diff options
| author | Andrzej Janik <[email protected]> | 2020-11-23 20:00:16 +0100 |
|---|---|---|
| committer | Andrzej Janik <[email protected]> | 2020-11-23 20:01:10 +0100 |
| commit | eb7c9aeeee4cf20a4d81a319f4ebf9329efb819f (patch) | |
| tree | af96435ca89cb36ffd414be99acdd1f91c701787 /ptx/lib/zluda_ptx_impl.cl | |
| parent | 0415f873ae15aceda87758f9fcce5d38aa68ca75 (diff) | |
| download | ZLUDA-eb7c9aeeee4cf20a4d81a319f4ebf9329efb819f.tar.gz ZLUDA-eb7c9aeeee4cf20a4d81a319f4ebf9329efb819f.zip | |
Rename everything
Diffstat (limited to 'ptx/lib/zluda_ptx_impl.cl')
| -rw-r--r-- | ptx/lib/zluda_ptx_impl.cl | 146 |
1 files changed, 146 insertions, 0 deletions
diff --git a/ptx/lib/zluda_ptx_impl.cl b/ptx/lib/zluda_ptx_impl.cl new file mode 100644 index 0000000..94d02ec --- /dev/null +++ b/ptx/lib/zluda_ptx_impl.cl @@ -0,0 +1,146 @@ +// 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
+// Additionally you should strip names:
+// spirv-opt --strip-debug zluda_ptx_impl.spv -o zluda_ptx_impl.spv
+
+#define FUNC(NAME) __zluda_ptx_impl__ ## NAME
+
+#define atomic_inc(NAME, SUCCESS, FAILURE, SCOPE, SPACE) \
+ uint FUNC(NAME)(SPACE uint* ptr, uint threshold) { \
+ uint expected = *ptr; \
+ uint desired; \
+ do { \
+ desired = (expected >= threshold) ? 0 : expected + 1; \
+ } while (!atomic_compare_exchange_strong_explicit((volatile SPACE atomic_uint*)ptr, &expected, desired, SUCCESS, FAILURE, SCOPE)); \
+ return expected; \
+ }
+
+#define atomic_dec(NAME, SUCCESS, FAILURE, SCOPE, SPACE) \
+ uint FUNC(NAME)(SPACE uint* ptr, uint threshold) { \
+ uint expected = *ptr; \
+ uint desired; \
+ do { \
+ desired = (expected == 0 || expected > threshold) ? threshold : expected - 1; \
+ } while (!atomic_compare_exchange_strong_explicit((volatile SPACE atomic_uint*)ptr, &expected, desired, SUCCESS, FAILURE, SCOPE)); \
+ return expected; \
+ }
+
+// We are doing all this mess instead of accepting memory_order and memory_scope parameters
+// because ocloc emits broken (failing spirv-dis) SPIR-V when memory_order or memory_scope is a parameter
+
+// atom.inc
+atomic_inc(atom_relaxed_cta_generic_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, );
+atomic_inc(atom_acquire_cta_generic_inc, memory_order_acquire, memory_order_acquire, memory_scope_work_group, );
+atomic_inc(atom_release_cta_generic_inc, memory_order_release, memory_order_acquire, memory_scope_work_group, );
+atomic_inc(atom_acq_rel_cta_generic_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, );
+
+atomic_inc(atom_relaxed_gpu_generic_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, );
+atomic_inc(atom_acquire_gpu_generic_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, );
+atomic_inc(atom_release_gpu_generic_inc, memory_order_release, memory_order_acquire, memory_scope_device, );
+atomic_inc(atom_acq_rel_gpu_generic_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, );
+
+atomic_inc(atom_relaxed_sys_generic_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, );
+atomic_inc(atom_acquire_sys_generic_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, );
+atomic_inc(atom_release_sys_generic_inc, memory_order_release, memory_order_acquire, memory_scope_device, );
+atomic_inc(atom_acq_rel_sys_generic_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, );
+
+atomic_inc(atom_relaxed_cta_global_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global);
+atomic_inc(atom_acquire_cta_global_inc, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global);
+atomic_inc(atom_release_cta_global_inc, memory_order_release, memory_order_acquire, memory_scope_work_group, __global);
+atomic_inc(atom_acq_rel_cta_global_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global);
+
+atomic_inc(atom_relaxed_gpu_global_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global);
+atomic_inc(atom_acquire_gpu_global_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, __global);
+atomic_inc(atom_release_gpu_global_inc, memory_order_release, memory_order_acquire, memory_scope_device, __global);
+atomic_inc(atom_acq_rel_gpu_global_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global);
+
+atomic_inc(atom_relaxed_sys_global_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global);
+atomic_inc(atom_acquire_sys_global_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, __global);
+atomic_inc(atom_release_sys_global_inc, memory_order_release, memory_order_acquire, memory_scope_device, __global);
+atomic_inc(atom_acq_rel_sys_global_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global);
+
+atomic_inc(atom_relaxed_cta_shared_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local);
+atomic_inc(atom_acquire_cta_shared_inc, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local);
