You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

param.h 81 kB

12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
12 years ago
5 years ago
12 years ago
12 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
5 years ago
12 years ago
12 years ago
12 years ago
5 years ago
5 years ago
12 years ago
5 years ago
5 years ago
12 years ago
5 years ago
12 years ago
5 years ago
5 years ago
5 years ago
12 years ago
6 years ago
12 years ago
12 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
12 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
12 years ago
12 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
12 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
Simplifying ARMv8 build parameters ARMv8 builds were a bit mixed up, with ThunderX2 code in ARMv8 mode (which is not right because TX2 is ARMv8.1) as well as requiring a few redundancies in the defines, making it harder to maintain and understand what core has what. A few other minor issues were also fixed. Tests were made on the following cores: A53, A57, A72, Falkor, ThunderX, ThunderX2, and XGene. Tests were: OpenBLAS/test, OpenBLAS/benchmark, BLAS-Tester. A summary: * Removed TX2 code from ARMv8 build, to make sure it is compatible with all ARMv8 cores, not just v8.1. Also, the TX2 code has actually harmed performance on big cores. * Commoned up ARMv8 architectures' defines in params.h, to make sure that all will benefit from ARMv8 settings, in addition to their own. * Adding a few more cores, using ARMv8's include strategy, to benefit from compiler optimisations using mtune. Also updated cache information from the manuals, making sure we set good conservative values by default. Removed Vulcan, as it's an alias to TX2. * Auto-detecting most of those cores, but also updating the forced compilation in getarch.c, to make sure the parameters are the same whether compiled natively or forced arch. Benefits: * ARMv8 build is now guaranteed to work on all ARMv8 cores * Improved performance for ARMv8 builds on some cores (A72, Falkor, ThunderX1 and 2: up to 11%) over current develop * Improved performance for *all* cores comparing to develop branch before TX2's patch (9% ~ 36%) * ThunderX1 builds are 14% faster than ARMv8 on TX1, 9% faster than current develop's branch and 8% faster than deveop before tx2 patches Issues: * Regression from current develop branch for A53 (-12%) and A57 (-3%) with ARMv8 builds, but still faster than before TX2's commit (+15% and +24% respectively). This can be improved with a simplification of TX2's code, to be done in future patches. At least the code is guaranteed to be ARMv8.0 now. Comments: * CortexA57 builds are unchanged on A57 hardware from develop's branch, which makes sense, as it's untouched. * CortexA72 builds improve over A57 on A72 hardware, even if they're using the same includes due to new compiler tunning in the makefile.
6 years ago
12 years ago
12 years ago
8 years ago
8 years ago
8 years ago
8 years ago
8 years ago
8 years ago
8 years ago
8 years ago
8 years ago
12 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462
  1. /*****************************************************************************
  2. Copyright (c) 2011-2014, The OpenBLAS Project
  3. All rights reserved.
  4. Redistribution and use in source and binary forms, with or without
  5. modification, are permitted provided that the following conditions are
  6. met:
  7. 1. Redistributions of source code must retain the above copyright
  8. notice, this list of conditions and the following disclaimer.
  9. 2. Redistributions in binary form must reproduce the above copyright
  10. notice, this list of conditions and the following disclaimer in
  11. the documentation and/or other materials provided with the
  12. distribution.
  13. 3. Neither the name of the OpenBLAS project nor the names of
  14. its contributors may be used to endorse or promote products
  15. derived from this software without specific prior written
  16. permission.
  17. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  18. AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19. IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20. ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
  21. LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22. DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  23. SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  24. CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  25. OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
  26. USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27. **********************************************************************************/
  28. /*********************************************************************/
  29. /* Copyright 2009, 2010 The University of Texas at Austin. */
  30. /* All rights reserved. */
  31. /* */
  32. /* Redistribution and use in source and binary forms, with or */
  33. /* without modification, are permitted provided that the following */
  34. /* conditions are met: */
  35. /* */
  36. /* 1. Redistributions of source code must retain the above */
  37. /* copyright notice, this list of conditions and the following */
  38. /* disclaimer. */
  39. /* */
  40. /* 2. Redistributions in binary form must reproduce the above */
  41. /* copyright notice, this list of conditions and the following */
  42. /* disclaimer in the documentation and/or other materials */
  43. /* provided with the distribution. */
  44. /* */
  45. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  46. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  47. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  48. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  49. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  50. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  51. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  52. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  53. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  54. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  55. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  56. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  57. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  58. /* POSSIBILITY OF SUCH DAMAGE. */
  59. /* */
  60. /* The views and conclusions contained in the software and */
  61. /* documentation are those of the authors and should not be */
  62. /* interpreted as representing official policies, either expressed */
  63. /* or implied, of The University of Texas at Austin. */
  64. /*********************************************************************/
  65. #ifndef PARAM_H
  66. #define PARAM_H
  67. #define SBGEMM_DEFAULT_UNROLL_N 4
  68. #define SBGEMM_DEFAULT_UNROLL_M 8
  69. #define SBGEMM_DEFAULT_UNROLL_MN 32
  70. #define SBGEMM_DEFAULT_P 256
  71. #define SBGEMM_DEFAULT_R 256
  72. #define SBGEMM_DEFAULT_Q 256
  73. #ifdef OPTERON
  74. #define SNUMOPT 4
  75. #define DNUMOPT 2
