mirrors/ZLUDA
0
0
Fork 0
mirror of https://github.com/vosen/ZLUDA.git synced 2025-08-05 07:41:25 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

Hack enough functionality that AMD GPU code builds

Browse source
This commit is contained in:
Andrzej Janik 2021-08-03 00:22:47 +02:00
commit 638786b0ec
5 changed files with 195 additions and 124 deletions
Download patch file Download diff file Expand all files Collapse all files

BIN
ptx/lib/zluda_ptx_impl.bc Normal file
View file

Binary file not shown.

200
ptx/lib/zluda_ptx_impl.cl
View file

@ -1,5 +1,6 @@
// Every time this file changes it must te rebuilt: // Every time this file changes it must te rebuilt:
// ocloc -file zluda_ptx_impl.cl -64 -options "-cl-std=CL2.0 -Dcl_intel_bit_instructions" -out_dir . -device kbl -output_no_suffix -spv_only // ocloc -file zluda_ptx_impl.cl -64 -options "-cl-std=CL2.0 -Dcl_intel_bit_instructions -DINTEL" -out_dir . -device kbl -output_no_suffix -spv_only
// /opt/amdgpu-pro/bin/clang -x cl -Xclang -finclude-default-header zluda_ptx_impl.cl -cl-std=CL2.0 -c -target amdgcn-amd-amdhsa -o zluda_ptx_impl.bc -emit-llvm
// Additionally you should strip names: // Additionally you should strip names:
// spirv-opt --strip-debug zluda_ptx_impl.spv -o zluda_ptx_impl.spv --target-env=spv1.3 // spirv-opt --strip-debug zluda_ptx_impl.spv -o zluda_ptx_impl.spv --target-env=spv1.3
@ -129,137 +130,142 @@ atomic_dec(atom_acq_rel_cta_shared_dec, memory_order_acq_rel, memory_order_acqui
atomic_dec(atom_relaxed_gpu_shared_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local); atomic_dec(atom_relaxed_gpu_shared_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local);
atomic_dec(atom_acquire_gpu_shared_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, __local); atomic_dec(atom_acquire_gpu_shared_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, __local);
atomic_dec(atom_release_gpu_shared_dec, memory_order_release, memory_order_acquire, memory_scope_device, __local);
atomic_dec(atom_acq_rel_gpu_shared_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local);
atomic_dec(atom_relaxed_sys_shared_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local);
atomic_dec(atom_acquire_sys_shared_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, __local);
atomic_dec(atom_release_sys_shared_dec, memory_order_release, memory_order_acquire, memory_scope_device, __local);
atomic_dec(atom_acq_rel_sys_shared_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local); atomic_dec(atom_acq_rel_sys_shared_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local);
// atom.add.f32 #ifdef INTEL
atomic_add(atom_relaxed_cta_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, , float, atomic_uint, uint); // atom.add.f32
atomic_add(atom_acquire_cta_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint); atomic_add(atom_relaxed_cta_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, , float, atomic_uint, uint);
atomic_add(atom_release_cta_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint); atomic_add(atom_acquire_cta_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint);
atomic_add(atom_acq_rel_cta_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint); atomic_add(atom_release_cta_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint);
atomic_add(atom_acq_rel_cta_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, , float, atomic_uint, uint);
atomic_add(atom_relaxed_gpu_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_relaxed_gpu_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_acquire_gpu_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_acquire_gpu_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_release_gpu_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_release_gpu_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_acq_rel_gpu_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_acq_rel_gpu_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_relaxed_sys_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_relaxed_sys_generic_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_acquire_sys_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_acquire_sys_generic_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_release_sys_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_release_sys_generic_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_acq_rel_sys_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint); atomic_add(atom_acq_rel_sys_generic_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , float, atomic_uint, uint);
atomic_add(atom_relaxed_cta_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global, float, atomic_uint, uint); atomic_add(atom_relaxed_cta_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global, float, atomic_uint, uint);
atomic_add(atom_acquire_cta_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint); atomic_add(atom_acquire_cta_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint);
atomic_add(atom_release_cta_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint); atomic_add(atom_release_cta_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint);
atomic_add(atom_acq_rel_cta_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint); atomic_add(atom_acq_rel_cta_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global, float, atomic_uint, uint);
atomic_add(atom_relaxed_gpu_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_relaxed_gpu_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_acquire_gpu_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_acquire_gpu_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_release_gpu_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_release_gpu_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_acq_rel_gpu_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_acq_rel_gpu_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_relaxed_sys_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_relaxed_sys_global_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_acquire_sys_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_acquire_sys_global_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_release_sys_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_release_sys_global_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_acq_rel_sys_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint); atomic_add(atom_acq_rel_sys_global_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, float, atomic_uint, uint);
atomic_add(atom_relaxed_cta_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local, float, atomic_uint, uint); atomic_add(atom_relaxed_cta_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local, float, atomic_uint, uint);
atomic_add(atom_acquire_cta_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint); atomic_add(atom_acquire_cta_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint);
atomic_add(atom_release_cta_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint); atomic_add(atom_release_cta_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint);
atomic_add(atom_acq_rel_cta_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint); atomic_add(atom_acq_rel_cta_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __local, float, atomic_uint, uint);
atomic_add(atom_relaxed_gpu_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_relaxed_gpu_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_acquire_gpu_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_acquire_gpu_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_release_gpu_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_release_gpu_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_acq_rel_gpu_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_acq_rel_gpu_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_relaxed_sys_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_relaxed_sys_shared_add_f32, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_acquire_sys_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_acquire_sys_shared_add_f32, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_release_sys_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_release_sys_shared_add_f32, memory_order_release, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_acq_rel_sys_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint); atomic_add(atom_acq_rel_sys_shared_add_f32, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, float, atomic_uint, uint);
