v4.4 update. (#2979)

This commit is contained in:
Junkai-Wu
2026-01-24 11:46:17 -05:00
committed by GitHub
parent 2fafefb7b9
commit 9fba3195f9
293 changed files with 46343 additions and 2995 deletions
+13
View File
@@ -123,6 +123,19 @@ function(cutlass_add_cutlass_library)
PRIVATE cutlass_library_internal_interface
)
# Use medium code model to handle large libraries (>2GB) on Linux x86_64
# Conservative approach: only apply on Linux with GCC/Clang/Intel compilers
# This addresses the "relocation truncated to fit" linker error when libraries exceed 2GB
if(CMAKE_SYSTEM_NAME STREQUAL "Linux" AND
CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64" AND
CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang|IntelLLVM")
message(STATUS "CUTLASS: Applying -mcmodel=medium for large library support (Linux x86_64, ${CMAKE_CXX_COMPILER_ID})")
target_compile_options(${__NAME}_objs PRIVATE
$<$<COMPILE_LANGUAGE:CXX>:-mcmodel=medium>
$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-mcmodel=medium>
)
endif()
if (CUTLASS_BUILD_MONO_LIBRARY AND __SUFFIX)
# If we're only building a single monolithic library then we
@@ -56,6 +56,10 @@ namespace library {
// 2. e3m2_e4m3_f32_f16_e5m2
// 3. e3m2_e4m3_f32_f16_f16
// 4. e3m2_e4m3_f32_f32_f32
// 5. e3m2_e5m2_f32_f16_e4m3
// 6. e3m2_e5m2_f32_f16_e5m2
// 7. e3m2_e5m2_f32_f16_f16
// 8. e3m2_e5m2_f32_f32_f32
void initialize_gemm_reference_operations_f6_f8_f32(Manifest &manifest) {
@@ -138,6 +142,86 @@ void initialize_gemm_reference_operations_f6_f8_f32(Manifest &manifest) {
float, // ElementAccumulator
float // ElementD
>(manifest);
// 5.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
half_t, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float_e4m3_t // ElementD
>(manifest);
// 6.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
half_t, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float_e5m2_t // ElementD
>(manifest);
// 7.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
half_t, // ElementC
float, // ElementScalar
float, // ElementAccumulator
half_t // ElementD
>(manifest);
// 8.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
float, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float // ElementD
>(manifest);
// 5.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
void, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float_e4m3_t // ElementD
>(manifest);
// 6.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
void, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float_e5m2_t // ElementD
>(manifest);
// 7.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
void, // ElementC
float, // ElementScalar
float, // ElementAccumulator
half_t // ElementD
>(manifest);
// 8.
make_gemm_real_canonical_layouts<
float_e3m2_t, // ElementA
float_e5m2_t, // ElementB
void, // ElementC
float, // ElementScalar
float, // ElementAccumulator
float // ElementD
>(manifest);
}
///////////////////////////////////////////////////////////////////////////////////////////////////