  76. #define GEMM_DEFAULT_OFFSET_A 64
  77. #define GEMM_DEFAULT_OFFSET_B 256
  78. #define GEMM_DEFAULT_ALIGN 0x01ffffUL
  79. #define SGEMM_DEFAULT_UNROLL_N 4
  80. #define DGEMM_DEFAULT_UNROLL_N 4
  81. #define QGEMM_DEFAULT_UNROLL_N 2
  82. #define CGEMM_DEFAULT_UNROLL_N 2
  83. #define ZGEMM_DEFAULT_UNROLL_N 2
  84. #define XGEMM_DEFAULT_UNROLL_N 1
  85. #ifdef ARCH_X86
  86. #define SGEMM_DEFAULT_UNROLL_M 4
  87. #define DGEMM_DEFAULT_UNROLL_M 2
  88. #define QGEMM_DEFAULT_UNROLL_M 2
  89. #define CGEMM_DEFAULT_UNROLL_M 2
  90. #define ZGEMM_DEFAULT_UNROLL_M 1
  91. #define XGEMM_DEFAULT_UNROLL_M 1
  92. #else
  93. #define SGEMM_DEFAULT_UNROLL_M 8
  94. #define DGEMM_DEFAULT_UNROLL_M 4
  95. #define QGEMM_DEFAULT_UNROLL_M 2
  96. #define CGEMM_DEFAULT_UNROLL_M 4
  97. #define ZGEMM_DEFAULT_UNROLL_M 2
  98. #define XGEMM_DEFAULT_UNROLL_M 1
  99. #endif
  100. #define SGEMM_DEFAULT_P sgemm_p
  101. #define DGEMM_DEFAULT_P dgemm_p
  102. #define QGEMM_DEFAULT_P qgemm_p
  103. #define CGEMM_DEFAULT_P cgemm_p
  104. #define ZGEMM_DEFAULT_P zgemm_p
  105. #define XGEMM_DEFAULT_P xgemm_p
  106. #define SGEMM_DEFAULT_R sgemm_r
  107. #define DGEMM_DEFAULT_R dgemm_r
  108. #define QGEMM_DEFAULT_R qgemm_r
  109. #define CGEMM_DEFAULT_R cgemm_r
  110. #define ZGEMM_DEFAULT_R zgemm_r
  111. #define XGEMM_DEFAULT_R xgemm_r
  112. #ifdef ALLOC_HUGETLB
  113. #define SGEMM_DEFAULT_Q 248
  114. #define DGEMM_DEFAULT_Q 248
  115. #define QGEMM_DEFAULT_Q 248
  116. #define CGEMM_DEFAULT_Q 248
  117. #define ZGEMM_DEFAULT_Q 248
  118. #define XGEMM_DEFAULT_Q 248
  119. #else
  120. #define SGEMM_DEFAULT_Q 240
  121. #define DGEMM_DEFAULT_Q 240
  122. #define QGEMM_DEFAULT_Q 240
  123. #define CGEMM_DEFAULT_Q 240
  124. #define ZGEMM_DEFAULT_Q 240
  125. #define XGEMM_DEFAULT_Q 240
  126. #endif
  127. #define SYMV_P 16
  128. #define HAVE_EXCLUSIVE_CACHE
  129. #endif
  130. #if defined(BARCELONA) || defined(SHANGHAI) || defined(BOBCAT)
  131. #define SNUMOPT 8
  132. #define DNUMOPT 4
  133. #define GEMM_DEFAULT_OFFSET_A 64
  134. #define GEMM_DEFAULT_OFFSET_B 832
  135. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  136. #define SGEMM_DEFAULT_UNROLL_N 4
  137. #define DGEMM_DEFAULT_UNROLL_N 4
  138. #define QGEMM_DEFAULT_UNROLL_N 2
  139. #define CGEMM_DEFAULT_UNROLL_N 2
  140. #define ZGEMM_DEFAULT_UNROLL_N 2
  141. #define XGEMM_DEFAULT_UNROLL_N 1
  142. #ifdef ARCH_X86
  143. #define SGEMM_DEFAULT_UNROLL_M 4
  144. #define DGEMM_DEFAULT_UNROLL_M 2
  145. #define QGEMM_DEFAULT_UNROLL_M 2
  146. #define CGEMM_DEFAULT_UNROLL_M 2
  147. #define ZGEMM_DEFAULT_UNROLL_M 1
  148. #define XGEMM_DEFAULT_UNROLL_M 1
  149. #else
  150. #define SGEMM_DEFAULT_UNROLL_M 8
  151. #define DGEMM_DEFAULT_UNROLL_M 4
  152. #define QGEMM_DEFAULT_UNROLL_M 2
  153. #define CGEMM_DEFAULT_UNROLL_M 4
  154. #define ZGEMM_DEFAULT_UNROLL_M 2
  155. #define XGEMM_DEFAULT_UNROLL_M 1
  156. #endif
  157. #if 0
  158. #define SGEMM_DEFAULT_P 496
  159. #define DGEMM_DEFAULT_P 248
  160. #define QGEMM_DEFAULT_P 124
  161. #define CGEMM_DEFAULT_P 248
  162. #define ZGEMM_DEFAULT_P 124
  163. #define XGEMM_DEFAULT_P 62
  164. #define SGEMM_DEFAULT_Q 248
  165. #define DGEMM_DEFAULT_Q 248
  166. #define QGEMM_DEFAULT_Q 248
  167. #define CGEMM_DEFAULT_Q 248
  168. #define ZGEMM_DEFAULT_Q 248
  169. #define XGEMM_DEFAULT_Q 248
  170. #else
  171. #define SGEMM_DEFAULT_P 448
  172. #define DGEMM_DEFAULT_P 224
  173. #define QGEMM_DEFAULT_P 112
  174. #define CGEMM_DEFAULT_P 224
  175. #define ZGEMM_DEFAULT_P 112
  176. #define XGEMM_DEFAULT_P 56
  177. #define SGEMM_DEFAULT_Q 224
  178. #define DGEMM_DEFAULT_Q 224
  179. #define QGEMM_DEFAULT_Q 224
  180. #define CGEMM_DEFAULT_Q 224
  181. #define ZGEMM_DEFAULT_Q 224
  182. #define XGEMM_DEFAULT_Q 224
  183. #endif
  184. #define SGEMM_DEFAULT_R sgemm_r
  185. #define QGEMM_DEFAULT_R qgemm_r
  186. #define DGEMM_DEFAULT_R dgemm_r
  187. #define CGEMM_DEFAULT_R cgemm_r
  188. #define ZGEMM_DEFAULT_R zgemm_r
  189. #define XGEMM_DEFAULT_R xgemm_r
  190. #define SYMV_P 16
  191. #define HAVE_EXCLUSIVE_CACHE
  192. #define GEMM_THREAD gemm_thread_mn
  193. #endif
  194. #ifdef BULLDOZER
  195. #define SNUMOPT 8
  196. #define DNUMOPT 4
  197. #define GEMM_DEFAULT_OFFSET_A 64
  198. #define GEMM_DEFAULT_OFFSET_B 832
  199. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  200. #define QGEMM_DEFAULT_UNROLL_N 2
  201. #define CGEMM_DEFAULT_UNROLL_N 2
  202. #define ZGEMM_DEFAULT_UNROLL_N 2
  203. #define XGEMM_DEFAULT_UNROLL_N 1
  204. #ifdef ARCH_X86
  205. #define SGEMM_DEFAULT_UNROLL_N 4
  206. #define DGEMM_DEFAULT_UNROLL_N 4
  207. #define SGEMM_DEFAULT_UNROLL_M 4
  208. #define DGEMM_DEFAULT_UNROLL_M 2
  209. #define QGEMM_DEFAULT_UNROLL_M 2
  210. #define CGEMM_DEFAULT_UNROLL_M 2
  211. #define ZGEMM_DEFAULT_UNROLL_M 1
  212. #define XGEMM_DEFAULT_UNROLL_M 1
  213. #else
  214. #define SGEMM_DEFAULT_UNROLL_N 2
  215. #define DGEMM_DEFAULT_UNROLL_N 2
  216. #define SGEMM_DEFAULT_UNROLL_M 16
  217. #define DGEMM_DEFAULT_UNROLL_M 8
  218. #define QGEMM_DEFAULT_UNROLL_M 2
  219. #define CGEMM_DEFAULT_UNROLL_M 4
  220. #define ZGEMM_DEFAULT_UNROLL_M 2
  221. #define XGEMM_DEFAULT_UNROLL_M 1
  222. #define CGEMM3M_DEFAULT_UNROLL_N 4
  223. #define CGEMM3M_DEFAULT_UNROLL_M 8
  224. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  225. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  226. #define DGEMM_DEFAULT_UNROLL_MN 16
  227. #define GEMV_UNROLL 8
  228. #endif
  229. #if defined(ARCH_X86_64)
  230. #define SGEMM_DEFAULT_P 768
  231. #define DGEMM_DEFAULT_P 384
  232. #else
  233. #define SGEMM_DEFAULT_P 448
  234. #define DGEMM_DEFAULT_P 224
  235. #endif
  236. #define QGEMM_DEFAULT_P 112
  237. #define CGEMM_DEFAULT_P 224
  238. #define ZGEMM_DEFAULT_P 112
  239. #define XGEMM_DEFAULT_P 56
  240. #if defined(ARCH_X86_64)
  241. #define SGEMM_DEFAULT_Q 168
  242. #define DGEMM_DEFAULT_Q 168
  243. #else
  244. #define SGEMM_DEFAULT_Q 224
  245. #define DGEMM_DEFAULT_Q 224
  246. #endif
  247. #define QGEMM_DEFAULT_Q 224
  248. #define CGEMM_DEFAULT_Q 224
  249. #define ZGEMM_DEFAULT_Q 224
  250. #define XGEMM_DEFAULT_Q 224
  251. #define CGEMM3M_DEFAULT_P 448
  252. #define ZGEMM3M_DEFAULT_P 224
  253. #define XGEMM3M_DEFAULT_P 112
  254. #define CGEMM3M_DEFAULT_Q 224
  255. #define ZGEMM3M_DEFAULT_Q 224
  256. #define XGEMM3M_DEFAULT_Q 224
  257. #define CGEMM3M_DEFAULT_R 12288
  258. #define ZGEMM3M_DEFAULT_R 12288
  259. #define XGEMM3M_DEFAULT_R 12288
  260. #define SGEMM_DEFAULT_R sgemm_r
  261. #define QGEMM_DEFAULT_R qgemm_r
  262. #define DGEMM_DEFAULT_R dgemm_r
  263. #define CGEMM_DEFAULT_R cgemm_r
  264. #define ZGEMM_DEFAULT_R zgemm_r
  265. #define XGEMM_DEFAULT_R xgemm_r
  266. #define SYMV_P 16
  267. #define HAVE_EXCLUSIVE_CACHE
  268. #define GEMM_THREAD gemm_thread_mn
  269. #endif
  270. #ifdef PILEDRIVER
  271. #define SNUMOPT 8
  272. #define DNUMOPT 4
  273. #define GEMM_DEFAULT_OFFSET_A 64
  274. #define GEMM_DEFAULT_OFFSET_B 832
