[AMD] Remove GlobalPrefetch and LocalPrefetch Knobs (#8295)

Removed `TRITON_HIP_GLOBAL_PREFETCH` and `TRITON_HIP_LOCAL_PREFETCH` to
simplify pipeliner as no one is practically using it. This prepares us from upcoming
restructuring of the pipeliner.

Author: knwng

include/triton/Tools/Sys/GetEnv.hpp

@@ -32,8 +32,6 @@ inline const std::set<std::string> CACHE_INVALIDATING_ENV_VARS = {
     "TRITON_DEFAULT_FP_FUSION",
     "TRITON_DISABLE_LINE_INFO",
     "TRITON_ENABLE_LLVM_DEBUG",
-    "TRITON_HIP_GLOBAL_PREFETCH",
-    "TRITON_HIP_LOCAL_PREFETCH",
     "TRITON_HIP_USE_ASYNC_COPY",
     "TRITON_HIP_USE_BLOCK_PINGPONG",
     "TRITON_HIP_USE_IN_THREAD_TRANSPOSE",

python/test/unit/test_knobs.py

@@ -69,15 +69,12 @@ class test_knobs(triton.knobs.base_knobs):
 
 
 def test_knobs_scope(fresh_knobs, monkeypatch):
-    fresh_knobs.amd.global_prefetch = 4
-    fresh_knobs.amd.local_prefetch = 3
+    fresh_knobs.amd.use_buffer_atomics = True
 
     # Update env *after* the __set__() does
-    monkeypatch.setenv("TRITON_HIP_LOCAL_PREFETCH", "17")
+    monkeypatch.setenv("AMDGCN_USE_BUFFER_ATOMICS", "0")
 
-    assert fresh_knobs.amd.global_prefetch == 4
-    assert fresh_knobs.amd.local_prefetch == 3
-    assert fresh_knobs.amd.use_buffer_ops
+    assert fresh_knobs.amd.use_buffer_atomics
 
     # Just to prove that use_buffer_ops is coming from env
     monkeypatch.setenv("AMDGCN_USE_BUFFER_OPS", "0")
@@ -86,17 +83,14 @@ def test_knobs_scope(fresh_knobs, monkeypatch):
     assert fresh_knobs.amd.use_buffer_ops
 
     with fresh_knobs.amd.scope():
-        fresh_knobs.amd.global_prefetch = 5
         # Use the environment
-        del fresh_knobs.amd.local_prefetch
+        del fresh_knobs.amd.use_buffer_atomics
         fresh_knobs.amd.use_buffer_ops = False
 
-        assert fresh_knobs.amd.global_prefetch == 5
-        assert fresh_knobs.amd.local_prefetch == 17
+        assert not fresh_knobs.amd.use_buffer_atomics
         assert not fresh_knobs.amd.use_buffer_ops
 
-    assert fresh_knobs.amd.global_prefetch == 4
-    assert fresh_knobs.amd.local_prefetch == 3
+    assert fresh_knobs.amd.use_buffer_atomics
     assert fresh_knobs.amd.use_buffer_ops
 
     # Just to prove that use_buffer_ops is coming from env
@@ -200,7 +194,7 @@ def test_set_knob_directly(fresh_knobs, monkeypatch):
     # Just in case, lets check all the other datatypes too
     fresh_knobs.language.default_fp_fusion = False
     fresh_knobs.amd.use_block_pingpong = True
-    fresh_knobs.amd.global_prefetch = 5
+    fresh_knobs.redis.port = 6380
     fresh_knobs.nvidia.mock_ptx_version = "42.0.1"
 
     from triton.runtime.cache import FileCacheManager
@@ -213,27 +207,27 @@ class TestManagerClass(FileCacheManager):
     monkeypatch.setenv("TRITON_CUDART_PATH", "/tmp/the/real/cudart")
     monkeypatch.setenv("TRITON_DEFAULT_FP_FUSION", "1")
     monkeypatch.setenv("TRITON_HIP_USE_BLOCK_PINGPONG", "0")
-    monkeypatch.setenv("TRITON_HIP_GLOBAL_PREFETCH", "2")
+    monkeypatch.setenv("TRITON_REDIS_PORT", "6381")
     monkeypatch.setenv("TRITON_MOCK_PTX_VERSION", "1.0.0")
     monkeypatch.setenv("TRITON_CACHE_MANAGER", "triton.runtime.cache:FileCacheManager")
 
     assert not fresh_knobs.language.default_fp_fusion
     assert fresh_knobs.amd.use_block_pingpong
-    assert fresh_knobs.amd.global_prefetch == 5
+    assert fresh_knobs.redis.port == 6380
     assert fresh_knobs.nvidia.mock_ptx_version == "42.0.1"
     assert fresh_knobs.cache.manager_class == TestManagerClass
 
     # Make sure both setting `.env` or deleting resets to env vars.
     fresh_knobs.language.default_fp_fusion = fresh_knobs.env
     fresh_knobs.amd.use_block_pingpong = fresh_knobs.env
-    fresh_knobs.amd.global_prefetch = fresh_knobs.env
+    fresh_knobs.redis.port = fresh_knobs.env
     del fresh_knobs.nvidia.mock_ptx_version
     del fresh_knobs.cache.manager_class
 
     assert fresh_knobs.build.backend_dirs == {"/tmp/the/real/cudart"}
     assert fresh_knobs.language.default_fp_fusion
     assert not fresh_knobs.amd.use_block_pingpong
-    assert fresh_knobs.amd.global_prefetch == 2
+    assert fresh_knobs.redis.port == 6381
     assert fresh_knobs.nvidia.mock_ptx_version == "1.0.0"
     assert fresh_knobs.cache.manager_class == FileCacheManager
 

python/triton/knobs.py

@@ -508,8 +508,6 @@ class amd_knobs(base_knobs):
     use_block_pingpong: env_opt_bool = env_opt_bool("TRITON_HIP_USE_BLOCK_PINGPONG")
     use_in_thread_transpose: env_opt_bool = env_opt_bool("TRITON_HIP_USE_IN_THREAD_TRANSPOSE")
 
-    global_prefetch: env_int = env_int("TRITON_HIP_GLOBAL_PREFETCH")
-    local_prefetch: env_int = env_int("TRITON_HIP_LOCAL_PREFETCH")
     use_async_copy: env_bool = env_bool("TRITON_HIP_USE_ASYNC_COPY")
     scalarize_packed_fops: env_bool = env_bool("AMDGCN_SCALARIZE_PACKED_FOPS")