amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-20 07:38:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Finer control of code path options (#67)

Browse Source 
Add macros to allow specific code options to be enabled or disabled,
controlled by options to configure and cmake. This expands on the
existing GEMM and/or TRSM functionality to enable/disable SUP handling
and replaces the hard coded #define in include files to enable small matrix
paths.

All options are enabled by default for all BLIS sub-configs but many of them
are currently only implemented in AMD specific framework code variants.

AMD-Internal: [CPUPL-6906]
---------

Co-authored-by: Varaganti, Kiran <Kiran.Varaganti@amd.com>
This commit is contained in:
Smyth, Edward
2025-07-08 10:59:23 +01:00
committed by GitHub
parent 9b02201b5b
commit 969ceb7413
18 changed files with 304 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
CMakeLists.txt
View File

@@ -277,7 +277,12 @@ option(ENABLE_BLAS "BLAS compatiblity layer" ON)
option(ENABLE_CBLAS "CBLAS compatiblity layer" OFF)
option(ENABLE_MIXED_DT "Mixed datatype support" ON)
option(ENABLE_MIXED_DT_EXTRA_MEM "Mixed datatype optimization requiring extra memory" ON)
option(ENABLE_SUP_HANDLING "Small matrix handling" ON)
option(ENABLE_MNK1_MATRIX "M, N or K = 1 matrix handling" ON)
option(ENABLE_TINY_MATRIX "Tiny matrix handling" ON)
option(ENABLE_SMALL_MATRIX "Small matrix handling" ON)
option(ENABLE_SUP_HANDLING "SUP matrix handling" ON)
option(ENABLE_SMALL_MATRIX_TRSM "TRSM Small matrix handling" ON)
option(ENABLE_TRSM_PREINVERSION "Enable TRSM preinversion" ON)
if(WIN32)
set(ENABLE_MEMKIND "no" CACHE STRING "libmemkind for manage memory pools")
set_property(CACHE ENABLE_MEMKIND PROPERTY STRINGS "no")
@@ -292,7 +297,6 @@ else()
during CMake invokation: auto, yes, no")
endif()
endif()
option(ENABLE_TRSM_PREINVERSION "Enable TRSM preinversion" ON)
option(ENABLE_AOCL_DYNAMIC "Dynamic selection of number of threads" ON)
set(FORCE_VERSION "no" CACHE STRING "Force configure to use an arbitrary version string")
if(WIN32)
@@ -608,14 +612,46 @@ else()
set(ENABLE_MIXED_DT_EXTRA_MEM_01 0)
set(ENABLE_MIXED_DT_01 0)
endif()
cmake_print_variables(ENABLE_SUP_HANDLING)
if(ENABLE_SUP_HANDLING)
cmake_print_variables(ENABLE_MNK1_MATRIX)
if(ENABLE_MNK1_MATRIX)
message(" M, N or K = 1 matrix handling is enabled.")
set(ENABLE_MNK1_MATRIX_01 1)
else()
message(" M, N or K = 1 matrix handling is disabled.")
set(ENABLE_MNK1_MATRIX_01 0)
endif()
cmake_print_variables(ENABLE_TINY_MATRIX)
if(ENABLE_TINY_MATRIX)
message(" Tiny matrix handling is enabled.")
set(ENABLE_TINY_MATRIX_01 1)
else()
message(" Tiny matrix handling is disabled.")
set(ENABLE_TINY_MATRIX_01 0)
endif()
cmake_print_variables(ENABLE_SMALL_MATRIX)
if(ENABLE_SMALL_MATRIX)
message(" Small matrix handling is enabled.")
set(ENABLE_SUP_HANDLING_01 1)
set(ENABLE_SMALL_MATRIX_01 1)
else()
message(" Small matrix handling is disabled.")
set(ENABLE_SMALL_MATRIX_01 0)
endif()
cmake_print_variables(ENABLE_SUP_HANDLING)
if(ENABLE_SUP_HANDLING)
message(" SUP matrix handling is enabled.")
set(ENABLE_SUP_HANDLING_01 1)
else()
message(" SUP matrix handling is disabled.")
set(ENABLE_SUP_HANDLING_01 0)
endif()
cmake_print_variables(ENABLE_SMALL_MATRIX_TRSM)
if(ENABLE_SMALL_MATRIX)
message(" TRSM Small matrix handling is enabled.")
set(ENABLE_SMALL_MATRIX_TRSM_01 1)
else()
message(" TRSM Small matrix handling is disabled.")
set(ENABLE_SMALL_MATRIX_TRSM_01 0)
endif()
cmake_print_variables(ENABLE_TRSM_PREINVERSION)
if(ENABLE_TRSM_PREINVERSION)
message(" trsm diagonal element pre-inversion is enabled.")

