Add more blas fns (#1544)

* Add more blas fns

* fix

* Update Project.toml

Author: wsmoses

Project.toml

@@ -16,6 +16,16 @@ Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 
+[weakdeps]
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
+SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
+StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
+
+[extensions]
+EnzymeChainRulesCoreExt = "ChainRulesCore"
+EnzymeSpecialFunctionsExt = "SpecialFunctions"
+EnzymeStaticArraysExt = "StaticArrays"
+
 [compat]
 CEnum = "0.4, 0.5"
 ChainRulesCore = "1"
@@ -29,17 +39,7 @@ SpecialFunctions = "1, 2"
 StaticArrays = "1"
 julia = "1.6"
 
-[extensions]
-EnzymeChainRulesCoreExt = "ChainRulesCore"
-EnzymeSpecialFunctionsExt = "SpecialFunctions"
-EnzymeStaticArraysExt = "StaticArrays"
-
 [extras]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
-
-[weakdeps]
-ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
-SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
-StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"

src/compiler.jl

@@ -5154,10 +5154,13 @@ function GPUCompiler.codegen(output::Symbol, job::CompilerJob{<:EnzymeTarget};
     check_ir(job, mod)
 
     disableFallback = String[]
+
+    ForwardModeDerivatives = ("dot","gemm","gemv","axpy","copy","scal")
+    ReverseModeDerivatives = ("dot","gemm","gemv","axpy","copy","scal", "trmv", "syrk", "trmm", "trsm")
     # Tablegen BLAS does not support forward mode yet
     if !(mode == API.DEM_ForwardMode && Enzyme.API.runtimeActivity())
         for ty in ("s", "d")
-            for func in ("dot","gemm","gemv","axpy","copy","scal")
+            for func in (mode == API.DEM_ForwardMode ? ForwardModeDerivatives : ReverseModeDerivatives)
                 for prefix in ("", "cblas_")
                     for ending in ("", "_", "64_", "_64_")
                         push!(disableFallback, prefix*ty*func*ending)