  275. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  276. #define QGEMM_DEFAULT_UNROLL_N 2
  277. #define CGEMM_DEFAULT_UNROLL_N 2
  278. #define ZGEMM_DEFAULT_UNROLL_N 2
  279. #define XGEMM_DEFAULT_UNROLL_N 1
  280. #ifdef ARCH_X86
  281. #define SGEMM_DEFAULT_UNROLL_N 4
  282. #define DGEMM_DEFAULT_UNROLL_N 4
  283. #define SGEMM_DEFAULT_UNROLL_M 4
  284. #define DGEMM_DEFAULT_UNROLL_M 2
  285. #define QGEMM_DEFAULT_UNROLL_M 2
  286. #define CGEMM_DEFAULT_UNROLL_M 2
  287. #define ZGEMM_DEFAULT_UNROLL_M 1
  288. #define XGEMM_DEFAULT_UNROLL_M 1
  289. #else
  290. #define SGEMM_DEFAULT_UNROLL_N 2
  291. #define DGEMM_DEFAULT_UNROLL_N 2
  292. #define SGEMM_DEFAULT_UNROLL_M 16
  293. #define DGEMM_DEFAULT_UNROLL_M 8
  294. #define QGEMM_DEFAULT_UNROLL_M 2
  295. #define CGEMM_DEFAULT_UNROLL_M 4
  296. #define ZGEMM_DEFAULT_UNROLL_M 2
  297. #define XGEMM_DEFAULT_UNROLL_M 1
  298. #define CGEMM3M_DEFAULT_UNROLL_N 4
  299. #define CGEMM3M_DEFAULT_UNROLL_M 8
  300. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  301. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  302. #define GEMV_UNROLL 8
  303. #endif
  304. #if defined(ARCH_X86_64)
  305. #define SGEMM_DEFAULT_P 768
  306. #define DGEMM_DEFAULT_P 768
  307. #define ZGEMM_DEFAULT_P 384
  308. #define CGEMM_DEFAULT_P 768
  309. #else
  310. #define SGEMM_DEFAULT_P 448
  311. #define DGEMM_DEFAULT_P 480
  312. #define ZGEMM_DEFAULT_P 112
  313. #define CGEMM_DEFAULT_P 224
  314. #endif
  315. #define QGEMM_DEFAULT_P 112
  316. #define XGEMM_DEFAULT_P 56
  317. #if defined(ARCH_X86_64)
  318. #define SGEMM_DEFAULT_Q 192
  319. #define DGEMM_DEFAULT_Q 168
  320. #define ZGEMM_DEFAULT_Q 168
  321. #define CGEMM_DEFAULT_Q 168
  322. #else
  323. #define SGEMM_DEFAULT_Q 224
  324. #define DGEMM_DEFAULT_Q 224
  325. #define ZGEMM_DEFAULT_Q 224
  326. #define CGEMM_DEFAULT_Q 224
  327. #endif
  328. #define QGEMM_DEFAULT_Q 224
  329. #define XGEMM_DEFAULT_Q 224
  330. #define CGEMM3M_DEFAULT_P 448
  331. #define ZGEMM3M_DEFAULT_P 224
  332. #define XGEMM3M_DEFAULT_P 112
  333. #define CGEMM3M_DEFAULT_Q 224
  334. #define ZGEMM3M_DEFAULT_Q 224
  335. #define XGEMM3M_DEFAULT_Q 224
  336. #define CGEMM3M_DEFAULT_R 12288
  337. #define ZGEMM3M_DEFAULT_R 12288
  338. #define XGEMM3M_DEFAULT_R 12288
  339. #define SGEMM_DEFAULT_R 12288
  340. #define QGEMM_DEFAULT_R qgemm_r
  341. #define DGEMM_DEFAULT_R 12288
  342. #define CGEMM_DEFAULT_R cgemm_r
  343. #define ZGEMM_DEFAULT_R zgemm_r
  344. #define XGEMM_DEFAULT_R xgemm_r
  345. #define SYMV_P 16
  346. #define HAVE_EXCLUSIVE_CACHE
  347. #define GEMM_THREAD gemm_thread_mn
  348. #endif
  349. #ifdef STEAMROLLER
  350. #define SNUMOPT 8
  351. #define DNUMOPT 4
  352. #define GEMM_DEFAULT_OFFSET_A 64
  353. #define GEMM_DEFAULT_OFFSET_B 832
  354. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  355. #define QGEMM_DEFAULT_UNROLL_N 2
  356. #define CGEMM_DEFAULT_UNROLL_N 2
  357. #define ZGEMM_DEFAULT_UNROLL_N 2
  358. #define XGEMM_DEFAULT_UNROLL_N 1
  359. #ifdef ARCH_X86
  360. #define SGEMM_DEFAULT_UNROLL_N 4
  361. #define DGEMM_DEFAULT_UNROLL_N 4
  362. #define SGEMM_DEFAULT_UNROLL_M 4
  363. #define DGEMM_DEFAULT_UNROLL_M 2
  364. #define QGEMM_DEFAULT_UNROLL_M 2
  365. #define CGEMM_DEFAULT_UNROLL_M 2
  366. #define ZGEMM_DEFAULT_UNROLL_M 1
  367. #define XGEMM_DEFAULT_UNROLL_M 1
  368. #else
  369. #define SGEMM_DEFAULT_UNROLL_N 2
  370. #define DGEMM_DEFAULT_UNROLL_N 2
  371. #define SGEMM_DEFAULT_UNROLL_M 16
  372. #define DGEMM_DEFAULT_UNROLL_M 8
  373. #define QGEMM_DEFAULT_UNROLL_M 2
  374. #define CGEMM_DEFAULT_UNROLL_M 4
  375. #define ZGEMM_DEFAULT_UNROLL_M 2
  376. #define XGEMM_DEFAULT_UNROLL_M 1
  377. #define CGEMM3M_DEFAULT_UNROLL_N 4
  378. #define CGEMM3M_DEFAULT_UNROLL_M 8
  379. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  380. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  381. #define GEMV_UNROLL 8
  382. #endif
  383. #if defined(ARCH_X86_64)
  384. #define SGEMM_DEFAULT_P 768
  385. #define DGEMM_DEFAULT_P 576
  386. #define ZGEMM_DEFAULT_P 288
  387. #define CGEMM_DEFAULT_P 576
  388. #else
  389. #define SGEMM_DEFAULT_P 448
  390. #define DGEMM_DEFAULT_P 480
  391. #define ZGEMM_DEFAULT_P 112
  392. #define CGEMM_DEFAULT_P 224
  393. #endif
  394. #define QGEMM_DEFAULT_P 112
  395. #define XGEMM_DEFAULT_P 56
  396. #if defined(ARCH_X86_64)
  397. #define SGEMM_DEFAULT_Q 192
  398. #define DGEMM_DEFAULT_Q 160
  399. #define ZGEMM_DEFAULT_Q 160
  400. #define CGEMM_DEFAULT_Q 160
  401. #else
  402. #define SGEMM_DEFAULT_Q 224
  403. #define DGEMM_DEFAULT_Q 224
  404. #define ZGEMM_DEFAULT_Q 224
  405. #define CGEMM_DEFAULT_Q 224
  406. #endif
  407. #define QGEMM_DEFAULT_Q 224
  408. #define XGEMM_DEFAULT_Q 224
  409. #define CGEMM3M_DEFAULT_P 448
  410. #define ZGEMM3M_DEFAULT_P 224
  411. #define XGEMM3M_DEFAULT_P 112
  412. #define CGEMM3M_DEFAULT_Q 224
  413. #define ZGEMM3M_DEFAULT_Q 224
  414. #define XGEMM3M_DEFAULT_Q 224
  415. #define CGEMM3M_DEFAULT_R 12288
  416. #define ZGEMM3M_DEFAULT_R 12288
  417. #define XGEMM3M_DEFAULT_R 12288
  418. #define SGEMM_DEFAULT_R 12288
  419. #define QGEMM_DEFAULT_R qgemm_r
  420. #define DGEMM_DEFAULT_R 12288
  421. #define CGEMM_DEFAULT_R cgemm_r
  422. #define ZGEMM_DEFAULT_R zgemm_r
  423. #define XGEMM_DEFAULT_R xgemm_r
  424. #define SYMV_P 16
  425. #define HAVE_EXCLUSIVE_CACHE
  426. #define GEMM_THREAD gemm_thread_mn
  427. #endif
  428. #ifdef EXCAVATOR
  429. #define SNUMOPT 8
  430. #define DNUMOPT 4
  431. #define GEMM_DEFAULT_OFFSET_A 64
  432. #define GEMM_DEFAULT_OFFSET_B 832
  433. #define GEMM_DEFAULT_ALIGN 0x0fffUL
  434. #define QGEMM_DEFAULT_UNROLL_N 2
  435. #define CGEMM_DEFAULT_UNROLL_N 2
  436. #define ZGEMM_DEFAULT_UNROLL_N 2
  437. #define XGEMM_DEFAULT_UNROLL_N 1
  438. #ifdef ARCH_X86
  439. #define SGEMM_DEFAULT_UNROLL_N 4
  440. #define DGEMM_DEFAULT_UNROLL_N 4
  441. #define SGEMM_DEFAULT_UNROLL_M 4
  442. #define DGEMM_DEFAULT_UNROLL_M 2
  443. #define QGEMM_DEFAULT_UNROLL_M 2
  444. #define CGEMM_DEFAULT_UNROLL_M 2
  445. #define ZGEMM_DEFAULT_UNROLL_M 1
  446. #define XGEMM_DEFAULT_UNROLL_M 1
  447. #else
  448. #define SGEMM_DEFAULT_UNROLL_N 2
  449. #define DGEMM_DEFAULT_UNROLL_N 2
  450. #define SGEMM_DEFAULT_UNROLL_M 16
  451. #define DGEMM_DEFAULT_UNROLL_M 8
  452. #define QGEMM_DEFAULT_UNROLL_M 2
  453. #define CGEMM_DEFAULT_UNROLL_M 4
  454. #define ZGEMM_DEFAULT_UNROLL_M 2
  455. #define XGEMM_DEFAULT_UNROLL_M 1
  456. #define CGEMM3M_DEFAULT_UNROLL_N 4
  457. #define CGEMM3M_DEFAULT_UNROLL_M 8
  458. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  459. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  460. #define GEMV_UNROLL 8
  461. #endif
  462. #if defined(ARCH_X86_64)
  463. #define SGEMM_DEFAULT_P 768
  464. #define DGEMM_DEFAULT_P 576
  465. #define ZGEMM_DEFAULT_P 288
  466. #define CGEMM_DEFAULT_P 576
  467. #else
  468. #define SGEMM_DEFAULT_P 448
  469. #define DGEMM_DEFAULT_P 480
  470. #define ZGEMM_DEFAULT_P 112
  471. #define CGEMM_DEFAULT_P 224