atomic_add(atom_relaxed_cta_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, , double, atomic_ulong, ulong); atomic_add(atom_relaxed_cta_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, , double, atomic_ulong, ulong);
atomic_add(atom_acquire_cta_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong); atomic_add(atom_acquire_cta_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong);
atomic_add(atom_release_cta_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong); atomic_add(atom_release_cta_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_cta_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong); atomic_add(atom_acq_rel_cta_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, , double, atomic_ulong, ulong);
atomic_add(atom_relaxed_gpu_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_relaxed_gpu_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_acquire_gpu_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_acquire_gpu_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_release_gpu_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_release_gpu_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_gpu_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_acq_rel_gpu_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_relaxed_sys_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_relaxed_sys_generic_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_acquire_sys_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_acquire_sys_generic_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_release_sys_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_release_sys_generic_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_sys_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong); atomic_add(atom_acq_rel_sys_generic_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, , double, atomic_ulong, ulong);
// atom.add.f64 // atom.add.f64
atomic_add(atom_relaxed_cta_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global, double, atomic_ulong, ulong); atomic_add(atom_relaxed_cta_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global, double, atomic_ulong, ulong);
atomic_add(atom_acquire_cta_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong); atomic_add(atom_acquire_cta_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong);
atomic_add(atom_release_cta_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong); atomic_add(atom_release_cta_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_cta_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_cta_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global, double, atomic_ulong, ulong);
atomic_add(atom_relaxed_gpu_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_relaxed_gpu_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_acquire_gpu_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_acquire_gpu_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_release_gpu_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_release_gpu_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_gpu_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_gpu_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_relaxed_sys_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_relaxed_sys_global_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_acquire_sys_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_acquire_sys_global_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_release_sys_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_release_sys_global_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_sys_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_sys_global_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global, double, atomic_ulong, ulong);
atomic_add(atom_relaxed_cta_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local, double, atomic_ulong, ulong); atomic_add(atom_relaxed_cta_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local, double, atomic_ulong, ulong);
atomic_add(atom_acquire_cta_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong); atomic_add(atom_acquire_cta_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong);
atomic_add(atom_release_cta_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong); atomic_add(atom_release_cta_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_cta_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_cta_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __local, double, atomic_ulong, ulong);
atomic_add(atom_relaxed_gpu_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_relaxed_gpu_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_acquire_gpu_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_acquire_gpu_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_release_gpu_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_release_gpu_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_gpu_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_gpu_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_relaxed_sys_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_relaxed_sys_shared_add_f64, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_acquire_sys_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_acquire_sys_shared_add_f64, memory_order_acquire, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_release_sys_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_release_sys_shared_add_f64, memory_order_release, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
atomic_add(atom_acq_rel_sys_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong); atomic_add(atom_acq_rel_sys_shared_add_f64, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local, double, atomic_ulong, ulong);
uint FUNC(bfe_u32)(uint base, uint pos, uint len) { uint FUNC(bfe_u32)(uint base, uint pos, uint len) {
return intel_ubfe(base, pos, len); return intel_ubfe(base, pos, len);
} }
ulong FUNC(bfe_u64)(ulong base, uint pos, uint len) { ulong FUNC(bfe_u64)(ulong base, uint pos, uint len) {
return intel_ubfe(base, pos, len); return intel_ubfe(base, pos, len);
} }
int FUNC(bfe_s32)(int base, uint pos, uint len) { int FUNC(bfe_s32)(int base, uint pos, uint len) {
return intel_sbfe(base, pos, len); return intel_sbfe(base, pos, len);
} }
long FUNC(bfe_s64)(long base, uint pos, uint len) { long FUNC(bfe_s64)(long base, uint pos, uint len) {
return intel_sbfe(base, pos, len); return intel_sbfe(base, pos, len);
} }
uint FUNC(bfi_b32)(uint insert, uint base, uint offset, uint count) { uint FUNC(bfi_b32)(uint insert, uint base, uint offset, uint count) {
return intel_bfi(base, insert, offset, count); return intel_bfi(base, insert, offset, count);
} }
ulong FUNC(bfi_b64)(ulong insert, ulong base, uint offset, uint count) { ulong FUNC(bfi_b64)(ulong insert, ulong base, uint offset, uint count) {
return intel_bfi(base, insert, offset, count); return intel_bfi(base, insert, offset, count);
} }
uint FUNC(brev_b32)(uint base) { uint FUNC(brev_b32)(uint base) {
return intel_bfrev(base); return intel_bfrev(base);
} }
ulong FUNC(brev_b64)(ulong base) { ulong FUNC(brev_b64)(ulong base) {
return intel_bfrev(base); return intel_bfrev(base);
} }
#else
uint FUNC(bfe_u32)(uint base, uint pos, uint len) {
return amd_bfe(base, pos, len);
}
extern __attribute__((const)) int __llvm_bitreverse_i32(int) __asm("llvm.bitreverse.i32");
uint FUNC(brev_b32)(uint base) {
return __llvm_bitreverse_i32(base);
}
#endif
void FUNC(__assertfail)( void FUNC(__assertfail)(
__private ulong* message, __private ulong* message,