+atomic_inc(atom_release_cta_shared_inc, memory_order_release, memory_order_acquire, memory_scope_work_group, __local);
+atomic_inc(atom_acq_rel_cta_shared_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __local);
+
+atomic_inc(atom_relaxed_gpu_shared_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local);
+atomic_inc(atom_acquire_gpu_shared_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, __local);
+atomic_inc(atom_release_gpu_shared_inc, memory_order_release, memory_order_acquire, memory_scope_device, __local);
+atomic_inc(atom_acq_rel_gpu_shared_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local);
+
+atomic_inc(atom_relaxed_sys_shared_inc, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __local);
+atomic_inc(atom_acquire_sys_shared_inc, memory_order_acquire, memory_order_acquire, memory_scope_device, __local);
+atomic_inc(atom_release_sys_shared_inc, memory_order_release, memory_order_acquire, memory_scope_device, __local);
+atomic_inc(atom_acq_rel_sys_shared_inc, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __local);
+
+// atom.dec
+atomic_dec(atom_relaxed_cta_generic_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, );
+atomic_dec(atom_acquire_cta_generic_dec, memory_order_acquire, memory_order_acquire, memory_scope_work_group, );
+atomic_dec(atom_release_cta_generic_dec, memory_order_release, memory_order_acquire, memory_scope_work_group, );
+atomic_dec(atom_acq_rel_cta_generic_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, );
+
+atomic_dec(atom_relaxed_gpu_generic_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, );
+atomic_dec(atom_acquire_gpu_generic_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, );
+atomic_dec(atom_release_gpu_generic_dec, memory_order_release, memory_order_acquire, memory_scope_device, );
+atomic_dec(atom_acq_rel_gpu_generic_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, );
+
+atomic_dec(atom_relaxed_sys_generic_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, );
+atomic_dec(atom_acquire_sys_generic_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, );
+atomic_dec(atom_release_sys_generic_dec, memory_order_release, memory_order_acquire, memory_scope_device, );
+atomic_dec(atom_acq_rel_sys_generic_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, );
+
+atomic_dec(atom_relaxed_cta_global_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __global);
+atomic_dec(atom_acquire_cta_global_dec, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __global);
+atomic_dec(atom_release_cta_global_dec, memory_order_release, memory_order_acquire, memory_scope_work_group, __global);
+atomic_dec(atom_acq_rel_cta_global_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __global);
+
+atomic_dec(atom_relaxed_gpu_global_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global);
+atomic_dec(atom_acquire_gpu_global_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, __global);
+atomic_dec(atom_release_gpu_global_dec, memory_order_release, memory_order_acquire, memory_scope_device, __global);
+atomic_dec(atom_acq_rel_gpu_global_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global);
+
+atomic_dec(atom_relaxed_sys_global_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_device, __global);
+atomic_dec(atom_acquire_sys_global_dec, memory_order_acquire, memory_order_acquire, memory_scope_device, __global);
+atomic_dec(atom_release_sys_global_dec, memory_order_release, memory_order_acquire, memory_scope_device, __global);
+atomic_dec(atom_acq_rel_sys_global_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_device, __global);
+
+atomic_dec(atom_relaxed_cta_shared_dec, memory_order_relaxed, memory_order_relaxed, memory_scope_work_group, __local);
+atomic_dec(atom_acquire_cta_shared_dec, memory_order_acquire, memory_order_acquire, memory_scope_work_group, __local);
+atomic_dec(atom_release_cta_shared_dec, memory_order_release, memory_order_acquire, memory_scope_work_group, __local);
+atomic_dec(atom_acq_rel_cta_shared_dec, memory_order_acq_rel, memory_order_acquire, memory_scope_work_group, __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_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);
+
+uint FUNC(bfe_u32)(uint base, uint pos, uint len) {
+ return intel_ubfe(base, pos, len);
+}
+
+ulong FUNC(bfe_u64)(ulong base, uint pos, uint len) {
+ return intel_ubfe(base, pos, len);
+}
+
+int FUNC(bfe_s32)(int base, uint pos, uint len) {
+ return intel_sbfe(base, pos, len);
+}
+
+long FUNC(bfe_s64)(long base, uint pos, uint len) {
+ return intel_sbfe(base, pos, len);
+}
+
+void FUNC(__assertfail)(
+ __private ulong* message,
+ __private ulong* file,
+ __private uint* line,
+ __private ulong* function,
+ __private ulong* charSize
+) {
+}
|