26
build/bli_config.h.in
View File

@@ -5,7 +5,7 @@
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018 - 2023, Advanced Micro Devices, Inc. All rights reserved.
Copyright (C) 2018 - 2025, Advanced Micro Devices, Inc. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -162,12 +162,36 @@
#endif
#endif
#if @enable_mnk1_matrix@
#define BLIS_ENABLE_MNK1_MATRIX
#else
#define BLIS_DISABLE_MNK1_MATRIX
#endif
#if @enable_tiny_matrix@
#define BLIS_ENABLE_TINY_MATRIX
#else
#define BLIS_DISABLE_TINY_MATRIX
#endif
#if @enable_small_matrix@
#define BLIS_ENABLE_SMALL_MATRIX
#else
#define BLIS_DISABLE_SMALL_MATRIX
#endif
#if @enable_sup_handling@
#define BLIS_ENABLE_SUP_HANDLING
#else
#define BLIS_DISABLE_SUP_HANDLING
#endif
#if @enable_small_matrix_trsm@
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
#else
#define BLIS_DISABLE_SMALL_MATRIX_TRSM
#endif
#if @enable_memkind@
#define BLIS_ENABLE_MEMKIND
#else

26
build/cmake/bli_config.h.in
View File

@@ -4,7 +4,7 @@
An object-based framework for developing high-performance BLAS-like
libraries.
Copyright (C) 2020 - 2024, Advanced Micro Devices, Inc. All rights reserved.
Copyright (C) 2020 - 2025, Advanced Micro Devices, Inc. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -160,12 +160,36 @@ ${KERNEL_LIST_DEFINES}
#endif
#endif
#if ${ENABLE_MNK1_MATRIX_01}
#define BLIS_ENABLE_MNK1_MATRIX
#else
#define BLIS_DISABLE_MNK1_MATRIX
#endif
#if ${ENABLE_TINY_MATRIX_01}
#define BLIS_ENABLE_TINY_MATRIX
#else
#define BLIS_DISABLE_TINY_MATRIX
#endif
#if ${ENABLE_SMALL_MATRIX_01}
#define BLIS_ENABLE_SMALL_MATRIX
#else
#define BLIS_DISABLE_SMALL_MATRIX
#endif
#if ${ENABLE_SUP_HANDLING_01}
#define BLIS_ENABLE_SUP_HANDLING
#else
#define BLIS_DISABLE_SUP_HANDLING
#endif
#if ${ENABLE_SMALL_MATRIX_TRSM_01}
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
#else
#define BLIS_DISABLE_SMALL_MATRIX_TRSM
#endif
#if ${ENABLE_MEMKIND_01}
#define BLIS_ENABLE_MEMKIND
#else

3
config/amdzen/bli_family_amdzen.h
View File

@@ -45,9 +45,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/x86_64/bli_family_x86_64.h
View File

@@ -43,9 +43,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/zen/bli_family_zen.h
View File

@@ -44,9 +44,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/zen2/bli_family_zen2.h
View File

@@ -42,9 +42,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/zen3/bli_family_zen3.h
View File

@@ -42,9 +42,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/zen4/bli_family_zen4.h
View File

@@ -43,9 +43,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

3
config/zen5/bli_family_zen5.h
View File

@@ -43,9 +43,6 @@
#define BLIS_THREAD_MAX_IR 1
#define BLIS_THREAD_MAX_JR 1
#define BLIS_ENABLE_SMALL_MATRIX
#define BLIS_ENABLE_SMALL_MATRIX_TRSM
// This will select the threshold below which small matrix code will be called.
#define BLIS_SMALL_MATRIX_THRES 700
#define BLIS_SMALL_M_RECT_MATRIX_THRES 160