2
ptx/src/test/spirv_run/mod.rs
View file

@ -251,7 +251,7 @@ fn run_spirv<Input: From<u8> + Copy + Debug, Output: From<u8> + Copy + Debug + D
let ctx = ze::Context::new(drv, None)?; let ctx = ze::Context::new(drv, None)?;
let queue = ze::CommandQueue::new(&ctx, dev)?; let queue = ze::CommandQueue::new(&ctx, dev)?;
let (module, maybe_log) = match module.should_link_ptx_impl { let (module, maybe_log) = match module.should_link_ptx_impl {
Some(ptx_impl) => ze::Module::build_link_spirv( Some((ptx_impl, _)) => ze::Module::build_link_spirv(
&ctx, &ctx,
dev, dev,
&[ptx_impl, byte_il], &[ptx_impl, byte_il],

7
ptx/src/translate.rs
View file

@ -7,7 +7,8 @@ use std::{borrow::Cow, collections::BTreeSet, ffi::CString, hash::Hash, iter, me
use rspirv::binary::Assemble; use rspirv::binary::Assemble;
static ZLUDA_PTX_IMPL: &'static [u8] = include_bytes!("../lib/zluda_ptx_impl.spv"); static ZLUDA_PTX_IMPL_INTEL: &'static [u8] = include_bytes!("../lib/zluda_ptx_impl.spv");
static ZLUDA_PTX_IMPL_AMD: &'static [u8] = include_bytes!("../lib/zluda_ptx_impl.bc");
static ZLUDA_PTX_PREFIX: &'static str = "__zluda_ptx_impl__"; static ZLUDA_PTX_PREFIX: &'static str = "__zluda_ptx_impl__";
quick_error! { quick_error! {
@ -405,7 +406,7 @@ impl TypeWordMap {
pub struct Module { pub struct Module {
pub spirv: dr::Module, pub spirv: dr::Module,
pub kernel_info: HashMap<String, KernelInfo>, pub kernel_info: HashMap<String, KernelInfo>,
pub should_link_ptx_impl: Option<&'static [u8]>, pub should_link_ptx_impl: Option<(&'static [u8], &'static [u8])>,
pub build_options: CString, pub build_options: CString,
} }
impl Module { impl Module {
@ -466,7 +467,7 @@ pub fn to_spirv_module<'a>(ast: ast::Module<'a>) -> Result<Module, TranslateErro
spirv, spirv,
kernel_info, kernel_info,
should_link_ptx_impl: if must_link_ptx_impl { should_link_ptx_impl: if must_link_ptx_impl {
Some(ZLUDA_PTX_IMPL) Some((ZLUDA_PTX_IMPL_INTEL, ZLUDA_PTX_IMPL_AMD))
} else { } else {
None None
}, },

