adaptive mixed precision

example/mixed_precision.jl

@@ -0,0 +1,14 @@
+using Dagger
+using LinearAlgebra
+using KernelFunctions
+using Distances
+
+k = GammaExponentialKernel(; γ=0.5, metric=Euclidean());
+x = randn(4000, 2000);
+A =  kernelmatrix(k, x);
+DA = view(A, Blocks(400, 400));
+MP  = fill("FP64", 5, 5);
+DMP = view(MP, Blocks(1, 1));
+
+Dagger.adaptive_mp!(DA, DMP, 10^-4);
+collect(DMP)

src/Dagger.jl

@@ -74,7 +74,7 @@ include("array/sort.jl")
 include("array/linalg.jl")
 include("array/mul.jl")
 include("array/cholesky.jl")
-
+include("array/adaptive_mp.jl")
 # Visualization
 include("visualization.jl")
 include("ui/gantt-common.jl")

src/array/adaptive_mp.jl

@@ -0,0 +1,88 @@
+function tile_precision(uplo, global_norm, scalar_factore, tolerance, A)
+    tile_sqr = 0.0
+
+    if uplo == 'G'
+        tile_sqr = mapreduce(LinearAlgebra.norm_sqr, +, A)
+    elseif  uplo == 'L'
+        tile_sqr= mapreduce(LinearAlgebra.norm_sqr, +, LowerTriangular(A))
+    elseif  uplo == 'U'
+        tile_sqr= mapreduce(LinearAlgebra.norm_sqr, +, UpperTriangular(A))
+    end
+    tile_norm = sqrt(tile_sqr)
+
+    cal = tile_norm * scalar_factore / global_norm 
+    decision_hp = tile_norm * scalar_factore / global_norm  < tolerance / eps(Float16); 
+    decision_sp = tile_norm * scalar_factore / global_norm < tolerance / eps(Float32); 
+    decision_fp8 = tile_norm * scalar_factore / global_norm  < tolerance / 0.0625; 
+
+    if decision_fp8
+        return "FP8"
+    elseif decision_hp
+        return "FP16"
+    elseif decision_sp
+        return "FP32"
+    else
+        return "FP64"
+    end
+end
+
+function adaptive_mp!(A::UpperTriangular{T,<:DArray{T,2}}, MP::UpperTriangular{String,<:DArray{String,2}}, tolerance::Float64) where T
+
+    Ac = parent(A).chunks
+    MPc= parent(MP).chunks
+    mt, nt = size(Ac)
+
+    global_norm = LinearAlgebra.norm2(A)
+
+    for m in range(1, mt)
+        for n in range(m, nt)
+            if m==n
+                MP[m, n] = Dagger.@spawn tile_precision('U', global_norm, max(mt, nt), tolerance, Ac[m, n])
+            else
+                MP[m, n] = Dagger.@spawn tile_precision('G', global_norm, max(mt, nt), tolerance, Ac[m, n])
+            end
+
+        end
+    end
+    return UpperTriangular(MP)
+end
+
+function adaptive_mp!(A::LowerTriangular{T,<:DArray{T,2}}, MP::LowerTriangular{String,<:DArray{String,2}}, tolerance::Float64) where T
+
+    Ac = parent(A).chunks
+    MPc= parent(MP).chunks
+    mt, nt = size(Ac)
+
+    global_norm = LinearAlgebra.norm2(A)
+
+    for m in range(1, mt)
+        for n in range(1, m)
+            if m==n
+                MP[m, n] = Dagger.@spawn tile_precision('L', global_norm, max(mt, nt), tolerance, Ac[m, n])
+            else
+                MP[m, n] = Dagger.@spawn tile_precision('G', global_norm, max(mt, nt), tolerance, Ac[m, n])
+            end
+
+        end
+    end
+    return LowerTriangular(MP)
+end
+
+
+function adaptive_mp!(A::DArray{T,2}, MP::DArray{String,2}, tolerance::Float64) where T
+
+    Ac = parent(A).chunks
+    MPc= parent(MP).chunks
+    mt, nt = size(Ac)
+
+    global_norm = LinearAlgebra.norm2(A)
+
+    for m in range(1, mt)
+        for n in range(1, nt)
+            MP[m, n] = Dagger.@spawn tile_precision('G', global_norm, max(mt, nt), tolerance, Ac[m, n])
+        end
+    end
+
+    return MP
+end
+