From 4b03428cb375182ea6bd05f40ea7f38a6c1f873c Mon Sep 17 00:00:00 2001 From: Tom Rix Date: Fri, 18 Apr 2025 07:30:37 -0700 Subject: [PATCH] tensile fedora gpus --- Tensile/AsmCaps.py | 176 ++++++++++++++++++ Tensile/Common.py | 10 +- Tensile/Source/CMakeLists.txt | 4 +- Tensile/Source/lib/include/Tensile/AMDGPU.hpp | 36 +++- .../include/Tensile/PlaceholderLibrary.hpp | 17 +- 5 files changed, 229 insertions(+), 14 deletions(-) diff --git a/Tensile/AsmCaps.py b/Tensile/AsmCaps.py index 548b31f28321..78ffa73bd81a 100644 --- a/Tensile/AsmCaps.py +++ b/Tensile/AsmCaps.py @@ -639,6 +639,50 @@ CACHED_ASM_CAPS = \ 'v_mov_b64': False, 'v_pk_fma_f16': True, 'v_pk_fmac_f16': False}, + (10, 3, 5): {'HasAddLshl': True, + 'HasAtomicAdd': False, + 'HasDirectToLdsDest': False, + 'HasDirectToLdsNoDest': True, + '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': False, + 'KernargPreloading': False, + 'MaxLgkmcnt': 15, + 'MaxVmcnt': 63, + 'SupportedISA': True, + 'SupportedSource': True, + 'VOP3v_dot4_i32_i8': True, + 'v_dot2_f32_f16': True, + 'v_dot2c_f32_f16': True, + 'v_dot4_i32_i8': False, + 'v_dot4c_i32_i8': True, + '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, 0, 0): {'HasAddLshl': True, 'HasAtomicAdd': True, 'HasDirectToLdsDest': False, @@ -771,6 +815,94 @@ CACHED_ASM_CAPS = \ 'v_mov_b64': False, 'v_pk_fma_f16': True, 'v_pk_fmac_f16': False}, + (11, 0, 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}, + (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, @@ -815,6 +947,50 @@ CACHED_ASM_CAPS = \ 'v_mov_b64': False, 'v_pk_fma_f16': True, 'v_pk_fmac_f16': False}, + (11, 5, 2): {'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 410c83656fd7..4d212d977c3d 100644 --- a/Tensile/Common.py +++ b/Tensile/Common.py @@ -245,9 +245,9 @@ globalParameters["MaxFileName"] = 64 # If a file name would be long globalParameters["SupportedISA"] = [(8,0,3), (9,0,0), (9,0,6), (9,0,8), (9,0,10), (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), + (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), (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 @@ architectureMap = { 'gfx1010':'navi10', 'gfx1011':'navi12', 'gfx1012':'navi14', 'gfx1030':'navi21', 'gfx1031':'navi22', 'gfx1032':'navi23', 'gfx1034':'navi24', 'gfx1035':'rembrandt', 'gfx1100':'navi31', 'gfx1101':'navi32', 'gfx1102':'navi33', - 'gfx1151':'gfx1151', + 'gfx1103':'gfx1103', 'gfx1150':'gfx1150', 'gfx1151':'gfx1151', 'gfx1152':'gfx1152', '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), (11,0,3), (11,5,0), (11,5,1), (11,5,2), (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 e02b209a262a..7f10ee319518 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 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 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 gfx1103 gfx1150 gfx1151 gfx1152 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 0ab8ced5cf5d..2317ce79f8f2 100644 --- a/Tensile/Source/lib/include/Tensile/AMDGPU.hpp +++ b/Tensile/Source/lib/include/Tensile/AMDGPU.hpp @@ -73,7 +73,11 @@ namespace Tensile gfx1035 = 1035, gfx1100 = 1100, gfx1101 = 1101, - gfx1102 = 1102 + gfx1102 = 1102, + gfx1103 = 1103, + gfx1150 = 1150, + gfx1151 = 1151, + gfx1152 = 1152, }; static std::string toString(Processor p) @@ -118,9 +122,17 @@ namespace Tensile return "gfx1101"; case AMDGPU::Processor::gfx1102: return "gfx1102"; - } - return ""; - } + case AMDGPU::Processor::gfx1103: + return "gfx1103"; + case AMDGPU::Processor::gfx1150: + return "gfx1150"; + case AMDGPU::Processor::gfx1151: + return "gfx1151"; + case AMDGPU::Processor::gfx1152: + return "gfx1152"; + } + return ""; + } AMDGPU::Processor toProcessorId(std::string const& deviceString) { @@ -184,6 +196,22 @@ namespace Tensile { return AMDGPU::Processor::gfx1102; } + else if(deviceString.find("gfx1103") != std::string::npos) + { + return AMDGPU::Processor::gfx1103; + } + 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 if(deviceString.find("gfx1152") != std::string::npos) + { + return AMDGPU::Processor::gfx1152; + } else { return static_cast(0); diff --git a/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp b/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp index 10898ec2d1d6..f838f15d3ac4 100644 --- a/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp +++ b/Tensile/Source/lib/include/Tensile/PlaceholderLibrary.hpp @@ -58,6 +58,10 @@ namespace Tensile gfx1100, gfx1101, gfx1102, + gfx1103, + gfx1151, + gfx1150, + gfx1152, All }; @@ -106,10 +110,17 @@ namespace Tensile return "TensileLibrary_*_gfx1101"; case LazyLoadingInit::gfx1102: return "TensileLibrary_*_gfx1102"; - case LazyLoadingInit::None: - return ""; + case LazyLoadingInit::gfx1103: + return "TensileLibrary_*_gfx1103"; + case LazyLoadingInit::gfx1150: + return "TensileLibrary_*_gfx1150"; + case LazyLoadingInit::gfx1151: + return "TensileLibrary_*_gfx1151"; + case LazyLoadingInit::gfx1152: + return "TensileLibrary_*_gfx1152"; + case LazyLoadingInit::None: + return ""; } - return ""; } template -- 2.48.1