94
zluda/src/impl/module.rs
View file

@ -7,6 +7,7 @@ use std::{
io::{self, Write}, io::{self, Write},
mem, mem,
os::raw::{c_char, c_int, c_uint}, os::raw::{c_char, c_int, c_uint},
path::PathBuf,
process::{Command, Stdio}, process::{Command, Stdio},
ptr, slice, ptr, slice,
}; };
@ -49,7 +50,7 @@ pub struct ModuleData {
pub struct SpirvModule { pub struct SpirvModule {
pub binaries: Vec<u32>, pub binaries: Vec<u32>,
pub kernel_info: HashMap<String, ptx::KernelInfo>, pub kernel_info: HashMap<String, ptx::KernelInfo>,
pub should_link_ptx_impl: Option<&'static [u8]>, pub should_link_ptx_impl: Option<(&'static [u8], &'static [u8])>,
pub build_options: CString, pub build_options: CString,
} }
@ -93,31 +94,93 @@ impl SpirvModule {
const LLVM_SPIRV: &'static str = "/home/vosen/amd/llvm-project/build/bin/llvm-spirv"; const LLVM_SPIRV: &'static str = "/home/vosen/amd/llvm-project/build/bin/llvm-spirv";
const AMDGPU: &'static str = "/opt/amdgpu-pro/"; const AMDGPU: &'static str = "/opt/amdgpu-pro/";
const AMDGPU_TARGET: &'static str = "amdgcn-amd-amdhsa";
const AMDGPU_BITCODE: [&'static str; 8] = [ const AMDGPU_BITCODE: [&'static str; 8] = [
"opencl", "opencl.bc",
"ocml", "ocml.bc",
"ockl", "ockl.bc",
"oclc_correctly_rounded_sqrt_off", "oclc_correctly_rounded_sqrt_off.bc",
"oclc_daz_opt_on", "oclc_daz_opt_on.bc",
"oclc_finite_only_off", "oclc_finite_only_off.bc",
"oclc_unsafe_math_off", "oclc_unsafe_math_off.bc",
"oclc_wavefrontsize64_off", "oclc_wavefrontsize64_off.bc",
]; ];
const AMDGPU_BITCODE_DEVICE_PREFIX: &'static str = "oclc_isa_version_"; const AMDGPU_BITCODE_DEVICE_PREFIX: &'static str = "oclc_isa_version_";
const AMDGPU_DEVICE: &'static str = "gfx1010"; const AMDGPU_DEVICE: &'static str = "gfx1010";
fn compile_amd(spirv_il: &[u8]) -> io::Result<()> { fn get_bitcode_paths() -> impl Iterator<Item = PathBuf> {
let generic_paths = Self::AMDGPU_BITCODE.iter().map(|x| {
let mut path = PathBuf::from(Self::AMDGPU);
path.push("amdgcn");
path.push("bitcode");
path.push(x);
path
});
let mut additional_path = PathBuf::from(Self::AMDGPU);
additional_path.push("amdgcn");
additional_path.push("bitcode");
additional_path.push(format!(
"{}{}{}",
Self::AMDGPU_BITCODE_DEVICE_PREFIX,
&Self::AMDGPU_DEVICE[3..],
".bc"
));
generic_paths.chain(std::iter::once(additional_path))
}
fn compile_amd(spirv_il: &[u8], ptx_lib: Option<&'static [u8]>) -> io::Result<()> {
let dir = tempfile::tempdir()?; let dir = tempfile::tempdir()?;
let mut spirv = NamedTempFile::new_in(&dir)?; let mut spirv = NamedTempFile::new_in(&dir)?;
let llvm = NamedTempFile::new_in(&dir)?; let llvm = NamedTempFile::new_in(&dir)?;
