Correct mma_sp_m16n8k16 implementation

eschnett · Jan 16, 2024 · 119f40b · 119f40b
1 parent 15eed5a
commit 119f40b
Showing 1 changed file with 3 additions and 5 deletions.
diff --git a/src/IndexSpaces.jl b/src/IndexSpaces.jl
@@ -1988,7 +1988,7 @@ CUDA.@device_override function mma_m16n8k16(A::NTuple{4,Float16x2}, B::NTuple{2,
     return (Float16x2(D[1] % UInt32), Float16x2(D[2] % UInt32))
 end
 
-mma_sp_m16n8k16(A::NTuple{4,Float16x2}, B::NTuple{2,Float16x2}, C::NTuple{2,Float16x2}) = C
+mma_sp_m16n8k16(A::NTuple{2,Float16x2}, B::NTuple{2,Float16x2}, C::NTuple{2,Float16x2}, e::Int2x16, f::Integer) = C
 CUDA.@device_override function mma_sp_m16n8k16(
     A::NTuple{2,Float16x2}, B::NTuple{2,Float16x2}, C::NTuple{2,Float16x2}, e::Int2x16, f::Integer
 )
@@ -2005,8 +2005,6 @@ CUDA.@device_override function mma_sp_m16n8k16(
         NTuple{8,Int32},
         A[1].val % Int32,
         A[2].val % Int32,
-        A[3].val % Int32,
-        A[4].val % Int32,
         B[1].val % Int32,
         B[2].val % Int32,
         C[1].val % Int32,
@@ -2698,8 +2696,8 @@ function mma_sp_row_col_m16n8k16_f16!(
                         e6
                     elseif threadgroup == 7i32
                         e7
-                    else
-                        @assert false
+                    # else
+                    #     @assert false
                     end
                     IndexSpaces.mma_sp_m16n8k16(($A0_name, $A1_name), ($B0_name, $B1_name), ($C0_name, $C1_name), e, 0i32)
                 end