117
configure vendored
View File

@@ -260,15 +260,58 @@ print_usage()
echo " only be enabled when mixed domain/precision support is"
echo " enabled."
echo " "
echo " --disable-sup-handling, --enable-sup-handling"
echo " --disable-mnk1-matrix, --enable-mnk1-matrix"
echo " "
echo " Disable (enabled by default) handling of matrix problem"
echo " where M, N or K = 1 via separate code branches. When disabled,"
echo " these operations will be performed by gemm rather than gemv"
echo " or other optimized implementations."
echo " "
echo " --disable-tiny-matrix, --enable-tiny-matrix"
echo " "
echo " Disable (enabled by default) handling of tiny"
echo " matrix problems via tiny code branches. When disabled,"
echo " these tiny level-3 operations will be performed by"
echo " the conventional implementation, which is optimized for"
echo " medium and large problems. Note that what qualifies as"
echo " \"tiny\" depends on thresholds that may vary by sub-"
echo " configuration."
echo " Currently only of relevance on configs that include"
echo " AMD Zen sub-configs"
echo " "
echo " --disable-small-matrix, --enable-small-matrix"
echo " "
echo " Disable (enabled by default) handling of small/skinny"
echo " matrix problems via separate code branches. When disabled,"
echo " matrix problems via small code branches. When disabled,"
echo " these small/skinny level-3 operations will be performed by"
echo " the conventional implementation, which is optimized for"
echo " medium and large problems. Note that what qualifies as"
echo " \"small\" depends on thresholds that may vary by sub-"
echo " configuration."
echo " Currently only of relevance on configs that include"
echo " AMD Zen sub-configs"
echo " "
echo " --disable-sup-handling, --enable-sup-handling"
echo " "
echo " Disable (enabled by default) handling of small/skinny"
echo " matrix problems via SUP code branches. When disabled,"
echo " these small/skinny level-3 operations will be performed by"
echo " the conventional implementation, which is optimized for"
echo " medium and large problems. Note that what qualifies as"
echo " \"SUP\" depends on thresholds that may vary by sub-"
echo " configuration."
echo " "
echo " --disable-small-matrix-trsm, --enable-small-matrix-trsm"
echo " "
echo " Disable (enabled by default) handling of small/skinny"
echo " TRSM problems via small code branches. When disabled,"
echo " these small/skinny level-3 operations will be performed by"
echo " the conventional implementation, which is optimized for"
echo " medium and large problems. Note that what qualifies as"
echo " \"small\" depends on thresholds that may vary by sub-"
echo " configuration."
echo " Currently only of relevance on configs that include"
echo " AMD Zen sub-configs"
echo " "
echo " -a NAME --enable-addon=NAME"
echo " "
@@ -2108,9 +2151,15 @@ main()
enable_cblas='no'
enable_mixed_dt='yes'
enable_mixed_dt_extra_mem='yes'
enable_mnk1_matrix='yes'
enable_tiny_matrix='yes'
enable_small_matrix='yes'
enable_sup_handling='yes'
enable_memkind='' # The default memkind value is determined later on.
enable_small_matrix_trsm='yes'
enable_trsm_preinversion='yes'
enable_memkind='' # The default memkind value is determined later on.
enable_aocl_dynamic='yes'
force_version='no'
complex_return='default'
@@ -2319,12 +2368,36 @@ main()
disable-mixed-dt-extra-mem)
enable_mixed_dt_extra_mem='no'
;;
enable-mnk1-matrix)
enable_mnk1_matrix='yes'
;;
disable-mnk1-matrix)
enable_mnk1_matrix='no'
;;
enable-tiny-matrix)
enable_tiny_matrix='yes'
;;
disable-tiny-matrix)
enable_tiny_matrix='no'
;;
enable-small-matrix)
enable_small_matrix='yes'
;;
disable-small-matrix)
enable_small_matrix='no'
;;
enable-sup-handling)
enable_sup_handling='yes'
;;
disable-sup-handling)
enable_sup_handling='no'
;;
enable-small-matrix-trsm)
enable_small_matrix_trsm='yes'
;;
disable-small-matrix-trsm)
enable_small_matrix_trsm='no'
;;
with-memkind)
enable_memkind='yes'
;;
@@ -3213,13 +3286,41 @@ main()
enable_mixed_dt_extra_mem_01=0
enable_mixed_dt_01=0
fi
if [ "x${enable_sup_handling}" = "xyes" ]; then
if [ "x${enable_mnk1_matrix}" = "xyes" ]; then
echo "${script_name}: M,N,K=1 matrix handling is enabled."
enable_mnk1_matrix_01=1
else
echo "${script_name}: M,N,K=1 matrix handling is disabled."
enable_mnk1_matrix_01=0
fi
if [ "x${enable_tiny_matrix}" = "xyes" ]; then
echo "${script_name}: tiny matrix handling is enabled."
enable_tiny_matrix_01=1
else
echo "${script_name}: tiny matrix handling is disabled."
enable_tiny_matrix_01=0
fi
if [ "x${enable_small_matrix}" = "xyes" ]; then
echo "${script_name}: small matrix handling is enabled."
enable_sup_handling_01=1
enable_small_matrix_01=1
else
echo "${script_name}: small matrix handling is disabled."
enable_small_matrix_01=0
fi
if [ "x${enable_sup_handling}" = "xyes" ]; then
echo "${script_name}: SUP matrix handling is enabled."
enable_sup_handling_01=1
else
echo "${script_name}: SUP matrix handling is disabled."
enable_sup_handling_01=0
fi
if [ "x${enable_small_matrix_trsm}" = "xyes" ]; then
echo "${script_name}: TRSM small matrix handling is enabled."
enable_small_matrix_trsm_01=1
else
echo "${script_name}: TRSM small matrix handling is disabled."
enable_small_matrix_trsm_01=0
fi
if [ "x${enable_trsm_preinversion}" = "xyes" ]; then
echo "${script_name}: trsm diagonal element pre-inversion is enabled."
enable_trsm_preinversion_01=1
@@ -3586,9 +3687,13 @@ main()
| sed -e "s/@enable_cblas@/${enable_cblas_01}/g" \
| sed -e "s/@enable_mixed_dt@/${enable_mixed_dt_01}/g" \
| sed -e "s/@enable_mixed_dt_extra_mem@/${enable_mixed_dt_extra_mem_01}/g" \
| sed -e "s/@enable_mnk1_matrix@/${enable_mnk1_matrix_01}/g" \
| sed -e "s/@enable_tiny_matrix@/${enable_tiny_matrix_01}/g" \
| sed -e "s/@enable_small_matrix@/${enable_small_matrix_01}/g" \
| sed -e "s/@enable_sup_handling@/${enable_sup_handling_01}/g" \
| sed -e "s/@enable_memkind@/${enable_memkind_01}/g" \
| sed -e "s/@enable_small_matrix_trsm@/${enable_small_matrix_trsm_01}/g" \
| sed -e "s/@enable_trsm_preinversion@/${enable_trsm_preinversion_01}/g" \
| sed -e "s/@enable_memkind@/${enable_memkind_01}/g" \
| sed -e "s/@enable_aocl_dynamic@/${enable_aocl_dynamic_01}/g" \
| sed -e "s/@enable_pragma_omp_simd@/${enable_pragma_omp_simd_01}/g" \
| sed -e "s/@enable_sandbox@/${enable_sandbox_01}/g" \

