diff --git a/ptx/lib/zluda_ptx_impl.bc b/ptx/lib/zluda_ptx_impl.bc index 1edcbd5..4e44151 100644 Binary files a/ptx/lib/zluda_ptx_impl.bc and b/ptx/lib/zluda_ptx_impl.bc differ diff --git a/ptx/lib/zluda_ptx_impl.cpp b/ptx/lib/zluda_ptx_impl.cpp index ecbe691..a6417f6 100644 --- a/ptx/lib/zluda_ptx_impl.cpp +++ b/ptx/lib/zluda_ptx_impl.cpp @@ -1,5 +1,5 @@ // Compile and disassemble: -// python3 ./cvt.py > cvt.h && /opt/rocm/llvm/bin/clang -std=c++17 -Xclang -no-opaque-pointers -Wall -Wextra -Wsign-compare -Wconversion -x hip zluda_ptx_impl.cpp -S -emit-llvm --cuda-device-only -nogpulib -O3 -Xclang -fallow-half-arguments-and-returns -o - | sed -e 's/define/define linkonce_odr/g' | sed -e '/@llvm.used/d' | sed -e 's/\"target-cpu\"=\"[^\"]*\"//g' | sed -e 's/\"target-features\"=\"[^\"]*\"//g' | sed -e 's/\"denormal-fp-math-f32\"=\"[^\"]*\"//g' | sed -e 's/!llvm.module.flags = !{!0, !1, !2, !3, !4}/!llvm.module.flags = !{ }/g' | sed -e 's/memory(none)/readnone/g' | sed -e 's/memory(argmem: readwrite, inaccessiblemem: readwrite)/inaccessiblemem_or_argmemonly/g' | sed -e 's/memory(read)/readonly/g' | sed -e 's/memory(argmem: readwrite)/argmemonly/g' | llvm-as-13 -o zluda_ptx_impl.bc && /opt/rocm/llvm/bin/llvm-dis zluda_ptx_impl.bc +// python3 ./cvt.py > cvt.h && /opt/rocm/llvm/bin/clang -std=c++20 -Xclang -no-opaque-pointers -Wall -Wextra -Wsign-compare -Wconversion -x hip zluda_ptx_impl.cpp -S -emit-llvm --cuda-device-only -nogpulib -O3 -Xclang -fallow-half-arguments-and-returns -o - | sed -e 's/define/define linkonce_odr/g' | sed -e '/@llvm.used/d' | sed -e 's/\"target-cpu\"=\"[^\"]*\"//g' | sed -e 's/\"target-features\"=\"[^\"]*\"//g' | sed -e 's/\"denormal-fp-math-f32\"=\"[^\"]*\"//g' | sed -e 's/!llvm.module.flags = !{!0, !1, !2, !3, !4}/!llvm.module.flags = !{ }/g' | sed -e 's/memory(none)/readnone/g' | sed -e 's/memory(argmem: readwrite, inaccessiblemem: readwrite)/inaccessiblemem_or_argmemonly/g' | sed -e 's/memory(read)/readonly/g' | sed -e 's/memory(argmem: readwrite)/argmemonly/g' | llvm-as-13 -o zluda_ptx_impl.bc && /opt/rocm/llvm/bin/llvm-dis zluda_ptx_impl.bc // Compile to binary: // /opt/rocm/llvm/bin/clang -x ir -target amdgcn-amd-amdhsa -Xlinker --no-undefined zluda_ptx_impl.bc -mno-wavefrontsize64 -mcpu=gfx1030 // Decompile: @@ -1260,6 +1260,7 @@ extern "C" default: return 0; } + return 2; case 'l': switch (s[1]) { @@ -1289,17 +1290,18 @@ extern "C" case 'X': case 'n': len = 8; - return 2; + break; default: return 0; } + return 3; default: return 0; } + return 2; default: return 0; } - return 1; } __device__ static bool parse_printf_specifier(const char *s, uint8_t &len) @@ -1407,16 +1409,16 @@ extern "C" { s += specifier_with_length; } - if (len > 0) - { - uint64_t value = read_valist(valist_ptr, valist_offset, len); - handle = __ockl_printf_append_args(handle, 1, value, 0, 0, 0, 0, 0, 0, 0); - } + } + + if (len > 0) + { + uint64_t value = read_valist(valist_ptr, valist_offset, len); + handle = __ockl_printf_append_args(handle, 1, value, 0, 0, 0, 0, 0, 0, 0); } } } return (uint32_t)__ockl_printf_append_args(handle, 0, 0, 0, 0, 0, 0, 0, 0, 1); - return 1; } int64_t __ockl_mul_hi_i64(int64_t x, int64_t y) __attribute__((device));