68 lines · plain
1! RUN: bbc -emit-hlfir -fcuda %s -o - | FileCheck %s2! RUN: bbc -emit-hlfir -fcuda %s -o - | fir-opt | FileCheck %s3 4! Test lowering of CUDA procedure calls.5 6module test_call7 use, intrinsic :: __fortran_builtins, only: __builtin_dim38contains9 attributes(global) subroutine dev_kernel0()10 end11 12 attributes(global) subroutine dev_kernel1(a)13 real :: a14 end15 16 subroutine host()17 real, device :: a18 integer(8) :: stream19 integer(4) :: nbytes20 21! CHECK-LABEL: func.func @_QMtest_callPhost()22! CHECK: %[[A:.*]]:2 = hlfir.declare %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QMtest_callFhostEa"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)23 24 call dev_kernel0<<<10, 20>>>()25! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}>>>()26 27 call dev_kernel0<<< __builtin_dim3(1,1,4), __builtin_dim3(32,1,1) >>>28! CHECK: %[[ADDR_DIM3_GRID:.*]] = fir.address_of(@_QQro._QM__fortran_builtinsT__builtin_dim3.{{.*}}) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>29! CHECK: %[[DIM3_GRID:.*]]:2 = hlfir.declare %[[ADDR_DIM3_GRID]] {fortran_attrs = #fir.var_attrs<parameter>, uniq_name = "_QQro._QM__fortran_builtinsT__builtin_dim3.0"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>, !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>)30! CHECK: %[[GRID_X:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"x"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>31! CHECK: %[[GRID_X_LOAD:.*]] = fir.load %[[GRID_X]] : !fir.ref<i32>32! CHECK: %[[GRID_Y:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"y"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>33! CHECK: %[[GRID_Y_LOAD:.*]] = fir.load %[[GRID_Y]] : !fir.ref<i32>34! CHECK: %[[GRID_Z:.*]] = hlfir.designate %[[DIM3_GRID]]#0{"z"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>35! CHECK: %[[GRID_Z_LOAD:.*]] = fir.load %[[GRID_Z]] : !fir.ref<i32>36! CHECK: %[[ADDR_DIM3_BLOCK:.*]] = fir.address_of(@_QQro._QM__fortran_builtinsT__builtin_dim3.{{.*}}) : !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>37! CHECK: %[[DIM3_BLOCK:.*]]:2 = hlfir.declare %[[ADDR_DIM3_BLOCK]] {fortran_attrs = #fir.var_attrs<parameter>, uniq_name = "_QQro._QM__fortran_builtinsT__builtin_dim3.1"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>, !fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>)38! CHECK: %[[BLOCK_X:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"x"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>39! CHECK: %[[BLOCK_X_LOAD:.*]] = fir.load %[[BLOCK_X]] : !fir.ref<i32>40! CHECK: %[[BLOCK_Y:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"y"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>41! CHECK: %[[BLOCK_Y_LOAD:.*]] = fir.load %[[BLOCK_Y]] : !fir.ref<i32>42! CHECK: %[[BLOCK_Z:.*]] = hlfir.designate %[[DIM3_BLOCK]]#0{"z"} : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_dim3{x:i32,y:i32,z:i32}>>) -> !fir.ref<i32>43! CHECK: %[[BLOCK_Z_LOAD:.*]] = fir.load %[[BLOCK_Z]] : !fir.ref<i32>44! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%[[GRID_X_LOAD]], %[[GRID_Y_LOAD]], %[[GRID_Z_LOAD]], %[[BLOCK_X_LOAD]], %[[BLOCK_Y_LOAD]], %[[BLOCK_Z_LOAD]]>>>()45 46 call dev_kernel0<<<10, 20, 2>>>()47! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}, %c2{{.*}}>>>() 48 49 call dev_kernel0<<<10, 20, 2, 0_8>>>()50! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel0<<<%c10{{.*}}, %c1{{.*}}, %c1{{.*}}, %c20{{.*}}, %c1{{.*}}, %c1{{.*}}, %c2{{.*}}, %{{.*}} : !fir.ref<i64>>>>()51 52 call dev_kernel1<<<1, 32>>>(a)53! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}>>>(%{{.*}}) : (!fir.ref<f32>)54 55 call dev_kernel1<<<*, 32>>>(a)56! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c-1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}>>>(%{{.*}})57 58 call dev_kernel1<<<*,32,0,stream>>>(a)59! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%c-1{{.*}}, %c1{{.*}}, %c1{{.*}}, %c32{{.*}}, %c1{{.*}}, %c1{{.*}}, %c0{{.*}}, %{{.*}} : !fir.ref<i64>>>>(%{{.*}}) : (!fir.ref<f32>)60 61 call dev_kernel1<<<*, 32, 0.8 * nbytes>>>(a)62! CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} fastmath<contract> : f3263! CHECK: %[[BYTES:.*]] = fir.convert %[[MUL]] : (f32) -> i3264! CHECK: cuf.kernel_launch @_QMtest_callPdev_kernel1<<<%{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}}, %[[BYTES]]>>>(%{{.*}}) : (!fir.ref<f32>)65 end66 67end68