spirv.write_all(spirv_il)?; spirv.write_all(spirv_il)?;
let mut cmd = Command::new(Self::LLVM_SPIRV) let to_llvm_cmd = Command::new(Self::LLVM_SPIRV)
.arg("-r") .arg("-r")
.arg("-o") .arg("-o")
.arg(llvm.path()) .arg(llvm.path())
.arg(spirv.path()) .arg(spirv.path())
.status()?; .status()?;
assert!(cmd.success()); assert!(to_llvm_cmd.success());
let linked_binary = NamedTempFile::new_in(&dir)?;
let mut llvm_link = PathBuf::from(Self::AMDGPU);
llvm_link.push("bin");
llvm_link.push("llvm-link");
let mut linker_cmd = Command::new(&llvm_link);
linker_cmd
.arg("--only-needed")
.arg("-o")
.arg(linked_binary.path())
.arg(llvm.path())
.args(Self::get_bitcode_paths());
if cfg!(debug_assertions) {
linker_cmd.arg("-v");
}
let status = linker_cmd.status()?;
assert!(status.success());
let mut ptx_lib_bitcode = NamedTempFile::new_in(&dir)?;
let compiled_binary = NamedTempFile::new_in(&dir)?;
let mut cland_exe = PathBuf::from(Self::AMDGPU);
cland_exe.push("bin");
cland_exe.push("clang");
let mut compiler_cmd = Command::new(&cland_exe);
compiler_cmd
.arg(format!("-mcpu={}", Self::AMDGPU_DEVICE))
.arg("-O3")
.arg("-Xlinker")
.arg("--no-undefined")
.arg("-target")
.arg(Self::AMDGPU_TARGET)
.arg("-o")
.arg(compiled_binary.path())
.arg(linked_binary.path());
if let Some(bitcode) = ptx_lib {
ptx_lib_bitcode.write_all(bitcode)?;
compiler_cmd.arg(ptx_lib_bitcode.path());
};
if cfg!(debug_assertions) {
compiler_cmd.arg("-v");
}
let status = compiler_cmd.status()?;
assert!(status.success());
Ok(()) Ok(())
} }
@ -132,10 +195,10 @@ impl SpirvModule {
self.binaries.len() * mem::size_of::<u32>(), self.binaries.len() * mem::size_of::<u32>(),
) )
}; };
Self::compile_amd(byte_il).unwrap();
let main_module = ocl_core::create_program_with_il(ctx, byte_il, None)?; let main_module = ocl_core::create_program_with_il(ctx, byte_il, None)?;
let main_module = match self.should_link_ptx_impl { let main_module = match self.should_link_ptx_impl {
None => { None => {
Self::compile_amd(byte_il, None).unwrap();
ocl_core::build_program( ocl_core::build_program(
&main_module, &main_module,
Some(&[dev]), Some(&[dev]),
@ -145,8 +208,9 @@ impl SpirvModule {
)?; )?;
main_module main_module
} }
Some(ptx_impl) => { Some((ptx_impl_intel, ptx_impl_amd)) => {
let ptx_impl_prog = ocl_core::create_program_with_il(ctx, ptx_impl, None)?; Self::compile_amd(byte_il, Some(ptx_impl_amd)).unwrap();
let ptx_impl_prog = ocl_core::create_program_with_il(ctx, ptx_impl_intel, None)?;
ocl_core::compile_program( ocl_core::compile_program(
&main_module, &main_module,
Some(&[dev]), Some(&[dev]),