Add gfx1150 and complete gfx1151 support (#2154)

Author: amd-mtrifuno

CHANGELOG.md

@@ -9,7 +9,7 @@
 - venv for Tensile create on Linux
 - Flag to keep build_tmp when running Tensile
 - Generalized profiling scripts
-- GFX1151 support
+- GFX1151 and GFX1150 support
 - Single-threaded support in TensileCreateLibrary
 - Logic to remove temporary build artifacts
 

Tensile/AsmCaps.py

@@ -771,6 +771,50 @@
               'v_mov_b64': False,
               'v_pk_fma_f16': True,
               'v_pk_fmac_f16': False},
+ (11, 5, 0): {'HasAddLshl': True,
+              'HasAtomicAdd': True,
+              'HasDirectToLdsDest': False,
+              'HasDirectToLdsNoDest': False,
+              'HasExplicitCO': True,
+              'HasExplicitNC': True,
+              'HasGLCModifier': True,
+              'HasNTModifier': False,
+              'HasLshlOr': True,
+              'HasMFMA': False,
+              'HasMFMA_b8': False,
+              'HasMFMA_bf16_1k': False,
+              'HasMFMA_bf16_original': False,
+              'HasMFMA_constSrc': False,
+              'HasMFMA_f64': False,
+              'HasMFMA_f8': False,
+              'HasMFMA_i8_908': False,
+              'HasMFMA_i8_940': False,
+              'HasMFMA_vgpr': False,
+              'HasMFMA_xf32': False,
+              'HasSMulHi': True,
+              'HasWMMA': True,
+              'KernargPreloading': False,
+              'MaxLgkmcnt': 15,
+              'MaxVmcnt': 63,
+              'SupportedISA': True,
+              'SupportedSource': True,
+              'VOP3v_dot4_i32_i8': False,
+              'v_dot2_f32_f16': True,
+              'v_dot2c_f32_f16': True,
+              'v_dot4_i32_i8': False,
+              'v_dot4c_i32_i8': False,
+              'v_fma_f16': True,
+              'v_fma_f32': True,
+              'v_fma_f64': True,
+              'v_fma_mix_f32': True,
+              'v_fmac_f16': False,
+              'v_fmac_f32': True,
+              'v_mac_f16': False,
+              'v_mac_f32': False,
+              'v_mad_mix_f32': False,
+              'v_mov_b64': False,
+              'v_pk_fma_f16': True,
+              'v_pk_fmac_f16': False},
  (11, 5, 1): {'HasAddLshl': True,
               'HasAtomicAdd': True,
               'HasDirectToLdsDest': False,

Tensile/Common.py

@@ -247,7 +247,7 @@ class DeveloperWarning(Warning):
                                     (9,4,0), (9,4,1), (9,4,2),
                                     (10,1,0), (10,1,1), (10,1,2), (10,3,0), (10,3,1),
                                     (11,0,0), (11,0,1), (11,0,2),
-                                    (11,5,1),
+                                    (11,5,0), (11,5,1),
                                     (12,0,0), (12,0,1)] # assembly kernels writer supports these architectures
 
 globalParameters["KeepBuildTmp"] = True                           # Do not remove build artifacts during the build process or build_tmp after build completes
@@ -324,7 +324,7 @@ class DeveloperWarning(Warning):
   'gfx1010':'navi10', 'gfx1011':'navi12', 'gfx1012':'navi14',
   'gfx1030':'navi21', 'gfx1031':'navi22', 'gfx1032':'navi23', 'gfx1034':'navi24', 'gfx1035':'rembrandt',
   'gfx1100':'navi31', 'gfx1101':'navi32', 'gfx1102':'navi33',
-  'gfx1151':'gfx1151',
+  'gfx1150':'strixpoint', 'gfx1151':'strixhalo',
   'gfx1200':'gfx1200',
   'gfx1201':'gfx1201'
 }
@@ -2466,7 +2466,7 @@ def assignGlobalParameters( config, capabilitiesCache: Optional[dict] = None ):
     if os.name == "nt":
       globalParameters["CurrentISA"] = (9,0,6)
       printWarning("Failed to detect ISA so forcing (gfx906) on windows")
-  isasWithDisabledHWMonitor = ((9,4,1), (9,4,2), (11,0,0), (11,0,1), (11,0,2), (12,0,0), (12,0,1))
+  isasWithDisabledHWMonitor = ((9,4,1), (9,4,2), (11,0,0), (11,0,1), (11,0,2), (11,5,0), (11,5,1), (12,0,0), (12,0,1))
   if globalParameters["CurrentISA"] in isasWithDisabledHWMonitor:
     isaString = ', '.join(map(gfxName, isasWithDisabledHWMonitor))
     printWarning(f"HardwareMonitor currently disabled for {isaString}")

