From b7e311847ff4e0b34a5a3dfc893613bdc54993eb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Beno=C3=AEt=20Legat?= <benoit.legat@gmail.com>
Date: Wed, 27 May 2026 23:57:03 +0200
Subject: [PATCH 1/2] Add GPU profiles

---
 perf/arraydiff.jl | 20 ++++++++++++++++++++
 perf/hand_cuda.jl | 24 ++++++++++++++++++++++++
 2 files changed, 44 insertions(+)

diff --git a/perf/arraydiff.jl b/perf/arraydiff.jl
index b79b44e..2581011 100644
--- a/perf/arraydiff.jl
+++ b/perf/arraydiff.jl
@@ -106,4 +106,24 @@ function neural(
     )
 end
 
+function profile_gpu(
+    ;T = Float32, h = 4096, d = 13, n = 178
+)
+    state = _build(T, h, d, n, true)
+    x = CUDA.CuVector{T}(vec(state.W1))
+    g = CUDA.zeros(T, h * d)
+    fill!(state.evaluator.backend.last_x, NaN)
+    CUDA.@sync CUDA.@allowscalar MOI.eval_objective_gradient(
+        state.evaluator,
+        g,
+        x,
+    )
+    fill!(state.evaluator.backend.last_x, NaN)
+    return CUDA.@profile CUDA.@sync CUDA.@allowscalar MOI.eval_objective_gradient(
+        state.evaluator,
+        g,
+        x,
+    )
+end
+
 end # module
diff --git a/perf/hand_cuda.jl b/perf/hand_cuda.jl
index b69de98..0149fde 100644
--- a/perf/hand_cuda.jl
+++ b/perf/hand_cuda.jl
@@ -111,4 +111,28 @@ function neural(
     end
 end
 
+function profile_gpu(;
+    T = Float32, h = 4096, d = 13, n = 178,
+    prealloc::Bool = true,
+)
+    Random.seed!(0)
+    W1 = randn(T, h, d)
+    W2 = randn(T, OUT_DIM, h)
+    X = randn(T, d, n)
+    y = randn(T, OUT_DIM, n)
+    W1g, W2g, Xg, yg = CuArray(W1), CuArray(W2), CuArray(X), CuArray(y)
+    CUDA.synchronize()
+    CUDA.@sync if prealloc
+        gradient!(Buffers{(typeof(W1g))}(h, d, n), W1g, W2g, Xg, yg)
+    else
+        gradient_alloc(W1g, W2g, Xg, yg)
+    end
+    return CUDA.@profile CUDA.@sync if prealloc
+        gradient!(Buffers{(typeof(W1g))}(h, d, n), W1g, W2g, Xg, yg)
+    else
+        gradient_alloc(W1g, W2g, Xg, yg)
+    end
+end
+
+
 end # module

From 8bc418dbc1842d75702e6224005d6e629b1c3da3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Beno=C3=AEt=20Legat?= <benoit.legat@gmail.com>
Date: Thu, 28 May 2026 20:30:05 +0200
Subject: [PATCH 2/2] Fix format

---
 perf/arraydiff.jl | 4 +---
 perf/hand_cuda.jl | 6 ++++--
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/perf/arraydiff.jl b/perf/arraydiff.jl
index 2581011..9d5b3f5 100644
--- a/perf/arraydiff.jl
+++ b/perf/arraydiff.jl
@@ -106,9 +106,7 @@ function neural(
     )
 end
 
-function profile_gpu(
-    ;T = Float32, h = 4096, d = 13, n = 178
-)
+function profile_gpu(; T = Float32, h = 4096, d = 13, n = 178)
     state = _build(T, h, d, n, true)
     x = CUDA.CuVector{T}(vec(state.W1))
     g = CUDA.zeros(T, h * d)
diff --git a/perf/hand_cuda.jl b/perf/hand_cuda.jl
index 0149fde..5dd8d68 100644
--- a/perf/hand_cuda.jl
+++ b/perf/hand_cuda.jl
@@ -112,7 +112,10 @@ function neural(
 end
 
 function profile_gpu(;
-    T = Float32, h = 4096, d = 13, n = 178,
+    T = Float32,
+    h = 4096,
+    d = 13,
+    n = 178,
     prealloc::Bool = true,
 )
     Random.seed!(0)
@@ -134,5 +137,4 @@ function profile_gpu(;
     end
 end
 
-
 end # module