Speedup unique_indices_length_kernel via binary search (#5766)#5766
Closed
AlbertDachiChen wants to merge 1 commit into
Closed
Speedup unique_indices_length_kernel via binary search (#5766)#5766AlbertDachiChen wants to merge 1 commit into
AlbertDachiChen wants to merge 1 commit into
Conversation
Contributor
|
@AlbertDachiChen has exported this pull request. If you are a Meta employee, you can view the originating Diff in D104827588. |
meta-codesync
Bot
changed the title
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 15, 2026
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
force-pushed
the
export-D104827588
branch
2 times, most recently
from
91b67c0 to
3a6d8ad
Compare
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 15, 2026
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
force-pushed
the
export-D104827588
branch
2 times, most recently
from
cc3955f to
662ccc9
Compare
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 15, 2026
Summary: Pull Request resolved: pytorch#5766 X-link: https://github.com/facebookresearch/FBGEMM/pull/2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
force-pushed
the
export-D104827588
branch
2 times, most recently
from
cfae1b3 to
402ab8c
Compare
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 18, 2026
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 18, 2026
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
added a commit
to AlbertDachiChen/FBGEMM
that referenced
this pull request
May 18, 2026
Summary: X-link: facebookresearch/FBGEMM#2695 The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape). The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in `linear_unique_indices`: since `at::_unique` is called with `sorted=True` and `linearize_index_wo_infos_kernel` writes `linear_indices[i] = hash_size_cumsum[t] + indices[i]`, feature t's unique linearized values occupy a contiguous slice of `linear_unique_indices`, namely `[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1]))`. The slice length is `num_unique_t`, which equals the `(max - min + 1)` reduction the old kernel computed. Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side `device_lower_bound` helper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize. The pipeline contract that ties the four kernels of `jagged_unique_indices_cuda` together (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why. Also adds `test_jagged_unique_indices_zch_huge_hash_size`, a regression test for the `ManagedCollisionCollection` shape that exposes `total_hash_size = INT64_MAX`. This shape is produced when a sharding group contains a single `HashZchManagedCollisionModule` with the default `input_hash_size=0`. `mc_modules._create_input_dists` then expands per-table hash size to `2**(63 - N) - 1` (per `torchrec/distributed/mc_modules.py:643`); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for the `cudaErrorIllegalInstruction` in S660690. **The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math on `total_hash_size`.** No public API change. Outputs of `jagged_unique_indices` are bit-identical to the previous version for all valid inputs. Reviewed By: q10 Differential Revision: D104827588
AlbertDachiChen
force-pushed
the
export-D104827588
branch
from
402ab8c to
f4727f2
Compare
Contributor
|
This pull request has been merged in 7121bf0. |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Summary:
X-link: https://github.com/facebookresearch/FBGEMM/pull/2695
The previous unique_indices_length_kernel computes per-feature unique-count via a BlockReduce-based min/max scan over the entire reverse_index array. With grid size = T (number of feature groups, typically 1-2 in production), only T SMs do work out of 132 on H100. Each block scans the full per-feature slice of reverse_index (~12M int64 = ~93MB for the prod IFR-MTML mc7 shape), bandwidth-bound on a single SM at ~30-50 GB/s. Total wall-clock is ~2-3 ms, dominating this op end-to-end (~60% of the ~5 ms baseline on the prod shape).
The kernel was reading 186 MB to compute 4 numbers (a min and max per feature group). It is wasteful because the information is already implicit in
linear_unique_indices: sinceat::_uniqueis called withsorted=Trueandlinearize_index_wo_infos_kernelwriteslinear_indices[i] = hash_size_cumsum[t] + indices[i], feature t's unique linearized values occupy a contiguous slice oflinear_unique_indices, namely[lower_bound(unique, hash_size_cumsum[t]), lower_bound(unique, hash_size_cumsum[t+1])). The slice length isnum_unique_t, which equals the(max - min + 1)reduction the old kernel computed.Replace the O(N) reduction with two O(log U) binary searches per feature group via a new device-side
device_lower_boundhelper. Block size 1024 -> 256 (no shared-memory reduction, no per-thread scratch). The per-block work is now ~336 B of reads (two binary searches, ~21 iterations each, 8 B per iteration), which trivially fits in cache; the T-block grid stops mattering because there is no work to parallelize.The pipeline contract that ties the four kernels of
jagged_unique_indices_cudatogether (linearize -> at::_unique -> delinearize -> length) is documented above the function so the next reader does not have to reverse-engineer it from the kernel bodies. The length kernel docstring states the local form of the invariant and points at the orchestrator for the why.Also adds
test_jagged_unique_indices_zch_huge_hash_size, a regression test for theManagedCollisionCollectionshape that exposestotal_hash_size = INT64_MAX. This shape is produced when a sharding group contains a singleHashZchManagedCollisionModulewith the defaultinput_hash_size=0.mc_modules._create_input_diststhen expands per-table hash size to2**(63 - N) - 1(pertorchrec/distributed/mc_modules.py:643); for N=0 (single-table group) that lands at INT64_MAX. This shape was not exercised by any existing test and was the trigger for thecudaErrorIllegalInstructionin S660690. The new length kernel handles it correctly (integer-only arithmetic at the boundary), but the test also serves as a trip-wire for downstream optimizations that introduce float-log2 math ontotal_hash_size.No public API change. Outputs of
jagged_unique_indicesare bit-identical to the previous version for all valid inputs.Reviewed By: q10
Differential Revision: D104827588