  472. #endif
  473. #define QGEMM_DEFAULT_P 112
  474. #define XGEMM_DEFAULT_P 56
  475. #if defined(ARCH_X86_64)
  476. #define SGEMM_DEFAULT_Q 192
  477. #define DGEMM_DEFAULT_Q 160
  478. #define ZGEMM_DEFAULT_Q 160
  479. #define CGEMM_DEFAULT_Q 160
  480. #else
  481. #define SGEMM_DEFAULT_Q 224
  482. #define DGEMM_DEFAULT_Q 224
  483. #define ZGEMM_DEFAULT_Q 224
  484. #define CGEMM_DEFAULT_Q 224
  485. #endif
  486. #define QGEMM_DEFAULT_Q 224
  487. #define XGEMM_DEFAULT_Q 224
  488. #define CGEMM3M_DEFAULT_P 448
  489. #define ZGEMM3M_DEFAULT_P 224
  490. #define XGEMM3M_DEFAULT_P 112
  491. #define CGEMM3M_DEFAULT_Q 224
  492. #define ZGEMM3M_DEFAULT_Q 224
  493. #define XGEMM3M_DEFAULT_Q 224
  494. #define CGEMM3M_DEFAULT_R 12288
  495. #define ZGEMM3M_DEFAULT_R 12288
  496. #define XGEMM3M_DEFAULT_R 12288
  497. #define SGEMM_DEFAULT_R 12288
  498. #define QGEMM_DEFAULT_R qgemm_r
  499. #define DGEMM_DEFAULT_R 12288
  500. #define CGEMM_DEFAULT_R cgemm_r
  501. #define ZGEMM_DEFAULT_R zgemm_r
  502. #define XGEMM_DEFAULT_R xgemm_r
  503. #define SYMV_P 16
  504. #define HAVE_EXCLUSIVE_CACHE
  505. #define GEMM_THREAD gemm_thread_mn
  506. #endif
  507. #ifdef ZEN
  508. #define SNUMOPT 16
  509. #define DNUMOPT 8
  510. #define GEMM_DEFAULT_OFFSET_A 0
  511. #define GEMM_DEFAULT_OFFSET_B 0
  512. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  513. #define SYMV_P 8
  514. #define SWITCH_RATIO 16
  515. #ifdef ARCH_X86
  516. #define SGEMM_DEFAULT_UNROLL_M 4
  517. #define DGEMM_DEFAULT_UNROLL_M 2
  518. #define QGEMM_DEFAULT_UNROLL_M 2
  519. #define CGEMM_DEFAULT_UNROLL_M 2
  520. #define ZGEMM_DEFAULT_UNROLL_M 1
  521. #define XGEMM_DEFAULT_UNROLL_M 1
  522. #define SGEMM_DEFAULT_UNROLL_N 4
  523. #define DGEMM_DEFAULT_UNROLL_N 4
  524. #define QGEMM_DEFAULT_UNROLL_N 2
  525. #define CGEMM_DEFAULT_UNROLL_N 2
  526. #define ZGEMM_DEFAULT_UNROLL_N 2
  527. #define XGEMM_DEFAULT_UNROLL_N 1
  528. #else
  529. #define SGEMM_DEFAULT_UNROLL_M 8
  530. #define DGEMM_DEFAULT_UNROLL_M 4
  531. #define QGEMM_DEFAULT_UNROLL_M 2
  532. #define CGEMM_DEFAULT_UNROLL_M 8
  533. #define ZGEMM_DEFAULT_UNROLL_M 4
  534. #define XGEMM_DEFAULT_UNROLL_M 1
  535. #define SGEMM_DEFAULT_UNROLL_N 4
  536. #define DGEMM_DEFAULT_UNROLL_N 8
  537. #define QGEMM_DEFAULT_UNROLL_N 2
  538. #define CGEMM_DEFAULT_UNROLL_N 2
  539. #define ZGEMM_DEFAULT_UNROLL_N 2
  540. #define XGEMM_DEFAULT_UNROLL_N 1
  541. /*
  542. #define SGEMM_DEFAULT_UNROLL_MN 32
  543. #define DGEMM_DEFAULT_UNROLL_MN 32
  544. */
  545. #endif
  546. #ifdef ARCH_X86
  547. #define SGEMM_DEFAULT_P 512
  548. #define SGEMM_DEFAULT_R sgemm_r
  549. #define DGEMM_DEFAULT_P 512
  550. #define DGEMM_DEFAULT_R dgemm_r
  551. #define QGEMM_DEFAULT_P 504
  552. #define QGEMM_DEFAULT_R qgemm_r
  553. #define CGEMM_DEFAULT_P 128
  554. #define CGEMM_DEFAULT_R 1024
  555. #define ZGEMM_DEFAULT_P 512
  556. #define ZGEMM_DEFAULT_R zgemm_r
  557. #define XGEMM_DEFAULT_P 252
  558. #define XGEMM_DEFAULT_R xgemm_r
  559. #define SGEMM_DEFAULT_Q 256
  560. #define DGEMM_DEFAULT_Q 256
  561. #define QGEMM_DEFAULT_Q 128
  562. #define CGEMM_DEFAULT_Q 256
  563. #define ZGEMM_DEFAULT_Q 192
  564. #define XGEMM_DEFAULT_Q 128
  565. #else
  566. #define SGEMM_DEFAULT_P 320
  567. #define DGEMM_DEFAULT_P 512
  568. #define CGEMM_DEFAULT_P 256
  569. #define ZGEMM_DEFAULT_P 192
  570. #ifdef WINDOWS_ABI
  571. #define SGEMM_DEFAULT_Q 320
  572. #define DGEMM_DEFAULT_Q 128
  573. #else
  574. #define SGEMM_DEFAULT_Q 320
  575. #define DGEMM_DEFAULT_Q 256
  576. #endif
  577. #define CGEMM_DEFAULT_Q 256
  578. #define ZGEMM_DEFAULT_Q 192
  579. #define SGEMM_DEFAULT_R sgemm_r
  580. #define DGEMM_DEFAULT_R 13824
  581. #define CGEMM_DEFAULT_R cgemm_r
  582. #define ZGEMM_DEFAULT_R zgemm_r
  583. #define QGEMM_DEFAULT_Q 128
  584. #define QGEMM_DEFAULT_P 504
  585. #define QGEMM_DEFAULT_R qgemm_r
  586. #define XGEMM_DEFAULT_P 252
  587. #define XGEMM_DEFAULT_R xgemm_r
  588. #define XGEMM_DEFAULT_Q 128
  589. #define CGEMM3M_DEFAULT_UNROLL_N 4
  590. #define CGEMM3M_DEFAULT_UNROLL_M 8
  591. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  592. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  593. #define CGEMM3M_DEFAULT_P 320
  594. #define ZGEMM3M_DEFAULT_P 256
  595. #define XGEMM3M_DEFAULT_P 112
  596. #define CGEMM3M_DEFAULT_Q 320
  597. #define ZGEMM3M_DEFAULT_Q 256
  598. #define XGEMM3M_DEFAULT_Q 224
  599. #define CGEMM3M_DEFAULT_R 12288
  600. #define ZGEMM3M_DEFAULT_R 12288
  601. #define XGEMM3M_DEFAULT_R 12288
  602. #endif
  603. #endif
  604. #ifdef ATHLON
  605. #define SNUMOPT 4
  606. #define DNUMOPT 2
  607. #define GEMM_DEFAULT_OFFSET_A 0
  608. #define GEMM_DEFAULT_OFFSET_B 384
  609. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  610. #define SGEMM_DEFAULT_UNROLL_N 4
  611. #define DGEMM_DEFAULT_UNROLL_N 4
  612. #define QGEMM_DEFAULT_UNROLL_N 2
  613. #define CGEMM_DEFAULT_UNROLL_N 2
  614. #define ZGEMM_DEFAULT_UNROLL_N 2
  615. #define XGEMM_DEFAULT_UNROLL_N 1
  616. #define SGEMM_DEFAULT_UNROLL_M 2
  617. #define DGEMM_DEFAULT_UNROLL_M 1
  618. #define QGEMM_DEFAULT_UNROLL_M 2
  619. #define CGEMM_DEFAULT_UNROLL_M 1
  620. #define ZGEMM_DEFAULT_UNROLL_M 1
  621. #define XGEMM_DEFAULT_UNROLL_M 1
  622. #define SGEMM_DEFAULT_R sgemm_r
  623. #define DGEMM_DEFAULT_R dgemm_r
  624. #define QGEMM_DEFAULT_R qgemm_r
  625. #define CGEMM_DEFAULT_R cgemm_r
  626. #define ZGEMM_DEFAULT_R zgemm_r
  627. #define XGEMM_DEFAULT_R xgemm_r
  628. #define SGEMM_DEFAULT_P 208
  629. #define DGEMM_DEFAULT_P 104
  630. #define QGEMM_DEFAULT_P 56
  631. #define CGEMM_DEFAULT_P 104
  632. #define ZGEMM_DEFAULT_P 56
  633. #define XGEMM_DEFAULT_P 28
  634. #define SGEMM_DEFAULT_Q 208
  635. #define DGEMM_DEFAULT_Q 208
  636. #define QGEMM_DEFAULT_Q 208
  637. #define CGEMM_DEFAULT_Q 208
  638. #define ZGEMM_DEFAULT_Q 208
  639. #define XGEMM_DEFAULT_Q 208
  640. #define SYMV_P 16
  641. #define HAVE_EXCLUSIVE_CACHE
  642. #endif
  643. #ifdef VIAC3
  644. #define SNUMOPT 2
  645. #define DNUMOPT 1
  646. #define GEMM_DEFAULT_OFFSET_A 0
  647. #define GEMM_DEFAULT_OFFSET_B 256
  648. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  649. #define SGEMM_DEFAULT_UNROLL_N 4
  650. #define DGEMM_DEFAULT_UNROLL_N 4
  651. #define QGEMM_DEFAULT_UNROLL_N 2
  652. #define CGEMM_DEFAULT_UNROLL_N 2
  653. #define ZGEMM_DEFAULT_UNROLL_N 2
  654. #define XGEMM_DEFAULT_UNROLL_N 1
  655. #define SGEMM_DEFAULT_UNROLL_M 2
  656. #define DGEMM_DEFAULT_UNROLL_M 1
  657. #define QGEMM_DEFAULT_UNROLL_M 2
  658. #define CGEMM_DEFAULT_UNROLL_M 1
  659. #define ZGEMM_DEFAULT_UNROLL_M 1
  660. #define XGEMM_DEFAULT_UNROLL_M 1
  661. #define SGEMM_DEFAULT_R sgemm_r
  662. #define DGEMM_DEFAULT_R dgemm_r
  663. #define QGEMM_DEFAULT_R qgemm_r
  664. #define CGEMM_DEFAULT_R cgemm_r
  665. #define ZGEMM_DEFAULT_R zgemm_r
  666. #define XGEMM_DEFAULT_R xgemm_r
  667. #define SGEMM_DEFAULT_P 128
  668. #define DGEMM_DEFAULT_P 128