2
frame/3/gemm/bli_tiny_gemm.h
View File

@@ -53,4 +53,4 @@ err_t PASTEMAC( ch, tfuncname ) \
); \
GENTFUNC( scomplex, c, gemm_tiny )
GENTFUNC( dcomplex, z, gemm_tiny )
GENTFUNC( dcomplex, z, gemm_tiny )

2
frame/compat/bla_gemm.c
View File

@@ -285,6 +285,7 @@ void PASTEF77S(ch,blasname) \
const inc_t rs_c = 1; \
const inc_t cs_c = *ldc; \
\
IF_BLIS_ENABLE_MNK1_MATRIX(\
if( n0 == 1 ) \
{ \
if(bli_is_notrans(blis_transa)) \
@@ -357,6 +358,7 @@ void PASTEF77S(ch,blasname) \
bli_finalize_auto(); \
return; \
} \
) /* End of IF_BLIS_ENABLE_MNK1_MATRIX */ \
\
const num_t dt = PASTEMAC(ch,type); \
\

52
frame/compat/bla_gemm_amd.c
View File

@@ -304,6 +304,7 @@ void PASTEF77S(ch,blasname) \
const inc_t rs_c = 1; \
const inc_t cs_c = *ldc; \
\
IF_BLIS_ENABLE_MNK1_MATRIX(\
if( n0 == 1 ) \
{ \
if(bli_is_notrans(blis_transa)) \
@@ -380,6 +381,7 @@ void PASTEF77S(ch,blasname) \
bli_finalize_auto(); \
return; \
} \
) /* End of IF_BLIS_ENABLE_MNK1_MATRIX */ \
\
const num_t dt = PASTEMAC(ch,type); \
\
@@ -534,6 +536,8 @@ void dgemm_blis_impl
const inc_t rs_c = 1;
const inc_t cs_c = *ldc;
#ifdef BLIS_ENABLE_MNK1_MATRIX
/* Call GEMV when m == 1 or n == 1 with the context set
to an uninitialized void pointer i.e. ((void *)0)*/
if (n0 == 1)
@@ -615,6 +619,8 @@ void dgemm_blis_impl
return;
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
// This function is invoked on all architectures including 'generic'.
// Non-AVX2+FMA3 platforms will use the kernels derived from the context.
if (bli_cpuid_is_avx2fma3_supported() == FALSE)
@@ -665,6 +671,8 @@ void dgemm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
/*
Invoking the API for input sizes with k = 1.
- The API is single-threaded.
@@ -715,6 +723,9 @@ void dgemm_blis_impl
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
#ifdef BLIS_ENABLE_TINY_MATRIX
/**
*Early check for tiny sizes.
*if inputs are in range of tiny gemm kernel,
@@ -746,6 +757,7 @@ void dgemm_blis_impl
bli_finalize_auto();
return;
}
#endif // End of BLIS_ENABLE_TINY_MATRIX
const num_t dt = BLIS_DOUBLE;
@@ -859,8 +871,9 @@ void dgemm_blis_impl
}
}
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX
#ifdef BLIS_ENABLE_SUP_HANDLING
err_t sup_status = BLIS_FAILURE;
sup_status = bli_gemmsup(&alphao, &ao, &bo, &betao, &co, NULL, NULL);
if ( sup_status == BLIS_SUCCESS )
@@ -871,6 +884,7 @@ void dgemm_blis_impl
bli_finalize_auto();
return;
}
#endif // End of BLIS_ENABLE_SUP_HANDLING
// fall back on native path when dgemm is not handled in sup path.
//bli_gemmnat(&alphao, &ao, &bo, &betao, &co, NULL, NULL);
@@ -1012,6 +1026,8 @@ void zgemm_blis_impl
const inc_t rs_c = 1;
const inc_t cs_c = *ldc;
#ifdef BLIS_ENABLE_MNK1_MATRIX
/* Call GEMV when m == 1 or n == 1 with the context set
to an uninitialized void pointer i.e. ((void *)0)*/
if (n0 == 1)
@@ -1093,6 +1109,8 @@ void zgemm_blis_impl
return;
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
// This function is invoked on all architectures including 'generic'.
// Non-AVX2+FMA3 platforms will use the kernels derived from the context.
if (bli_cpuid_is_avx2fma3_supported() == FALSE)
