Make isExpensiveLoadOrStore consider blocked pointers load and stores
#2570
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Signed-off-by: Tiotto, Ettore <[email protected]>
Contributor
Author
Signed-off-by: Tiotto, Ettore <[email protected]>
etiotto
commented
Oct 28, 2024
| // CHECK: %[[VAL_41:.*]]:3 = scf.for %{{.*}} = %{{.*}} to %{{.*}} step %{{.*}} iter_args(%{{.*}} = %{{.*}}, %{{.*}} = %[[VAL_36]], %{{.*}} = %[[VAL_40]]) -> (tensor<64x256xf32, #[[DPAS]]>, !tt.ptr<tensor<64x32xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #[[DPAS]], kWidth = 2}>>>, !tt.ptr<tensor<32x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #[[DPAS]], kWidth = 2}>>>) : i32 { | ||
| // CHECK: %[[VAL_46:.*]] = tt.load %{{.*}} {boundaryCheck = array<i32: 0, 1>} : !tt.ptr<tensor<64x32xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #[[DPAS]], kWidth = 2}>>> | ||
| // CHECK: %[[VAL_47:.*]] = tt.load %{{.*}} {boundaryCheck = array<i32: 0, 1>} : !tt.ptr<tensor<32x256xf16, #triton_gpu.dot_op<{opIdx = 1, parent = #[[DPAS]], kWidth = 2}>>> | ||
| // CHECK: %[[VAL_46:.*]] = tt.load %{{.*}} {boundaryCheck = array<i32: 0, 1>, triton_intel_gpu.block_io = "row_major"} : !tt.ptr<tensor<64x32xf16, #triton_gpu.dot_op<{opIdx = 0, parent = #[[DPAS]], kWidth = 2}>>> |
Contributor
Author
There was a problem hiding this comment.
Note: adding triton_intel_gpu.block_io is consistent with our optimization pipeline (in our pipeline this is done before the 2nd invocation of RemoveLayoutConversion)
| %1 = tt.get_program_id y : i32 | ||
| // CHECK: %[[VAL_0:.*]] = tt.make_tensor_ptr {{.*}} : <tensor<256x32xbf16, #triton_gpu.dot_op<{opIdx = 0, parent = #[[$DPAS]], kWidth = 2}>>> | ||
| // CHECK: %[[VAL_1:.*]] = tt.make_tensor_ptr {{.*}} : <tensor<32x256xbf16, #triton_gpu.dot_op<{opIdx = 1, parent = #[[$DPAS]], kWidth = 2}>>> | ||
| // CHECK: %[[VAL_0:.*]] = tt.make_tensor_ptr {{.*}} : <tensor<256x32xbf16, {{.*}}>> |
Contributor
Author
There was a problem hiding this comment.
The actual layout is not important in these tests.
Contributor
Author
|
@chengjunlu can you do a code review please ? |
chengjunlu
approved these changes
Oct 31, 2024
Contributor
LGTM. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The
isExpensiveLoadOrStorefunction (third_party/intel/lib/TritonIntelGPUTransforms/Utility.cpp) fails to consider block pointers and consequently always returnsfalsefor loads (and stores) operations that use a block pointer.In turn, this causes the
RemoveLayoutConversionpass to never consider loads using block pointers asanchoroperations.This PR changes
isExpensiveLoadOrStoreso that block pointer loads can be properly recognized. TheRemoveLayourConversionpass is then able to consider those loads as anchor operations and preserve their layout.Because
RemoveLayoutConversionis invoked at several points in the optimization pipeline, the change in third_party/intel/lib/TritonIntelGPUTransforms/Utility.cpp alone causes performance degradation in a couple of GEMM like benchmarks, specifically when operand A oftl.dotis transposed and when the input oftl.dotis first fed into an exponential.These 2 performance degradation have ben fixed by an enhancing the
MaterializeBlockPointerandMatmulLoopPipelineoptimizations, so that they can retrieve the dot layout of block pointer loads transitively from its users (in those benchmarks the blocked layout of block ptrs loads is transitively converted to a dot layout).