  669. #define QGEMM_DEFAULT_P 128
  670. #define CGEMM_DEFAULT_P 128
  671. #define ZGEMM_DEFAULT_P 128
  672. #define XGEMM_DEFAULT_P 128
  673. #define SGEMM_DEFAULT_Q 512
  674. #define DGEMM_DEFAULT_Q 256
  675. #define QGEMM_DEFAULT_Q 256
  676. #define CGEMM_DEFAULT_Q 256
  677. #define ZGEMM_DEFAULT_Q 128
  678. #define XGEMM_DEFAULT_Q 128
  679. #define SYMV_P 16
  680. #endif
  681. #ifdef NANO
  682. #define SNUMOPT 4
  683. #define DNUMOPT 2
  684. #define GEMM_DEFAULT_OFFSET_A 64
  685. #define GEMM_DEFAULT_OFFSET_B 256
  686. #define GEMM_DEFAULT_ALIGN 0x01ffffUL
  687. #ifdef ARCH_X86
  688. #define SGEMM_DEFAULT_UNROLL_N 4
  689. #define DGEMM_DEFAULT_UNROLL_N 4
  690. #define QGEMM_DEFAULT_UNROLL_N 2
  691. #define CGEMM_DEFAULT_UNROLL_N 2
  692. #define ZGEMM_DEFAULT_UNROLL_N 2
  693. #define XGEMM_DEFAULT_UNROLL_N 1
  694. #define SGEMM_DEFAULT_UNROLL_M 4
  695. #define DGEMM_DEFAULT_UNROLL_M 2
  696. #define QGEMM_DEFAULT_UNROLL_M 2
  697. #define CGEMM_DEFAULT_UNROLL_M 2
  698. #define ZGEMM_DEFAULT_UNROLL_M 1
  699. #define XGEMM_DEFAULT_UNROLL_M 1
  700. #else
  701. #define SGEMM_DEFAULT_UNROLL_N 8
  702. #define DGEMM_DEFAULT_UNROLL_N 4
  703. #define QGEMM_DEFAULT_UNROLL_N 2
  704. #define CGEMM_DEFAULT_UNROLL_N 4
  705. #define ZGEMM_DEFAULT_UNROLL_N 2
  706. #define XGEMM_DEFAULT_UNROLL_N 1
  707. #define SGEMM_DEFAULT_UNROLL_M 4
  708. #define DGEMM_DEFAULT_UNROLL_M 4
  709. #define QGEMM_DEFAULT_UNROLL_M 2
  710. #define CGEMM_DEFAULT_UNROLL_M 2
  711. #define ZGEMM_DEFAULT_UNROLL_M 2
  712. #define XGEMM_DEFAULT_UNROLL_M 1
  713. #endif
  714. #define SGEMM_DEFAULT_P 288
  715. #define DGEMM_DEFAULT_P 288
  716. #define QGEMM_DEFAULT_P 288
  717. #define CGEMM_DEFAULT_P 288
  718. #define ZGEMM_DEFAULT_P 288
  719. #define XGEMM_DEFAULT_P 288
  720. #define SGEMM_DEFAULT_R sgemm_r
  721. #define DGEMM_DEFAULT_R dgemm_r
  722. #define QGEMM_DEFAULT_R qgemm_r
  723. #define CGEMM_DEFAULT_R cgemm_r
  724. #define ZGEMM_DEFAULT_R zgemm_r
  725. #define XGEMM_DEFAULT_R xgemm_r
  726. #define SGEMM_DEFAULT_Q 256
  727. #define DGEMM_DEFAULT_Q 128
  728. #define QGEMM_DEFAULT_Q 64
  729. #define CGEMM_DEFAULT_Q 128
  730. #define ZGEMM_DEFAULT_Q 64
  731. #define XGEMM_DEFAULT_Q 32
  732. #define SYMV_P 16
  733. #define HAVE_EXCLUSIVE_CACHE
  734. #endif
  735. #if defined(PENTIUM) || defined(PENTIUM2) || defined(PENTIUM3)
  736. #ifdef HAVE_SSE
  737. #define SNUMOPT 2
  738. #else
  739. #define SNUMOPT 1
  740. #endif
  741. #define DNUMOPT 1
  742. #define GEMM_DEFAULT_OFFSET_A 0
  743. #define GEMM_DEFAULT_OFFSET_B 0
  744. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  745. #ifdef HAVE_SSE
  746. #define SGEMM_DEFAULT_UNROLL_M 8
  747. #define CGEMM_DEFAULT_UNROLL_M 4
  748. #else
  749. #define SGEMM_DEFAULT_UNROLL_M 4
  750. #define CGEMM_DEFAULT_UNROLL_M 2
  751. #endif
  752. #define DGEMM_DEFAULT_UNROLL_M 2
  753. #define SGEMM_DEFAULT_UNROLL_N 2
  754. #define DGEMM_DEFAULT_UNROLL_N 2
  755. #define QGEMM_DEFAULT_UNROLL_M 2
  756. #define QGEMM_DEFAULT_UNROLL_N 2
  757. #define CGEMM_DEFAULT_UNROLL_N 1
  758. #define ZGEMM_DEFAULT_UNROLL_M 1
  759. #define ZGEMM_DEFAULT_UNROLL_N 1
  760. #define XGEMM_DEFAULT_UNROLL_M 1
  761. #define XGEMM_DEFAULT_UNROLL_N 1
  762. #define SGEMM_DEFAULT_P sgemm_p
  763. #define SGEMM_DEFAULT_Q 256
  764. #define SGEMM_DEFAULT_R sgemm_r
  765. #define DGEMM_DEFAULT_P dgemm_p
  766. #define DGEMM_DEFAULT_Q 256
  767. #define DGEMM_DEFAULT_R dgemm_r
  768. #define QGEMM_DEFAULT_P qgemm_p
  769. #define QGEMM_DEFAULT_Q 256
  770. #define QGEMM_DEFAULT_R qgemm_r
  771. #define CGEMM_DEFAULT_P cgemm_p
  772. #define CGEMM_DEFAULT_Q 256
  773. #define CGEMM_DEFAULT_R cgemm_r
  774. #define ZGEMM_DEFAULT_P zgemm_p
  775. #define ZGEMM_DEFAULT_Q 256
  776. #define ZGEMM_DEFAULT_R zgemm_r
  777. #define XGEMM_DEFAULT_P xgemm_p
  778. #define XGEMM_DEFAULT_Q 256
  779. #define XGEMM_DEFAULT_R xgemm_r
  780. #define SYMV_P 4
  781. #endif
  782. #ifdef PENTIUMM
  783. #define SNUMOPT 2
  784. #define DNUMOPT 1
  785. #define GEMM_DEFAULT_OFFSET_A 0
  786. #define GEMM_DEFAULT_OFFSET_B 0
  787. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  788. #ifdef CORE_YONAH
  789. #define SGEMM_DEFAULT_UNROLL_M 4
  790. #define SGEMM_DEFAULT_UNROLL_N 4
  791. #define DGEMM_DEFAULT_UNROLL_M 2
  792. #define DGEMM_DEFAULT_UNROLL_N 4
  793. #define QGEMM_DEFAULT_UNROLL_M 2
  794. #define QGEMM_DEFAULT_UNROLL_N 2
  795. #define CGEMM_DEFAULT_UNROLL_M 2
  796. #define CGEMM_DEFAULT_UNROLL_N 2
  797. #define ZGEMM_DEFAULT_UNROLL_M 1
  798. #define ZGEMM_DEFAULT_UNROLL_N 2
  799. #define XGEMM_DEFAULT_UNROLL_M 1
  800. #define XGEMM_DEFAULT_UNROLL_N 1
  801. #else
  802. #define SGEMM_DEFAULT_UNROLL_M 8
  803. #define SGEMM_DEFAULT_UNROLL_N 2
  804. #define DGEMM_DEFAULT_UNROLL_M 2
  805. #define DGEMM_DEFAULT_UNROLL_N 2
  806. #define QGEMM_DEFAULT_UNROLL_M 2
  807. #define QGEMM_DEFAULT_UNROLL_N 2
  808. #define CGEMM_DEFAULT_UNROLL_M 4
  809. #define CGEMM_DEFAULT_UNROLL_N 1
  810. #define ZGEMM_DEFAULT_UNROLL_M 1
  811. #define ZGEMM_DEFAULT_UNROLL_N 1
  812. #define XGEMM_DEFAULT_UNROLL_M 1
  813. #define XGEMM_DEFAULT_UNROLL_N 1
  814. #endif
  815. #define SGEMM_DEFAULT_P sgemm_p
  816. #define SGEMM_DEFAULT_Q 256
  817. #define SGEMM_DEFAULT_R sgemm_r
  818. #define DGEMM_DEFAULT_P dgemm_p
  819. #define DGEMM_DEFAULT_Q 256
  820. #define DGEMM_DEFAULT_R dgemm_r
  821. #define QGEMM_DEFAULT_P qgemm_p
  822. #define QGEMM_DEFAULT_Q 256
  823. #define QGEMM_DEFAULT_R qgemm_r
  824. #define CGEMM_DEFAULT_P cgemm_p
  825. #define CGEMM_DEFAULT_Q 256
  826. #define CGEMM_DEFAULT_R cgemm_r
  827. #define ZGEMM_DEFAULT_P zgemm_p
  828. #define ZGEMM_DEFAULT_Q 256
  829. #define ZGEMM_DEFAULT_R zgemm_r
  830. #define XGEMM_DEFAULT_P xgemm_p
  831. #define XGEMM_DEFAULT_Q 256
  832. #define XGEMM_DEFAULT_R xgemm_r
  833. #define SYMV_P 4
  834. #endif
  835. #ifdef CORE_NORTHWOOD
  836. #define SNUMOPT 4
  837. #define DNUMOPT 2
  838. #define GEMM_DEFAULT_OFFSET_A 0
  839. #define GEMM_DEFAULT_OFFSET_B 32
  840. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  841. #define SYMV_P 8
  842. #define SGEMM_DEFAULT_UNROLL_M 8
  843. #define DGEMM_DEFAULT_UNROLL_M 4
  844. #define QGEMM_DEFAULT_UNROLL_M 2
  845. #define CGEMM_DEFAULT_UNROLL_M 4
  846. #define ZGEMM_DEFAULT_UNROLL_M 2
  847. #define XGEMM_DEFAULT_UNROLL_M 1
  848. #define SGEMM_DEFAULT_UNROLL_N 2
  849. #define DGEMM_DEFAULT_UNROLL_N 2
  850. #define QGEMM_DEFAULT_UNROLL_N 2
  851. #define CGEMM_DEFAULT_UNROLL_N 1
  852. #define ZGEMM_DEFAULT_UNROLL_N 1
  853. #define XGEMM_DEFAULT_UNROLL_N 1
  854. #define SGEMM_DEFAULT_P sgemm_p
  855. #define SGEMM_DEFAULT_R sgemm_r
  856. #define DGEMM_DEFAULT_P dgemm_p
  857. #define DGEMM_DEFAULT_R dgemm_r
  858. #define QGEMM_DEFAULT_P qgemm_p
  859. #define QGEMM_DEFAULT_R qgemm_r
  860. #define CGEMM_DEFAULT_P cgemm_p
  861. #define CGEMM_DEFAULT_R cgemm_r
  862. #define ZGEMM_DEFAULT_P zgemm_p
  863. #define ZGEMM_DEFAULT_R zgemm_r