@@ -1143,6 +1161,8 @@ void zgemm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
/*
Invoking the API for input sizes with k = 1.
- The API is single-threaded.
@@ -1240,6 +1260,11 @@ void zgemm_blis_impl
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
#ifdef BLIS_ENABLE_TINY_MATRIX
// May also be used in small path below
bool is_parallel = bli_thread_get_is_parallel(); // Check if parallel zgemm is invoked.
// Tiny gemm dispatch
@@ -1271,6 +1296,8 @@ void zgemm_blis_impl
}
#endif
#endif // End of BLIS_ENABLE_TINY_MATRIX
const num_t dt = BLIS_DCOMPLEX;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -1297,6 +1324,11 @@ void zgemm_blis_impl
#ifdef BLIS_ENABLE_SMALL_MATRIX
/* Check if we have already defined this above */
#ifndef BLIS_ENABLE_TINY_MATRIX
bool is_parallel = bli_thread_get_is_parallel(); // Check if parallel zgemm is invoked.
#endif
/* Query the architecture ID */
arch_t arch_id = bli_arch_query_id();
@@ -1389,8 +1421,9 @@ void zgemm_blis_impl
}
}
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX
#ifdef BLIS_ENABLE_SUP_HANDLING
err_t sup_status = BLIS_FAILURE;
sup_status = bli_gemmsup(&alphao, &ao, &bo, &betao, &co, NULL, NULL);
if ( sup_status == BLIS_SUCCESS )
@@ -1401,6 +1434,7 @@ void zgemm_blis_impl
bli_finalize_auto();
return;
}
#endif // End of BLIS_ENABLE_SUP_HANDLING
// fall back on native path when zgemm is not handled in sup path.
//bli_gemmnat(&alphao, &ao, &bo, &betao, &co, NULL, NULL);
@@ -1555,6 +1589,8 @@ void cgemm_blis_impl
const inc_t rs_c = 1;
const inc_t cs_c = *ldc;
#ifdef BLIS_ENABLE_MNK1_MATRIX
/* Call GEMV when m == 1 or n == 1 with the context set
to an uninitialized void pointer i.e. ((void *)0)*/
if (n0 == 1)
@@ -1636,6 +1672,8 @@ void cgemm_blis_impl
return;
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
// This function is invoked on all architectures including 'generic'.
// Non-AVX2+FMA3 platforms will use the kernels derived from the context.
if (bli_cpuid_is_avx2fma3_supported() == FALSE)
@@ -1686,6 +1724,8 @@ void cgemm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
/*
Invoking the API for input sizes with k = 1.
- The API is single-threaded.
@@ -1719,6 +1759,10 @@ void cgemm_blis_impl
}
#endif
#endif // End of BLIS_ENABLE_MNK1_MATRIX
#ifdef BLIS_ENABLE_TINY_MATRIX
bool is_parallel = bli_thread_get_is_parallel(); // Check if parallel cgemm is invoked.
// Tiny gemm dispatch
@@ -1750,6 +1794,8 @@ void cgemm_blis_impl
}
#endif
#endif // End of BLIS_ENABLE_TINY_MATRIX
const num_t dt = BLIS_SCOMPLEX;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -1774,6 +1820,7 @@ void cgemm_blis_impl
bli_obj_set_conjtrans( blis_transa, &ao );
bli_obj_set_conjtrans( blis_transb, &bo );
#ifdef BLIS_ENABLE_SUP_HANDLING
err_t sup_status = BLIS_FAILURE;
sup_status = bli_gemmsup(&alphao, &ao, &bo, &betao, &co, NULL, NULL);
if ( sup_status == BLIS_SUCCESS )
@@ -1784,6 +1831,7 @@ void cgemm_blis_impl
bli_finalize_auto();
return;
}
#endif // End of BLIS_ENABLE_SUP_HANDLING
// fall back on native path when cgemm is not handled in sup path.
//bli_gemmnat(&alphao, &ao, &bo, &betao, &co, NULL, NULL);

