Add gfx900:xnack-, gfx1032, gfx1034, gfx1035 (#1595)

* Allow gfx900:xnack- The miscommunication relating to whether gfx900 should be used with or without gfx900:xnack- was never properly resolved and so many of the libraries that depend on rocBLAS use gfx900:xnack- in their default configuration. This change allows both gfx900 and gfx900:xnack- to be accepted by Tensile, thereby allowing users that are building from source to choose whatever they prefer. * Add gfx1032, gfx1034, gfx1035 This change allows users to build rocBLAS from source for Navi23, Navi24 and Rembrandt.
ROCm · Sep 23, 2022 · 41236e3 · 41236e3
1 parent 6cfd8a7
commit 41236e3
Show file tree

Hide file tree

Showing 5 changed files with 22 additions and 4 deletions.
diff --git a/Tensile/Common.py b/Tensile/Common.py
@@ -271,12 +271,12 @@
 
 # Translate GPU targets to filter filenames in Tensile_LOGIC directory
 architectureMap = {
-  'all':'_','gfx000':'none', 'gfx803':'r9nano', 'gfx900':'vega10',
+  'all':'_','gfx000':'none', 'gfx803':'r9nano', 'gfx900':'vega10', 'gfx900:xnack-':'vega10',
   'gfx906':'vega20', 'gfx906:xnack+':'vega20', 'gfx906:xnack-':'vega20',
   'gfx908':'arcturus','gfx908:xnack+':'arcturus', 'gfx908:xnack-':'arcturus',
   'gfx90a':'aldebaran', 'gfx90a:xnack+':'aldebaran', 'gfx90a:xnack-':'aldebaran',
   'gfx1010':'navi10', 'gfx1011':'navi12', 'gfx1012':'navi14',
-  'gfx1030':'navi21', 'gfx1031':'navi22',
+  'gfx1030':'navi21', 'gfx1031':'navi22', 'gfx1032':'navi23', 'gfx1034':'navi24', 'gfx1035':'rembrandt',
   'gfx1100':'navi31', 'gfx1101':'navi32', 'gfx1102':'navi33'
 }
 

diff --git a/Tensile/Source/CMakeLists.txt b/Tensile/Source/CMakeLists.txt
@@ -51,9 +51,9 @@ if(NOT DEFINED CXX_VERSION_STRING)
 endif()
 
 if(CMAKE_CXX_COMPILER STREQUAL "hipcc")
-  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906:xnack- gfx908:xnack- gfx90a:xnack- gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1100 gfx1101 gfx1102 CACHE STRING "GPU architectures")
+  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906:xnack- gfx908:xnack- gfx90a:xnack- gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 CACHE STRING "GPU architectures")
 else()
-  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906 gfx908 gfx90a gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1100 gfx1101 gfx1102 CACHE STRING "GPU architectures")
+  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906 gfx908 gfx90a gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 CACHE STRING "GPU architectures")
 endif()
 
 include(CMakeDependentOption)

diff --git a/Tensile/Source/lib/include/Tensile/AMDGPU.hpp b/Tensile/Source/lib/include/Tensile/AMDGPU.hpp
@@ -64,6 +64,9 @@ namespace Tensile
             gfx1012 = 1012,
             gfx1030 = 1030,
             gfx1031 = 1031,
+            gfx1032 = 1032,
+            gfx1034 = 1034,
+            gfx1035 = 1035,
             gfx1100 = 1100,
             gfx1101 = 1101,
             gfx1102 = 1102
@@ -93,6 +96,12 @@ namespace Tensile
                 return "gfx1030";
             case AMDGPU::Processor::gfx1031:
                 return "gfx1031";
+            case AMDGPU::Processor::gfx1032:
+                return "gfx1032";
+            case AMDGPU::Processor::gfx1034:
+                return "gfx1034";
+            case AMDGPU::Processor::gfx1035:
+                return "gfx1035";
             case AMDGPU::Processor::gfx1100:
                 return "gfx1100";
             case AMDGPU::Processor::gfx1101:

diff --git a/Tensile/Source/lib/include/Tensile/Serialization/Predicates.hpp b/Tensile/Source/lib/include/Tensile/Serialization/Predicates.hpp
@@ -216,6 +216,9 @@ namespace Tensile
                 iot::enumCase(io, value, "gfx1012", AMDGPU::Processor::gfx1012);
                 iot::enumCase(io, value, "gfx1030", AMDGPU::Processor::gfx1030);
                 iot::enumCase(io, value, "gfx1031", AMDGPU::Processor::gfx1031);
+                iot::enumCase(io, value, "gfx1032", AMDGPU::Processor::gfx1032);
+                iot::enumCase(io, value, "gfx1034", AMDGPU::Processor::gfx1034);
+                iot::enumCase(io, value, "gfx1035", AMDGPU::Processor::gfx1035);
                 iot::enumCase(io, value, "gfx1100", AMDGPU::Processor::gfx1100);
                 iot::enumCase(io, value, "gfx1101", AMDGPU::Processor::gfx1101);
                 iot::enumCase(io, value, "gfx1102", AMDGPU::Processor::gfx1102);

diff --git a/pytest.ini b/pytest.ini
@@ -90,6 +90,9 @@ markers =
  xfail-gfx1012: architecture
  xfail-gfx1030: architecture
  xfail-gfx1031: architecture
+ xfail-gfx1032: architecture
+ xfail-gfx1034: architecture
+ xfail-gfx1035: architecture
  xfail-gfx1100: architecture
  xfail-gfx1101: architecture
  xfail-gfx1102: architecture
@@ -103,6 +106,9 @@ markers =
  skip-gfx1012: architecture
  skip-gfx1030: architecture
  skip-gfx1031: architecture
+ skip-gfx1032: architecture
+ skip-gfx1034: architecture
+ skip-gfx1035: architecture
  skip-gfx1100: architecture
  skip-gfx1101: architecture
  skip-gfx1102: architecture