  864. #define XGEMM_DEFAULT_P xgemm_p
  865. #define XGEMM_DEFAULT_R xgemm_r
  866. #define SGEMM_DEFAULT_Q 128
  867. #define DGEMM_DEFAULT_Q 128
  868. #define QGEMM_DEFAULT_Q 128
  869. #define CGEMM_DEFAULT_Q 128
  870. #define ZGEMM_DEFAULT_Q 128
  871. #define XGEMM_DEFAULT_Q 128
  872. #endif
  873. #ifdef CORE_PRESCOTT
  874. #define SNUMOPT 4
  875. #define DNUMOPT 2
  876. #ifndef __64BIT__
  877. #define GEMM_DEFAULT_OFFSET_A 128
  878. #define GEMM_DEFAULT_OFFSET_B 192
  879. #else
  880. #define GEMM_DEFAULT_OFFSET_A 0
  881. #define GEMM_DEFAULT_OFFSET_B 256
  882. #endif
  883. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  884. #define SYMV_P 8
  885. #ifdef ARCH_X86
  886. #define SGEMM_DEFAULT_UNROLL_M 4
  887. #define DGEMM_DEFAULT_UNROLL_M 2
  888. #define QGEMM_DEFAULT_UNROLL_M 2
  889. #define CGEMM_DEFAULT_UNROLL_M 2
  890. #define ZGEMM_DEFAULT_UNROLL_M 1
  891. #define XGEMM_DEFAULT_UNROLL_M 1
  892. #else
  893. #define SGEMM_DEFAULT_UNROLL_M 8
  894. #define DGEMM_DEFAULT_UNROLL_M 4
  895. #define QGEMM_DEFAULT_UNROLL_M 2
  896. #define CGEMM_DEFAULT_UNROLL_M 4
  897. #define ZGEMM_DEFAULT_UNROLL_M 2
  898. #define XGEMM_DEFAULT_UNROLL_M 1
  899. #endif
  900. #define SGEMM_DEFAULT_UNROLL_N 4
  901. #define DGEMM_DEFAULT_UNROLL_N 4
  902. #define QGEMM_DEFAULT_UNROLL_N 2
  903. #define CGEMM_DEFAULT_UNROLL_N 2
  904. #define ZGEMM_DEFAULT_UNROLL_N 2
  905. #define XGEMM_DEFAULT_UNROLL_N 1
  906. #define SGEMM_DEFAULT_P sgemm_p
  907. #define SGEMM_DEFAULT_R sgemm_r
  908. #define DGEMM_DEFAULT_P dgemm_p
  909. #define DGEMM_DEFAULT_R dgemm_r
  910. #define QGEMM_DEFAULT_P qgemm_p
  911. #define QGEMM_DEFAULT_R qgemm_r
  912. #define CGEMM_DEFAULT_P cgemm_p
  913. #define CGEMM_DEFAULT_R cgemm_r
  914. #define ZGEMM_DEFAULT_P zgemm_p
  915. #define ZGEMM_DEFAULT_R zgemm_r
  916. #define XGEMM_DEFAULT_P xgemm_p
  917. #define XGEMM_DEFAULT_R xgemm_r
  918. #define SGEMM_DEFAULT_Q 128
  919. #define DGEMM_DEFAULT_Q 128
  920. #define QGEMM_DEFAULT_Q 128
  921. #define CGEMM_DEFAULT_Q 128
  922. #define ZGEMM_DEFAULT_Q 128
  923. #define XGEMM_DEFAULT_Q 128
  924. #endif
  925. #ifdef CORE2
  926. #define SNUMOPT 8
  927. #define DNUMOPT 4
  928. #define GEMM_DEFAULT_OFFSET_A 448
  929. #define GEMM_DEFAULT_OFFSET_B 128
  930. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  931. #define SYMV_P 8
  932. #define SWITCH_RATIO 4
  933. #ifdef ARCH_X86
  934. #define SGEMM_DEFAULT_UNROLL_M 8
  935. #define DGEMM_DEFAULT_UNROLL_M 4
  936. #define QGEMM_DEFAULT_UNROLL_M 2
  937. #define CGEMM_DEFAULT_UNROLL_M 4
  938. #define ZGEMM_DEFAULT_UNROLL_M 2
  939. #define XGEMM_DEFAULT_UNROLL_M 1
  940. #define SGEMM_DEFAULT_UNROLL_N 2
  941. #define DGEMM_DEFAULT_UNROLL_N 2
  942. #define QGEMM_DEFAULT_UNROLL_N 2
  943. #define CGEMM_DEFAULT_UNROLL_N 1
  944. #define ZGEMM_DEFAULT_UNROLL_N 1
  945. #define XGEMM_DEFAULT_UNROLL_N 1
  946. #define MASK(a, b) ((((a) + (b) - 1) / (b)) * (b))
  947. #else
  948. #define SGEMM_DEFAULT_UNROLL_M 8
  949. #define DGEMM_DEFAULT_UNROLL_M 4
  950. #define QGEMM_DEFAULT_UNROLL_M 2
  951. #define CGEMM_DEFAULT_UNROLL_M 4
  952. #define ZGEMM_DEFAULT_UNROLL_M 2
  953. #define XGEMM_DEFAULT_UNROLL_M 1
  954. #define SGEMM_DEFAULT_UNROLL_N 4
  955. #define DGEMM_DEFAULT_UNROLL_N 4
  956. #define QGEMM_DEFAULT_UNROLL_N 2
  957. #define CGEMM_DEFAULT_UNROLL_N 2
  958. #define ZGEMM_DEFAULT_UNROLL_N 2
  959. #define XGEMM_DEFAULT_UNROLL_N 1
  960. #endif
  961. #define SGEMM_DEFAULT_P sgemm_p
  962. #define SGEMM_DEFAULT_R sgemm_r
  963. #define DGEMM_DEFAULT_P dgemm_p
  964. #define DGEMM_DEFAULT_R dgemm_r
  965. #define QGEMM_DEFAULT_P qgemm_p
  966. #define QGEMM_DEFAULT_R qgemm_r
  967. #define CGEMM_DEFAULT_P cgemm_p
  968. #define CGEMM_DEFAULT_R cgemm_r
  969. #define ZGEMM_DEFAULT_P zgemm_p
  970. #define ZGEMM_DEFAULT_R zgemm_r
  971. #define XGEMM_DEFAULT_P xgemm_p
  972. #define XGEMM_DEFAULT_R xgemm_r
  973. #define SGEMM_DEFAULT_Q 256
  974. #define DGEMM_DEFAULT_Q 256
  975. #define QGEMM_DEFAULT_Q 256
  976. #define CGEMM_DEFAULT_Q 256
  977. #define ZGEMM_DEFAULT_Q 256
  978. #define XGEMM_DEFAULT_Q 256
  979. #endif
  980. #ifdef PENRYN
  981. #define SNUMOPT 8
  982. #define DNUMOPT 4
  983. #define GEMM_DEFAULT_OFFSET_A 128
  984. #define GEMM_DEFAULT_OFFSET_B 0
  985. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  986. #define SYMV_P 8
  987. #define SWITCH_RATIO 4
  988. #ifdef ARCH_X86
  989. #define SGEMM_DEFAULT_UNROLL_M 4
  990. #define DGEMM_DEFAULT_UNROLL_M 2
  991. #define QGEMM_DEFAULT_UNROLL_M 2
  992. #define CGEMM_DEFAULT_UNROLL_M 2
  993. #define ZGEMM_DEFAULT_UNROLL_M 1
  994. #define XGEMM_DEFAULT_UNROLL_M 1
  995. #define SGEMM_DEFAULT_UNROLL_N 4
  996. #define DGEMM_DEFAULT_UNROLL_N 4
  997. #define QGEMM_DEFAULT_UNROLL_N 2
  998. #define CGEMM_DEFAULT_UNROLL_N 2
  999. #define ZGEMM_DEFAULT_UNROLL_N 2
  1000. #define XGEMM_DEFAULT_UNROLL_N 1
  1001. #else
  1002. #define SGEMM_DEFAULT_UNROLL_M 8
  1003. #define DGEMM_DEFAULT_UNROLL_M 4
  1004. #define QGEMM_DEFAULT_UNROLL_M 2
  1005. #define CGEMM_DEFAULT_UNROLL_M 4
  1006. #define ZGEMM_DEFAULT_UNROLL_M 2
  1007. #define XGEMM_DEFAULT_UNROLL_M 1
  1008. #define SGEMM_DEFAULT_UNROLL_N 4
  1009. #define DGEMM_DEFAULT_UNROLL_N 4
  1010. #define QGEMM_DEFAULT_UNROLL_N 2
  1011. #define CGEMM_DEFAULT_UNROLL_N 2
  1012. #define ZGEMM_DEFAULT_UNROLL_N 2
  1013. #define XGEMM_DEFAULT_UNROLL_N 1
  1014. #endif
  1015. #define SGEMM_DEFAULT_P sgemm_p
  1016. #define SGEMM_DEFAULT_R sgemm_r
  1017. #define DGEMM_DEFAULT_P dgemm_p
  1018. #define DGEMM_DEFAULT_R dgemm_r
  1019. #define QGEMM_DEFAULT_P qgemm_p
  1020. #define QGEMM_DEFAULT_R qgemm_r
  1021. #define CGEMM_DEFAULT_P cgemm_p
  1022. #define CGEMM_DEFAULT_R cgemm_r
  1023. #define ZGEMM_DEFAULT_P zgemm_p
  1024. #define ZGEMM_DEFAULT_R zgemm_r
  1025. #define XGEMM_DEFAULT_P xgemm_p
  1026. #define XGEMM_DEFAULT_R xgemm_r
  1027. #define SGEMM_DEFAULT_Q 512
  1028. #define DGEMM_DEFAULT_Q 256
  1029. #define QGEMM_DEFAULT_Q 128
  1030. #define CGEMM_DEFAULT_Q 512
  1031. #define ZGEMM_DEFAULT_Q 256
  1032. #define XGEMM_DEFAULT_Q 128
  1033. #define GETRF_FACTOR 0.75
  1034. #endif
  1035. #ifdef DUNNINGTON
  1036. #define SNUMOPT 8
  1037. #define DNUMOPT 4
  1038. #define GEMM_DEFAULT_OFFSET_A 128
  1039. #define GEMM_DEFAULT_OFFSET_B 0
  1040. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1041. #define SYMV_P 8
  1042. #define SWITCH_RATIO 4
  1043. #ifdef ARCH_X86
  1044. #define SGEMM_DEFAULT_UNROLL_M 4
  1045. #define DGEMM_DEFAULT_UNROLL_M 2
  1046. #define QGEMM_DEFAULT_UNROLL_M 2
  1047. #define CGEMM_DEFAULT_UNROLL_M 2
  1048. #define ZGEMM_DEFAULT_UNROLL_M 1
  1049. #define XGEMM_DEFAULT_UNROLL_M 1
  1050. #define SGEMM_DEFAULT_UNROLL_N 4
  1051. #define DGEMM_DEFAULT_UNROLL_N 4
  1052. #define QGEMM_DEFAULT_UNROLL_N 2
  1053. #define CGEMM_DEFAULT_UNROLL_N 2
  1054. #define ZGEMM_DEFAULT_UNROLL_N 2
  1055. #define XGEMM_DEFAULT_UNROLL_N 1
  1056. #else