Tensile/Source/CMakeLists.txt

@@ -51,9 +51,9 @@ if(CMAKE_CXX_COMPILER MATCHES ".*/hipcc$" OR CMAKE_CXX_COMPILER MATCHES ".*clang
 endif()
 
 if(CMAKE_CXX_COMPILER STREQUAL "hipcc")
-  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")
+  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906:xnack- gfx908:xnack- gfx90a:xnack- gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 gfx1150 gfx1151 CACHE STRING "GPU architectures")
 else()
-  set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906 gfx908 gfx90a gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 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 gfx1150 gfx1151 CACHE STRING "GPU architectures")
 endif()
 
 include(CMakeDependentOption)

Tensile/Source/lib/include/Tensile/AMDGPU.hpp

@@ -73,7 +73,9 @@ namespace Tensile
             gfx1035 = 1035,
             gfx1100 = 1100,
             gfx1101 = 1101,
-            gfx1102 = 1102
+            gfx1102 = 1102,
+            gfx1150 = 1150,
+            gfx1151 = 1151
         };
 
         static std::string toString(Processor p)
@@ -118,6 +120,10 @@ namespace Tensile
                 return "gfx1101";
             case AMDGPU::Processor::gfx1102:
                 return "gfx1102";
+            case AMDGPU::Processor::gfx1150:
+                return "gfx1150";
+            case AMDGPU::Processor::gfx1151:
+                return "gfx1151";
             }
             return "";
         }
@@ -184,6 +190,14 @@ namespace Tensile
             {
                 return AMDGPU::Processor::gfx1102;
             }
+            else if(deviceString.find("gfx1150") != std::string::npos)
+            {
+                return AMDGPU::Processor::gfx1150;
+            }
+            else if(deviceString.find("gfx1151") != std::string::npos)
+            {
+                return AMDGPU::Processor::gfx1151;
+            }
             else
             {
                 return static_cast<AMDGPU::Processor>(0);

Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp

@@ -58,6 +58,8 @@ namespace Tensile
         gfx1100,
         gfx1101,
         gfx1102,
+        gfx1150,
+        gfx1151,
         All
     };
 
@@ -106,6 +108,10 @@ namespace Tensile
             return "TensileLibrary_*_gfx1101";
         case LazyLoadingInit::gfx1102:
             return "TensileLibrary_*_gfx1102";
+        case LazyLoadingInit::gfx1150:
+            return "TensileLibrary_*_gfx1150";
+        case LazyLoadingInit::gfx1151:
+            return "TensileLibrary_*_gfx1151";
         case LazyLoadingInit::None:
             return "";
         }

Tensile/Source/lib/include/Tensile/Serialization/Predicates.hpp

@@ -232,6 +232,8 @@ namespace Tensile
                 iot::enumCase(io, value, "gfx1100", AMDGPU::Processor::gfx1100);
                 iot::enumCase(io, value, "gfx1101", AMDGPU::Processor::gfx1101);
                 iot::enumCase(io, value, "gfx1102", AMDGPU::Processor::gfx1102);
+                iot::enumCase(io, value, "gfx1150", AMDGPU::Processor::gfx1150);
+                iot::enumCase(io, value, "gfx1151", AMDGPU::Processor::gfx1151);
             }
         };
 

Tensile/Tests/disabled/direct_to_lds/dtl_dgemm.yaml

@@ -1,5 +1,5 @@
 TestParameters:
-  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1151] # not supported by arch
+  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1150, skip-gfx1151] # not supported by arch
   # Failing on latest ROCm build, re-enable when passing
 
 GlobalParameters:

Tensile/Tests/disabled/direct_to_lds/dtl_dgemm_lite.yaml

@@ -1,5 +1,5 @@
 TestParameters:
-  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1151] # not supported by arch
+  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1150, skip-gfx1151] # not supported by arch
   # Failing on latest ROCm build, re-enable when passing
 
 GlobalParameters:

Tensile/Tests/disabled/direct_to_lds/dtl_tsgr_dgemm.yaml

@@ -1,5 +1,5 @@
 TestParameters:
-  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1151] # not supported by arch
+  marks: [xfail, skip-gfx900, skip-gfx906, skip-gfx908, skip-gfx1010, skip-gfx1011, skip-gfx1012, skip-gfx1030, skip-gfx1100, skip-gfx1101, skip-gfx1102, skip-gfx1150, skip-gfx1151] # not supported by arch
   # Failing on latest ROCm build, re-enable when passing
 
 GlobalParameters: