fix: task switching in AMDGPU complex batched_matmul (#178)

* ci(buildkite): add downstream testing for NeuralOperators * perf: restore old batched_mul * fix: disable threading for certain devices * revert: "perf: restore old batched_mul" This reverts commit a8c0f3b4615f96a8773577e16fac61ba310d8123.
LuxDL · Oct 25, 2024 · c63829b · c63829b
1 parent ceb36a1
commit c63829b
Show file tree

Hide file tree

Showing 3 changed files with 40 additions and 8 deletions.
diff --git a/lib/LuxLib/.buildkite/testing.yml b/lib/LuxLib/.buildkite/testing.yml
@@ -38,7 +38,6 @@ steps:
                 - src
                 - ext
         env:
-          RETESTITEMS_NWORKERS: 2
           BACKEND_GROUP: "AMDGPU"
         agents:
           queue: "juliagpu"
@@ -126,6 +125,7 @@ steps:
             repo:
               - "Boltz"
               - "Lux"
+              - "NeuralOperators"
 
   - group: ":telescope: Downstream AMD GPU"
     steps:
@@ -143,15 +143,14 @@ steps:
           queue: "juliagpu"
           rocm: "*"
           rocmgpu: "*"
-        env:
-          RETESTITEMS_NWORKERS: 2
         if: build.message !~ /\[skip tests\]/ && build.message !~ /\[skip downstream\]/ && build.message !~ /\[skip ci\]/ && build.branch != "main"
         timeout_in_minutes: 240
         matrix:
           setup:
             repo:
               - "Boltz"
               - "Lux"
+              - "NeuralOperators"
 
 env:
   JULIA_PKG_SERVER: ""

diff --git a/lib/LuxLib/Project.toml b/lib/LuxLib/Project.toml
@@ -1,7 +1,7 @@
 name = "LuxLib"
 uuid = "82251201-b29d-42c6-8e01-566dec8acb11"
 authors = ["Avik Pal <avikpal@mit.edu> and contributors"]
-version = "1.3.4"
+version = "1.3.5"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

diff --git a/lib/LuxLib/src/impl/batched_mul.jl b/lib/LuxLib/src/impl/batched_mul.jl
@@ -70,15 +70,15 @@ end
 function batched_matmul_loopvec_impl! end
 
 function fallback_batched_matmul(
-        dev, x::AbstractArray{xT, 3}, y::AbstractArray{yT, 3}) where {xT, yT}
+        opmode, x::AbstractArray{xT, 3}, y::AbstractArray{yT, 3}) where {xT, yT}
     z = similar(x, promote_type(eltype(x), eltype(y)), size(x, 1),
         size(y, 2), max(size(x, 3), size(y, 3)))
-    fallback_batched_matmul!(z, dev, x, y)
+    fallback_batched_matmul!(z, opmode, x, y)
     return z
 end
 
 function fallback_batched_matmul!(
-        z::AbstractArray{zT, 3}, dev, x::AbstractArray{xT, 3},
+        z::AbstractArray{zT, 3}, opmode, x::AbstractArray{xT, 3},
         y::AbstractArray{yT, 3}) where {zT, xT, yT}
     # XXX: bring back once the enzyme segfault is fixed
     # @warn "Using fallback Batched Matrix Multiply routine for $(dev) with A: size = \
@@ -90,6 +90,36 @@ function fallback_batched_matmul!(
         throw(DimensionMismatch(lazy"size(x) = $(size(x)), size(y) = $(size(y)) inconsistent for batched_matmul."))
     end
 
+    if use_threaded_batched_matmul(get_device_type(x))
+        unsafe_fallback_threaded_batched_matmul!(z, x, y)
+    else
+        unsafe_fallback_serial_batched_matmul!(z, x, y)
+    end
+
+    return
+end
+
+function unsafe_fallback_serial_batched_matmul!(
+        z::AbstractArray{zT, 3}, x::AbstractArray{xT, 3},
+        y::AbstractArray{yT, 3}) where {zT, xT, yT}
+    if size(x, 3) == size(y, 3)
+        for L in axes(z, 3)
+            mul!(batchview(z, L), batchview(x, L), batchview(y, L))
+        end
+    elseif size(x, 3) == 1
+        for L in axes(z, 3)
+            mul!(batchview(z, L), batchview(x, 1), batchview(y, L))
+        end
+    else # has to be size(y, 3) == 1
+        for L in axes(z, 3)
+            mul!(batchview(z, L), batchview(x, L), batchview(y, 1))
+        end
+    end
+end
+
+function unsafe_fallback_threaded_batched_matmul!(
+        z::AbstractArray{zT, 3}, x::AbstractArray{xT, 3},
+        y::AbstractArray{yT, 3}) where {zT, xT, yT}
     old_threads = maybe_reduce_BLAS_threads(z)
 
     if size(x, 3) == size(y, 3)
@@ -107,10 +137,13 @@ function fallback_batched_matmul!(
     end
 
     reset_BLAS_threads(old_threads)
-
     return
 end
 
+use_threaded_batched_matmul(::Type) = false
+use_threaded_batched_matmul(::Type{CUDADevice}) = true
+use_threaded_batched_matmul(::Type{CPUDevice}) = true
+
 function CRC.rrule(::typeof(batched_matmul), x::AbstractArray{xT, 3},
         y::AbstractArray{yT, 3}) where {xT, yT}
     ∇batched_matmul = @closure Δ_ -> begin