  1057. #define SGEMM_DEFAULT_UNROLL_M 8
  1058. #define DGEMM_DEFAULT_UNROLL_M 4
  1059. #define QGEMM_DEFAULT_UNROLL_M 2
  1060. #define CGEMM_DEFAULT_UNROLL_M 4
  1061. #define ZGEMM_DEFAULT_UNROLL_M 2
  1062. #define XGEMM_DEFAULT_UNROLL_M 1
  1063. #define SGEMM_DEFAULT_UNROLL_N 4
  1064. #define DGEMM_DEFAULT_UNROLL_N 4
  1065. #define QGEMM_DEFAULT_UNROLL_N 2
  1066. #define CGEMM_DEFAULT_UNROLL_N 2
  1067. #define ZGEMM_DEFAULT_UNROLL_N 2
  1068. #define XGEMM_DEFAULT_UNROLL_N 1
  1069. #endif
  1070. #define SGEMM_DEFAULT_P sgemm_p
  1071. #define SGEMM_DEFAULT_R sgemm_r
  1072. #define DGEMM_DEFAULT_P dgemm_p
  1073. #define DGEMM_DEFAULT_R dgemm_r
  1074. #define QGEMM_DEFAULT_P qgemm_p
  1075. #define QGEMM_DEFAULT_R qgemm_r
  1076. #define CGEMM_DEFAULT_P cgemm_p
  1077. #define CGEMM_DEFAULT_R cgemm_r
  1078. #define ZGEMM_DEFAULT_P zgemm_p
  1079. #define ZGEMM_DEFAULT_R zgemm_r
  1080. #define XGEMM_DEFAULT_P xgemm_p
  1081. #define XGEMM_DEFAULT_R xgemm_r
  1082. #define SGEMM_DEFAULT_Q 768
  1083. #define DGEMM_DEFAULT_Q 384
  1084. #define QGEMM_DEFAULT_Q 192
  1085. #define CGEMM_DEFAULT_Q 768
  1086. #define ZGEMM_DEFAULT_Q 384
  1087. #define XGEMM_DEFAULT_Q 192
  1088. #define GETRF_FACTOR 0.75
  1089. #define GEMM_THREAD gemm_thread_mn
  1090. #endif
  1091. #ifdef NEHALEM
  1092. #define SNUMOPT 8
  1093. #define DNUMOPT 4
  1094. #define GEMM_DEFAULT_OFFSET_A 32
  1095. #define GEMM_DEFAULT_OFFSET_B 0
  1096. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1097. #define SYMV_P 8
  1098. #define SWITCH_RATIO 4
  1099. #ifdef ARCH_X86
  1100. #define SGEMM_DEFAULT_UNROLL_M 4
  1101. #define DGEMM_DEFAULT_UNROLL_M 2
  1102. #define QGEMM_DEFAULT_UNROLL_M 2
  1103. #define CGEMM_DEFAULT_UNROLL_M 2
  1104. #define ZGEMM_DEFAULT_UNROLL_M 1
  1105. #define XGEMM_DEFAULT_UNROLL_M 1
  1106. #define SGEMM_DEFAULT_UNROLL_N 4
  1107. #define DGEMM_DEFAULT_UNROLL_N 4
  1108. #define QGEMM_DEFAULT_UNROLL_N 2
  1109. #define CGEMM_DEFAULT_UNROLL_N 2
  1110. #define ZGEMM_DEFAULT_UNROLL_N 2
  1111. #define XGEMM_DEFAULT_UNROLL_N 1
  1112. #else
  1113. #define SGEMM_DEFAULT_UNROLL_M 4
  1114. #define DGEMM_DEFAULT_UNROLL_M 2
  1115. #define QGEMM_DEFAULT_UNROLL_M 2
  1116. #define CGEMM_DEFAULT_UNROLL_M 2
  1117. #define ZGEMM_DEFAULT_UNROLL_M 1
  1118. #define XGEMM_DEFAULT_UNROLL_M 1
  1119. #define SGEMM_DEFAULT_UNROLL_N 8
  1120. #define DGEMM_DEFAULT_UNROLL_N 8
  1121. #define QGEMM_DEFAULT_UNROLL_N 2
  1122. #define CGEMM_DEFAULT_UNROLL_N 4
  1123. #define ZGEMM_DEFAULT_UNROLL_N 4
  1124. #define XGEMM_DEFAULT_UNROLL_N 1
  1125. #endif
  1126. #define SGEMM_DEFAULT_P 504
  1127. #define SGEMM_DEFAULT_R sgemm_r
  1128. #define DGEMM_DEFAULT_P 504
  1129. #define DGEMM_DEFAULT_R dgemm_r
  1130. #define QGEMM_DEFAULT_P 504
  1131. #define QGEMM_DEFAULT_R qgemm_r
  1132. #define CGEMM_DEFAULT_P 252
  1133. #define CGEMM_DEFAULT_R cgemm_r
  1134. #define ZGEMM_DEFAULT_P 252
  1135. #define ZGEMM_DEFAULT_R zgemm_r
  1136. #define XGEMM_DEFAULT_P 252
  1137. #define XGEMM_DEFAULT_R xgemm_r
  1138. #define SGEMM_DEFAULT_Q 512
  1139. #define DGEMM_DEFAULT_Q 256
  1140. #define QGEMM_DEFAULT_Q 128
  1141. #define CGEMM_DEFAULT_Q 512
  1142. #define ZGEMM_DEFAULT_Q 256
  1143. #define XGEMM_DEFAULT_Q 128
  1144. #define GETRF_FACTOR 0.72
  1145. #endif
  1146. #ifdef SANDYBRIDGE
  1147. #define SNUMOPT 8
  1148. #define DNUMOPT 4
  1149. #define GEMM_DEFAULT_OFFSET_A 0
  1150. #define GEMM_DEFAULT_OFFSET_B 0
  1151. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1152. #define SYMV_P 8
  1153. #define SWITCH_RATIO 4
  1154. #ifdef ARCH_X86
  1155. #define SGEMM_DEFAULT_UNROLL_M 4
  1156. #define DGEMM_DEFAULT_UNROLL_M 2
  1157. #define QGEMM_DEFAULT_UNROLL_M 2
  1158. #define CGEMM_DEFAULT_UNROLL_M 2
  1159. #define ZGEMM_DEFAULT_UNROLL_M 1
  1160. #define XGEMM_DEFAULT_UNROLL_M 1
  1161. #define SGEMM_DEFAULT_UNROLL_N 4
  1162. #define DGEMM_DEFAULT_UNROLL_N 4
  1163. #define QGEMM_DEFAULT_UNROLL_N 2
  1164. #define CGEMM_DEFAULT_UNROLL_N 2
  1165. #define ZGEMM_DEFAULT_UNROLL_N 2
  1166. #define XGEMM_DEFAULT_UNROLL_N 1
  1167. #else
  1168. #define SGEMM_DEFAULT_UNROLL_M 16
  1169. #define DGEMM_DEFAULT_UNROLL_M 8
  1170. #define QGEMM_DEFAULT_UNROLL_M 2
  1171. #define CGEMM_DEFAULT_UNROLL_M 8
  1172. #define ZGEMM_DEFAULT_UNROLL_M 1
  1173. #define XGEMM_DEFAULT_UNROLL_M 1
  1174. #define SGEMM_DEFAULT_UNROLL_N 4
  1175. #define DGEMM_DEFAULT_UNROLL_N 4
  1176. #define QGEMM_DEFAULT_UNROLL_N 2
  1177. #define CGEMM_DEFAULT_UNROLL_N 2
  1178. #define ZGEMM_DEFAULT_UNROLL_N 4
  1179. #define XGEMM_DEFAULT_UNROLL_N 1
  1180. #endif
  1181. #define SGEMM_DEFAULT_P 768
  1182. #define SGEMM_DEFAULT_R sgemm_r
  1183. /*#define SGEMM_DEFAULT_R 1024*/
  1184. #define DGEMM_DEFAULT_P 512
  1185. #define DGEMM_DEFAULT_R dgemm_r
  1186. /*#define DGEMM_DEFAULT_R 1024*/
  1187. #define QGEMM_DEFAULT_P 504
  1188. #define QGEMM_DEFAULT_R qgemm_r
  1189. #define CGEMM_DEFAULT_P 768
  1190. #define CGEMM_DEFAULT_R cgemm_r
  1191. /*#define CGEMM_DEFAULT_R 1024*/
  1192. #define ZGEMM_DEFAULT_P 512
  1193. #define ZGEMM_DEFAULT_R zgemm_r
  1194. /*#define ZGEMM_DEFAULT_R 1024*/
  1195. #define XGEMM_DEFAULT_P 252
  1196. #define XGEMM_DEFAULT_R xgemm_r
  1197. #define SGEMM_DEFAULT_Q 384
  1198. #define DGEMM_DEFAULT_Q 256
  1199. #define QGEMM_DEFAULT_Q 128
  1200. #define CGEMM_DEFAULT_Q 512
  1201. #define ZGEMM_DEFAULT_Q 192
  1202. #define XGEMM_DEFAULT_Q 128
  1203. #define CGEMM3M_DEFAULT_UNROLL_N 8
  1204. #define CGEMM3M_DEFAULT_UNROLL_M 4
  1205. #define ZGEMM3M_DEFAULT_UNROLL_N 8
  1206. #define ZGEMM3M_DEFAULT_UNROLL_M 2
  1207. #define CGEMM3M_DEFAULT_P 448
  1208. #define ZGEMM3M_DEFAULT_P 224
  1209. #define XGEMM3M_DEFAULT_P 112
  1210. #define CGEMM3M_DEFAULT_Q 224
  1211. #define ZGEMM3M_DEFAULT_Q 224
  1212. #define XGEMM3M_DEFAULT_Q 224
  1213. #define CGEMM3M_DEFAULT_R 12288
  1214. #define ZGEMM3M_DEFAULT_R 12288
  1215. #define XGEMM3M_DEFAULT_R 12288
  1216. #define GETRF_FACTOR 0.72
  1217. #endif
  1218. #ifdef HASWELL
  1219. #define SNUMOPT 16
  1220. #define DNUMOPT 8
  1221. #define GEMM_DEFAULT_OFFSET_A 0
  1222. #define GEMM_DEFAULT_OFFSET_B 0
  1223. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1224. #define SYMV_P 8
  1225. #if defined(XDOUBLE) || defined(DOUBLE)
  1226. #define SWITCH_RATIO 4
  1227. #define GEMM_PREFERED_SIZE 4
  1228. #else
  1229. #define SWITCH_RATIO 8
  1230. #define GEMM_PREFERED_SIZE 8
  1231. #endif
  1232. #ifdef ARCH_X86
  1233. #define SGEMM_DEFAULT_UNROLL_M 4
  1234. #define DGEMM_DEFAULT_UNROLL_M 2
  1235. #define QGEMM_DEFAULT_UNROLL_M 2
  1236. #define CGEMM_DEFAULT_UNROLL_M 2
  1237. #define ZGEMM_DEFAULT_UNROLL_M 1
  1238. #define XGEMM_DEFAULT_UNROLL_M 1
  1239. #define SGEMM_DEFAULT_UNROLL_N 4
  1240. #define DGEMM_DEFAULT_UNROLL_N 4
  1241. #define QGEMM_DEFAULT_UNROLL_N 2
  1242. #define CGEMM_DEFAULT_UNROLL_N 2
  1243. #define ZGEMM_DEFAULT_UNROLL_N 2
  1244. #define XGEMM_DEFAULT_UNROLL_N 1
  1245. #else