4
frame/compat/bla_gemv_amd.c
View File

@@ -317,6 +317,8 @@ void dgemv_blis_impl
return;
}
#ifdef BLIS_ENABLE_TINY_MATRIX
/**
* DGEMV Tiny Path
* If the matrix dimensions are within 8x8 then calculate the result
@@ -341,6 +343,8 @@ void dgemv_blis_impl
return;
}
#endif // End of BLIS_ENABLE_TINY_MATRIX
/* Call variants based on transpose value. */
if((bli_does_notrans(blis_transa) && bli_is_col_stored( rs_a, cs_a ))
|| (bli_does_trans(blis_transa) && bli_is_row_stored( rs_a, cs_a )))

4
frame/compat/bla_trsm.c
View File

@@ -5,7 +5,7 @@
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2019 - 2023, Advanced Micro Devices, Inc. All rights reserved.
Copyright (C) 2019 - 2025, Advanced Micro Devices, Inc. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -247,6 +247,7 @@ void PASTEF77S(ch,blasname) \
/* If Transpose(A) uplo = higher then uplo = lower */ \
/* ----------------------------------------------------------- */ \
\
IF_BLIS_ENABLE_MNK1_MATRIX(\
if( n0 == 1 ) \
{ \
if( blis_side == BLIS_LEFT ) \
@@ -375,6 +376,7 @@ void PASTEF77S(ch,blasname) \
return; \
} \
} \
) /* End of IF_BLIS_ENABLE_MNK1_MATRIX */ \
\
const struc_t struca = BLIS_TRIANGULAR; \
\

