From 6d6d64c3065b3485ff764c4e368849ba0d41d0e4 Mon Sep 17 00:00:00 2001 From: Tom Rix Date: Sun, 27 Jul 2025 13:06:25 -0700 Subject: [PATCH] tensile gfx1153 cut-n-paste from gfx1152 Signed-off-by: Tom Rix --- Tensile/AsmCaps.py | 44 +++++++++++++++++++ Tensile/Common.py | 6 +-- Tensile/Source/CMakeLists.txt | 4 +- Tensile/Source/lib/include/Tensile/AMDGPU.hpp | 7 +++ .../include/Tensile/PlaceholderLibrary.hpp | 3 ++ 5 files changed, 59 insertions(+), 5 deletions(-) diff --git a/Tensile/AsmCaps.py b/Tensile/AsmCaps.py index ea0518752ac4..83514e6d8f03 100644 --- a/Tensile/AsmCaps.py +++ b/Tensile/AsmCaps.py @@ -1035,6 +1035,50 @@ CACHED_ASM_CAPS = \ 'v_mov_b64': False, 'v_pk_fma_f16': True, 'v_pk_fmac_f16': False}, + (11, 5, 3): {'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}, (12, 0, 0): {'HasAddLshl': True, 'HasAtomicAdd': False, 'HasDirectToLdsDest': False, diff --git a/Tensile/Common.py b/Tensile/Common.py index 107dcb272c61..f66e7a274953 100644 --- a/Tensile/Common.py +++ b/Tensile/Common.py @@ -247,7 +247,7 @@ globalParameters["SupportedISA"] = [(8,0,3), (9,4,0), (9,4,1), (9,4,2), (9,5,0), (10,1,0), (10,1,1), (10,1,2), (10,3,0), (10,3,1), (10,3,5), (11,0,0), (11,0,1), (11,0,2), (11,0,3), - (11,5,0), (11,5,1), (11,5,2), + (11,5,0), (11,5,1), (11,5,2), (11,5,3), (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 @@ -325,7 +325,7 @@ architectureMap = { 'gfx1010':'navi10', 'gfx1011':'navi12', 'gfx1012':'navi14', 'gfx1030':'navi21', 'gfx1031':'navi22', 'gfx1032':'navi23', 'gfx1034':'navi24', 'gfx1035':'rembrandt', 'gfx1100':'navi31', 'gfx1101':'navi32', 'gfx1102':'navi33', - 'gfx1103':'gfx1103', 'gfx1150':'gfx1150', 'gfx1151':'gfx1151', 'gfx1152':'gfx1152', + 'gfx1103':'gfx1103', 'gfx1150':'gfx1150', 'gfx1151':'gfx1151', 'gfx1152':'gfx1152', 'gfx1153':'gfx1153', 'gfx1200':'gfx1200', 'gfx1201':'gfx1201' } @@ -2467,7 +2467,7 @@ 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), (9,5,0), (11,0,0), (11,0,1), (11,0,2), (11,0,3), (11,5,0), (11,5,1), (11,5,2), (12,0,0), (12,0,1)) + isasWithDisabledHWMonitor = ((9,4,1), (9,4,2), (9,5,0), (11,0,0), (11,0,1), (11,0,2), (11,0,3), (11,5,0), (11,5,1), (11,5,2), (11,5,3), (12,0,0), (12,0,1)) if globalParameters["CurrentISA"] in isasWithDisabledHWMonitor: isaString = ', '.join(map(gfxName, isasWithDisabledHWMonitor)) printWarning(f"HardwareMonitor currently disabled for {isaString}") diff --git a/Tensile/Source/CMakeLists.txt b/Tensile/Source/CMakeLists.txt index 7f10ee319518..1002b29bb981 100644 --- a/Tensile/Source/CMakeLists.txt +++ b/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 gfx1103 gfx1150 gfx1151 gfx1152 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 gfx1103 gfx1150 gfx1151 gfx1152 gfx1153 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 gfx1103 gfx1150 gfx1151 gfx1152 CACHE STRING "GPU architectures") + set(TENSILE_GPU_ARCHS gfx803 gfx900 gfx906 gfx908 gfx90a gfx1010 gfx1011 gfx1012 gfx1030 gfx1031 gfx1032 gfx1034 gfx1035 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1153 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 index e65a4831e082..5e0a43c0fdab 100644 --- a/Tensile/Source/lib/include/Tensile/AMDGPU.hpp +++ b/Tensile/Source/lib/include/Tensile/AMDGPU.hpp @@ -79,6 +79,7 @@ namespace Tensile gfx1150 = 1150, gfx1151 = 1151, gfx1152 = 1152, + gfx1153 = 1153, }; static std::string toString(Processor p) @@ -133,6 +134,8 @@ namespace Tensile return "gfx1151"; case AMDGPU::Processor::gfx1152: return "gfx1152"; + case AMDGPU::Processor::gfx1153: + return "gfx1153"; } return ""; } @@ -219,6 +222,10 @@ namespace Tensile { return AMDGPU::Processor::gfx1152; } + else if(deviceString.find("gfx1153") != std::string::npos) + { + return AMDGPU::Processor::gfx1153; + } else { return static_cast(0); diff --git a/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp b/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp index ba9719f77bb2..3d1d43e0c9d0 100644 --- a/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp +++ b/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp @@ -63,6 +63,7 @@ namespace Tensile gfx1151, gfx1150, gfx1152, + gfx1153, All }; @@ -121,6 +122,8 @@ namespace Tensile return "TensileLibrary_*_gfx1151"; case LazyLoadingInit::gfx1152: return "TensileLibrary_*_gfx1152"; + case LazyLoadingInit::gfx1153: + return "TensileLibrary_*_gfx1153"; case LazyLoadingInit::None: return ""; } -- 2.50.1