load (vector.transfer_read)
261252 second_contract = contract_ops [1 ]
262253 v_load = transform .get_producer_of_operand (
263254 anytype , second_contract , operand_number = 1
@@ -268,12 +259,9 @@ def bundle_xegpu_fused_attention_schedule(
268259 mulf_op = match_and_split (func , ops = {"arith.mulf" }, nhandles = 1 )[0 ]
269260 scale = transform .get_producer_of_operand (anytype , mulf_op , operand_number = 1 )
270261
271- if stop_at_stage == "bufferized" :
272- raise PipelineInterrupt ()
273-
274- # Generate fused attention computation with inner tiling
275- # This replaces the second vector.contract (attention_weights @ V) with a tiled
276- # loop that implements online softmax for efficient memory usage
262+ # Apply the fused attention optimization. This replaces the second vector.contract
263+ # (attention_weights @ V) with a tiled loop that implements online softmax for
264+ # efficient memory usage
277265 tile_size = parameters .get (
278266 "inner_loop_tile_size" , 64
279267 ) # Tile size for reduction dimension (K/V sequence length)
@@ -291,20 +279,20 @@ def bundle_xegpu_fused_attention_schedule(
291279 if stop_at_stage == "inner-tiled" :
292280 raise PipelineInterrupt ()
293281
294- # convert forall to parallel
282+ # Convert forall to parallel
295283 wg_loops = match_and_split (mod , ops = {"scf.forall" })
296284 for wg_loop in wg_loops :
297285 wg_loop = loop .loop_forall_to_parallel ([anytype ], wg_loop )
298286 func = transform .get_parent_op (anytype , wg_loop )
299287
300- # convert scf.parallel to gpu.launch
288+ # Convert scf.parallel to gpu.launch
301289 func = apply_registered_pass (func , "gpu-map-parallel-loops" )
302290 func = apply_registered_pass (func , "convert-parallel-loops-to-gpu" )
303291 func = apply_registered_pass (func , "lower-affine" )
304292 transform .apply_cse (func )
305293 canonicalize (func )
306294
307- # set the number of threads for the gpu.launch operation
295+ # Set the number of threads for the gpu.launch operation
308296 launch_op = match_and_split (func , ops = {"gpu.launch" })
309297 wg_rows = parameters ["wg_rows" ]
310298 sg_rows = parameters ["sg_rows" ]
@@ -313,7 +301,7 @@ def bundle_xegpu_fused_attention_schedule(
313301 num_threads = num_subgroups * subgroup_size
314302 xegpu .set_gpu_launch_threads (launch_op [0 ], threads = [num_threads , 1 , 1 ])
315303
316- # outline gpu func
304+ # Outline gpu func
317305 func = apply_registered_pass (func , "lower-affine" )
318306 canonicalize (func )
319307 func = apply_registered_pass (func , "gpu-launch-sink-index-computations" )
@@ -323,22 +311,17 @@ def bundle_xegpu_fused_attention_schedule(
323311 if stop_at_stage == "gpu-outlining" :
324312 raise PipelineInterrupt ()
325313
326- # set xevm target
314+ # Set xevm target
327315 mod = apply_registered_pass (
328316 mod ,
329317 "xevm-attach-target" ,
330318 options = {"O" : "3" , "chip" : "bmg" },
331319 )
332320
333- # for each gpu function in the gpu module, change memref.alloca address
334- # space to 3 (SLM) and convert vector to xegpu.
321+ # Convert vectot to xegpu
335322 gpu_mod_ops = match_and_split (mod , ops = {"gpu.module" })
336323 for gpu_mod in gpu_mod_ops :
337324 gpu_func = match (gpu_mod , ops = {"gpu.func" })
338- allocas = match_and_split (gpu_func , ops = {"memref.alloca" }, nhandles = 3 )
339- for alloca in allocas :
340- # print("Updating address space for alloca:")
341- update_address_space (alloca , address_space = 3 )
342325 gpu_func = apply_registered_pass (gpu_func , "convert-vector-to-xegpu" )
343326 transform .apply_cse (gpu_func )
344327 gpu_func = apply_registered_pass (gpu_func , "loop-invariant-code-motion" )
0 commit comments