27
frame/compat/bla_trsm_amd.c
View File

@@ -308,6 +308,7 @@ void PASTEF77S(ch,blasname) \
/* If Transpose(A) uplo = higher then uplo = lower */ \
/* ----------------------------------------------------------- */ \
\
IF_BLIS_ENABLE_MNK1_MATRIX(\
if( n0 == 1 ) \
{ \
if( blis_side == BLIS_LEFT ) \
@@ -442,6 +443,7 @@ void PASTEF77S(ch,blasname) \
return; \
} \
} \
) /* End of IF_BLIS_ENABLE_MNK1_MATRIX */ \
\
const struc_t struca = BLIS_TRIANGULAR; \
\
@@ -588,6 +590,8 @@ void strsm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
if( n0 == 1 )
{
if( blis_side == BLIS_LEFT )
@@ -732,6 +736,9 @@ void strsm_blis_impl
return;
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
const struc_t struca = BLIS_TRIANGULAR;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -789,7 +796,7 @@ void strsm_blis_impl
}
} // bli_cpuid_is_avx2fma3_supported
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX_TRSM
//bli_trsmnat
//(
@@ -932,6 +939,8 @@ void dtrsm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
if( n0 == 1 )
{
if( blis_side == BLIS_LEFT )
@@ -1077,6 +1086,8 @@ void dtrsm_blis_impl
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
const struc_t struca = BLIS_TRIANGULAR;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -1287,7 +1298,7 @@ void dtrsm_blis_impl
}
} // bli_cpuid_is_avx2fma3_supported
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX
//bli_trsmnat
//(
@@ -1431,6 +1442,8 @@ void ztrsm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
if( n0 == 1 )
{
if( blis_side == BLIS_LEFT )
@@ -1635,6 +1648,8 @@ void ztrsm_blis_impl
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
const struc_t struca = BLIS_TRIANGULAR;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -1810,7 +1825,7 @@ void ztrsm_blis_impl
}
} // bli_cpuid_is_avx2fma3_supported
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX
//bli_trsmnat
//(
@@ -1954,6 +1969,8 @@ void ctrsm_blis_impl
return;
}
#ifdef BLIS_ENABLE_MNK1_MATRIX
if( n0 == 1 )
{
if( blis_side == BLIS_LEFT )
@@ -2158,6 +2175,8 @@ void ctrsm_blis_impl
}
}
#endif // End of BLIS_ENABLE_MNK1_MATRIX
const struc_t struca = BLIS_TRIANGULAR;
obj_t alphao = BLIS_OBJECT_INITIALIZER_1X1;
@@ -2215,7 +2234,7 @@ void ctrsm_blis_impl
}
} // bli_cpuid_is_avx2fma3_supported
#endif //#ifdef BLIS_ENABLE_SMALL_MATRIX
#endif // End of BLIS_ENABLE_SMALL_MATRIX
//bli_trsmnat
//(

20
frame/include/bli_config_macro_defs.h
View File

@@ -5,7 +5,7 @@
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2019 - 2023, Advanced Micro Devices, Inc. All rights reserved.
Copyright (C) 2019 - 2025, Advanced Micro Devices, Inc. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -269,3 +269,21 @@
#endif
// -- CODE PATH ENABLEMENT --------------------------------------------------
#ifdef BLIS_ENABLE_MNK1_MATRIX
#define IF_BLIS_ENABLE_MNK1_MATRIX(...) __VA_ARGS__
#else
#define IF_BLIS_ENABLE_MNK1_MATRIX(...)
#endif
#ifdef BLIS_ENABLE_TINY_MATRIX
#define IF_BLIS_ENABLE_TINY_MATRIX(...) __VA_ARGS__
#else
#define IF_BLIS_ENABLE_TINY_MATRIX(...)
#endif
#ifdef BLIS_ENABLE_SMALL_MATRIX
#define IF_BLIS_ENABLE_SMALL_MATRIX(...) __VA_ARGS__
#else
#define IF_BLIS_ENABLE_SMALL_MATRIX(...)
#endif