feat: new syrk passes + lowering

Project.toml

@@ -1,7 +1,7 @@
 name = "Reactant"
 uuid = "3c362404-f566-11ee-1572-e11a4b42c853"
 authors = ["William Moses <wmoses@mit.edu>", "Valentin Churavy <vchuravy@mit.edu>", "Sergio Sánchez Ramírez <sergio.sanchez.ramirez@bsc.es>", "Paul Berg <paul@plutojl.org>", "Avik Pal <avikpal@mit.edu>", "Mosè Giordano <mose@gnu.org>"]
-version = "0.2.180"
+version = "0.2.181"
 
 [deps]
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
@@ -105,7 +105,7 @@ PythonCall = "0.9.25"
 Random = "1.10"
 Random123 = "1.7"
 ReactantCore = "0.1.16"
-Reactant_jll = "0.0.265"
+Reactant_jll = "0.0.267"
 ScopedValues = "1.3.0"
 Scratch = "1.2"
 Sockets = "1.10"

src/Compiler.jl

@@ -703,6 +703,8 @@ function optimization_passes(
     recognize_comms::Bool=true,
     lower_comms::Bool=true,
     backend::String="gpu",
+    is_sharded::Bool=false,
+    raise_shlo_to_blas_lapack::Bool=true,
 )
     (; max_constant_threshold) = compile_options
 
@@ -909,8 +911,19 @@ function optimization_passes(
         "transpose_symmetric_simplify",
         "divide_negated_operands_simplify",
         "multiply_negated_operands_simplify",
+        "transpose_syrk_to_syrk",
+        "fuse_mul_into_syrk",
+        "fuse_add_into_syrk",
+        "factor_scalars_in_dot_general",
     ]
 
+    if !is_sharded
+        # these passes don't have optimized sharding implementations
+        if raise_shlo_to_blas_lapack
+            append!(transform_passes_list, ["dot_general_to_syrk"])
+        end
+    end
+
     if !compile_options.disable_auto_batching_passes
         append!(
             transform_passes_list,
@@ -1693,10 +1706,10 @@ function compile_mlir!(
     end
 
     opt_passes = optimization_passes(
-        compile_options; sroa=true, recognize_comms, lower_comms, backend
+        compile_options; sroa=true, recognize_comms, lower_comms, backend, is_sharded
     )
     opt_passes2 = optimization_passes(
-        compile_options; sroa=false, recognize_comms, lower_comms, backend
+        compile_options; sroa=false, recognize_comms, lower_comms, backend, is_sharded
     )
 
     raise_passes = if raise isa String
@@ -1718,6 +1731,7 @@ function compile_mlir!(
                 recognize_comms,
                 lower_comms,
                 backend,
+                is_sharded,
             )
             result = result * "," * opt_passes3
         end
@@ -1728,6 +1742,8 @@ function compile_mlir!(
 
     blas_int_width = sizeof(BlasInt) * 8
     lower_enzymexla_linalg_pass = "lower-enzymexla-linalg{backend=$backend \
+                                   blas_int_width=$blas_int_width},\
+                                   lower-enzymexla-blas{backend=$backend \
                                    blas_int_width=$blas_int_width},\
                                    lower-enzymexla-lapack{backend=$backend \
                                    blas_int_width=$blas_int_width}"
@@ -2012,6 +2028,8 @@ function compile_mlir!(
                 recognize_comms,
                 lower_comms,
                 backend,
+                is_sharded,
+                raise_shlo_to_blas_lapack=false,
             ),
             "post_op_transpose_reshape",
         )
@@ -2154,7 +2172,15 @@ function compile_mlir!(
                 run_pass_pipeline!(
                     mod,
                     join(
-                        [opt_passes, "canonicalize", "cse", "canonicalize", opt_passes2],
+                        [
+                            opt_passes,
+                            "canonicalize",
+                            "cse",
+                            "canonicalize",
+                            opt_passes2,
+                            lower_enzymexla_linalg_pass,
+                            jit,
+                        ],
                         ",",
                     ),
                     "mid_pad_opts",

src/stdlibs/LinearAlgebra.jl

@@ -46,6 +46,21 @@ function __init__()
             (BLAS.@blasfunc(dgesvj_), :enzymexla_lapack_dgesvj_),
             (BLAS.@blasfunc(cgesvj_), :enzymexla_lapack_cgesvj_),
             (BLAS.@blasfunc(zgesvj_), :enzymexla_lapack_zgesvj_),
+            # syrk
+            (BLAS.@blasfunc(ssyrk_), :enzymexla_blas_ssyrk_),
+            (BLAS.@blasfunc(dsyrk_), :enzymexla_blas_dsyrk_),
+            (BLAS.@blasfunc(csyrk_), :enzymexla_blas_csyrk_),
+            (BLAS.@blasfunc(zsyrk_), :enzymexla_blas_zsyrk_),
+            # trmm
+            (BLAS.@blasfunc(strmm_), :enzymexla_blas_strmm_),
+            (BLAS.@blasfunc(dtrmm_), :enzymexla_blas_dtrmm_),
+            (BLAS.@blasfunc(ctrmm_), :enzymexla_blas_ctrmm_),
+            (BLAS.@blasfunc(ztrmm_), :enzymexla_blas_ztrmm_),
+            # symm
+            (BLAS.@blasfunc(ssymm_), :enzymexla_blas_ssymm_),
+            (BLAS.@blasfunc(dsymm_), :enzymexla_blas_dsymm_),
+            (BLAS.@blasfunc(csymm_), :enzymexla_blas_csymm_),
+            (BLAS.@blasfunc(zsymm_), :enzymexla_blas_zsymm_),
         ]
             sym = Libdl.dlsym(libblastrampoline_handle, cname)
             @ccall MLIR.API.mlir_c.EnzymeJaXMapSymbol(

test/integration/linear_algebra.jl

@@ -723,3 +723,24 @@ end
     @jit LinearAlgebra.normalize!(x_ra)
     @test x_ra ≈ x
 end
+
+raise_to_syrk(x, y) = 3 .* (x * transpose(x)) .+ 5 .* y
+raise_to_syrk2(x, y) = 3 .* (transpose(x) * x) .+ 5 .* y
+
+@testset "syrk optimizations" begin
+    @testset for elty in (Float32, Float64, ComplexF32, ComplexF64)
+        x = Reactant.TestUtils.construct_test_array(elty, 4, 5)
+        y1 = Reactant.TestUtils.construct_test_array(elty, 4, 4)
+        y2 = Reactant.TestUtils.construct_test_array(elty, 5, 5)
+        x_ra = Reactant.to_rarray(x)
+
+        @testset for (fn, y) in ((raise_to_syrk, y1), (raise_to_syrk2, y2))
+            y_ra = Reactant.to_rarray(y)
+
+            hlo = @code_hlo optimize = :before_jit fn(x_ra, y_ra)
+            @test occursin("enzymexla.blas.syrk", repr(hlo))
+
+            @test @jit(fn(x_ra, y_ra)) ≈ fn(x, y)
+        end
+    end
+end