Incremental progress

Author: kshyatt

src/pullbacks/lq.jl

@@ -59,7 +59,7 @@ function lq_pullback!(
     ΔQ̃ = zero!(similar(Q, (p, n)))
     if !iszerotangent(ΔQ)
         ΔQ1 = view(ΔQ, 1:p, :)
-        copy!(ΔQ̃, ΔQ1)
+        ΔQ̃ .= ΔQ1
         if p < size(Q, 1)
             Q2 = view(Q, (p + 1):size(Q, 1), :)
             ΔQ2 = view(ΔQ, (p + 1):size(Q, 1), :)

src/pullbacks/qr.jl

@@ -59,7 +59,7 @@ function qr_pullback!(
 
     ΔQ̃ = zero!(similar(Q, (m, p)))
     if !iszerotangent(ΔQ)
-        copy!(ΔQ̃, view(ΔQ, :, 1:p))
+        ΔQ̃ .= view(ΔQ, :, 1:p)
         if p < size(Q, 2)
             Q2 = view(Q, :, (p + 1):size(Q, 2))
             ΔQ2 = view(ΔQ, :, (p + 1):size(Q, 2))
@@ -91,9 +91,9 @@ function qr_pullback!(
     M = zero!(similar(R, (p, p)))
     if !iszerotangent(ΔR)
         ΔR11 = view(ΔR, 1:p, 1:p)
-        M = mul!(M, ΔR11, R11', 1, 1)
+        M += ΔR11 * R11'
     end
-    M = mul!(M, Q1', ΔQ̃, -1, 1)
+    M -= Q1' * ΔQ̃
     view(M, lowertriangularind(M)) .= conj.(view(M, uppertriangularind(M)))
     if eltype(M) <: Complex
         Md = diagview(M)

test/mooncake.jl

@@ -16,7 +16,7 @@ for T in BLASFloats, n in (17, m, 23)
     TestSuite.seed_rng!(123)
     if CUDA.functional()
         TestSuite.test_mooncake(CuMatrix{T}, (m, n); atol = m * n * TestSuite.precision(T), rtol = m * n * TestSuite.precision(T))
-        #n == m && TestSuite.test_mooncake(Diagonal{T, CuVector{T}}, m; atol = m * TestSuite.precision(T), rtol = m * TestSuite.precision(T))
+        n == m && TestSuite.test_mooncake(Diagonal{T, CuVector{T}}, m; atol = m * TestSuite.precision(T), rtol = m * TestSuite.precision(T))
     end
     #=if AMDGPU.functional()
         TestSuite.test_mooncake(ROCMatrix{T}, (m, n); atol = m * n * TestSuite.precision(T), rtol = m * n * TestSuite.precision(T))