  1246. #define SGEMM_DEFAULT_UNROLL_M 8
  1247. #define DGEMM_DEFAULT_UNROLL_M 4
  1248. #define QGEMM_DEFAULT_UNROLL_M 2
  1249. #define CGEMM_DEFAULT_UNROLL_M 8
  1250. #define ZGEMM_DEFAULT_UNROLL_M 4
  1251. #define XGEMM_DEFAULT_UNROLL_M 1
  1252. #define SGEMM_DEFAULT_UNROLL_N 4
  1253. #define DGEMM_DEFAULT_UNROLL_N 8
  1254. #define QGEMM_DEFAULT_UNROLL_N 2
  1255. #define CGEMM_DEFAULT_UNROLL_N 2
  1256. #define ZGEMM_DEFAULT_UNROLL_N 2
  1257. #define XGEMM_DEFAULT_UNROLL_N 1
  1258. /*
  1259. #define SGEMM_DEFAULT_UNROLL_MN 32
  1260. #define DGEMM_DEFAULT_UNROLL_MN 32
  1261. */
  1262. #endif
  1263. #ifdef ARCH_X86
  1264. #define SGEMM_DEFAULT_P 512
  1265. #define SGEMM_DEFAULT_R sgemm_r
  1266. #define DGEMM_DEFAULT_P 512
  1267. #define DGEMM_DEFAULT_R dgemm_r
  1268. #define QGEMM_DEFAULT_P 504
  1269. #define QGEMM_DEFAULT_R qgemm_r
  1270. #define CGEMM_DEFAULT_P 128
  1271. #define CGEMM_DEFAULT_R 1024
  1272. #define ZGEMM_DEFAULT_P 512
  1273. #define ZGEMM_DEFAULT_R zgemm_r
  1274. #define XGEMM_DEFAULT_P 252
  1275. #define XGEMM_DEFAULT_R xgemm_r
  1276. #define SGEMM_DEFAULT_Q 256
  1277. #define DGEMM_DEFAULT_Q 256
  1278. #define QGEMM_DEFAULT_Q 128
  1279. #define CGEMM_DEFAULT_Q 256
  1280. #define ZGEMM_DEFAULT_Q 192
  1281. #define XGEMM_DEFAULT_Q 128
  1282. #else
  1283. #define SGEMM_DEFAULT_P 320
  1284. #define DGEMM_DEFAULT_P 512
  1285. #define CGEMM_DEFAULT_P 256
  1286. #define ZGEMM_DEFAULT_P 192
  1287. #ifdef WINDOWS_ABI
  1288. #define SGEMM_DEFAULT_Q 320
  1289. #define DGEMM_DEFAULT_Q 128
  1290. #else
  1291. #define SGEMM_DEFAULT_Q 320
  1292. #define DGEMM_DEFAULT_Q 256
  1293. #endif
  1294. #define CGEMM_DEFAULT_Q 256
  1295. #define ZGEMM_DEFAULT_Q 192
  1296. #define SGEMM_DEFAULT_R sgemm_r
  1297. #define DGEMM_DEFAULT_R 13824
  1298. #define CGEMM_DEFAULT_R cgemm_r
  1299. #define ZGEMM_DEFAULT_R zgemm_r
  1300. #define QGEMM_DEFAULT_Q 128
  1301. #define QGEMM_DEFAULT_P 504
  1302. #define QGEMM_DEFAULT_R qgemm_r
  1303. #define XGEMM_DEFAULT_P 252
  1304. #define XGEMM_DEFAULT_R xgemm_r
  1305. #define XGEMM_DEFAULT_Q 128
  1306. #define CGEMM3M_DEFAULT_UNROLL_N 4
  1307. #define CGEMM3M_DEFAULT_UNROLL_M 8
  1308. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  1309. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  1310. #define CGEMM3M_DEFAULT_P 320
  1311. #define ZGEMM3M_DEFAULT_P 256
  1312. #define XGEMM3M_DEFAULT_P 112
  1313. #define CGEMM3M_DEFAULT_Q 320
  1314. #define ZGEMM3M_DEFAULT_Q 256
  1315. #define XGEMM3M_DEFAULT_Q 224
  1316. #define CGEMM3M_DEFAULT_R 12288
  1317. #define ZGEMM3M_DEFAULT_R 12288
  1318. #define XGEMM3M_DEFAULT_R 12288
  1319. #endif
  1320. #endif
  1321. #ifdef SKYLAKEX
  1322. #define SNUMOPT 16
  1323. #define DNUMOPT 8
  1324. #define GEMM_DEFAULT_OFFSET_A 0
  1325. #define GEMM_DEFAULT_OFFSET_B 0
  1326. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1327. #define SYMV_P 8
  1328. #if defined(XDOUBLE) || defined(DOUBLE)
  1329. #define SWITCH_RATIO 8
  1330. #define GEMM_PREFERED_SIZE 8
  1331. #else
  1332. #define SWITCH_RATIO 16
  1333. #define GEMM_PREFERED_SIZE 16
  1334. #endif
  1335. #define USE_SGEMM_KERNEL_DIRECT 1
  1336. #ifdef ARCH_X86
  1337. #define SGEMM_DEFAULT_UNROLL_M 4
  1338. #define DGEMM_DEFAULT_UNROLL_M 2
  1339. #define QGEMM_DEFAULT_UNROLL_M 2
  1340. #define CGEMM_DEFAULT_UNROLL_M 2
  1341. #define ZGEMM_DEFAULT_UNROLL_M 1
  1342. #define XGEMM_DEFAULT_UNROLL_M 1
  1343. #define SGEMM_DEFAULT_UNROLL_N 4
  1344. #define DGEMM_DEFAULT_UNROLL_N 4
  1345. #define QGEMM_DEFAULT_UNROLL_N 2
  1346. #define CGEMM_DEFAULT_UNROLL_N 2
  1347. #define ZGEMM_DEFAULT_UNROLL_N 2
  1348. #define XGEMM_DEFAULT_UNROLL_N 1
  1349. #else
  1350. #define SGEMM_DEFAULT_UNROLL_M 16
  1351. #define DGEMM_DEFAULT_UNROLL_M 16
  1352. #define QGEMM_DEFAULT_UNROLL_M 2
  1353. #define CGEMM_DEFAULT_UNROLL_M 8
  1354. #define ZGEMM_DEFAULT_UNROLL_M 4
  1355. #define XGEMM_DEFAULT_UNROLL_M 1
  1356. #define SGEMM_DEFAULT_UNROLL_N 4
  1357. #define DGEMM_DEFAULT_UNROLL_N 2
  1358. #define QGEMM_DEFAULT_UNROLL_N 2
  1359. #define CGEMM_DEFAULT_UNROLL_N 2
  1360. #define ZGEMM_DEFAULT_UNROLL_N 2
  1361. #define XGEMM_DEFAULT_UNROLL_N 1
  1362. #define SGEMM_DEFAULT_UNROLL_MN 32
  1363. #define DGEMM_DEFAULT_UNROLL_MN 32
  1364. #endif
  1365. #ifdef ARCH_X86
  1366. #define SGEMM_DEFAULT_P 512
  1367. #define SGEMM_DEFAULT_R sgemm_r
  1368. #define DGEMM_DEFAULT_P 512
  1369. #define DGEMM_DEFAULT_R dgemm_r
  1370. #define QGEMM_DEFAULT_P 504
  1371. #define QGEMM_DEFAULT_R qgemm_r
  1372. #define CGEMM_DEFAULT_P 128
  1373. #define CGEMM_DEFAULT_R 1024
  1374. #define ZGEMM_DEFAULT_P 512
  1375. #define ZGEMM_DEFAULT_R zgemm_r
  1376. #define XGEMM_DEFAULT_P 252
  1377. #define XGEMM_DEFAULT_R xgemm_r
  1378. #define SGEMM_DEFAULT_Q 256
  1379. #define DGEMM_DEFAULT_Q 256
  1380. #define QGEMM_DEFAULT_Q 128
  1381. #define CGEMM_DEFAULT_Q 256
  1382. #define ZGEMM_DEFAULT_Q 192
  1383. #define XGEMM_DEFAULT_Q 128
  1384. #else
  1385. #define SGEMM_DEFAULT_P 448
  1386. #define DGEMM_DEFAULT_P 192
  1387. #define CGEMM_DEFAULT_P 384
  1388. #define ZGEMM_DEFAULT_P 256
  1389. #define SGEMM_DEFAULT_Q 448
  1390. #define DGEMM_DEFAULT_Q 384
  1391. #define CGEMM_DEFAULT_Q 192
  1392. #define ZGEMM_DEFAULT_Q 128
  1393. #define SGEMM_DEFAULT_R sgemm_r
  1394. #define DGEMM_DEFAULT_R 8640
  1395. #define CGEMM_DEFAULT_R cgemm_r
  1396. #define ZGEMM_DEFAULT_R zgemm_r
  1397. #define QGEMM_DEFAULT_Q 128
  1398. #define QGEMM_DEFAULT_P 504
  1399. #define QGEMM_DEFAULT_R qgemm_r
  1400. #define XGEMM_DEFAULT_P 252
  1401. #define XGEMM_DEFAULT_R xgemm_r
  1402. #define XGEMM_DEFAULT_Q 128
  1403. #define CGEMM3M_DEFAULT_UNROLL_N 4
  1404. #define CGEMM3M_DEFAULT_UNROLL_M 8
  1405. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  1406. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  1407. #define CGEMM3M_DEFAULT_P 320
  1408. #define ZGEMM3M_DEFAULT_P 256
  1409. #define XGEMM3M_DEFAULT_P 112
  1410. #define CGEMM3M_DEFAULT_Q 320
  1411. #define ZGEMM3M_DEFAULT_Q 256
  1412. #define XGEMM3M_DEFAULT_Q 224
  1413. #define CGEMM3M_DEFAULT_R 12288
  1414. #define ZGEMM3M_DEFAULT_R 12288
  1415. #define XGEMM3M_DEFAULT_R 12288
  1416. #endif
  1417. #endif
  1418. #ifdef COOPERLAKE
  1419. #define SNUMOPT 16
  1420. #define DNUMOPT 8
  1421. #define GEMM_DEFAULT_OFFSET_A 0
  1422. #define GEMM_DEFAULT_OFFSET_B 0
  1423. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1424. #define SYMV_P 8
  1425. #if defined(XDOUBLE) || defined(DOUBLE)
  1426. #define SWITCH_RATIO 8
  1427. #define GEMM_PREFERED_SIZE 8
  1428. #else
  1429. #define SWITCH_RATIO 16
  1430. #define GEMM_PREFERED_SIZE 16
  1431. #endif
  1432. #define USE_SGEMM_KERNEL_DIRECT 1
  1433. #ifdef ARCH_X86
  1434. #define SGEMM_DEFAULT_UNROLL_M 4
  1435. #define DGEMM_DEFAULT_UNROLL_M 2
  1436. #define QGEMM_DEFAULT_UNROLL_M 2
  1437. #define CGEMM_DEFAULT_UNROLL_M 2
  1438. #define ZGEMM_DEFAULT_UNROLL_M 1
  1439. #define XGEMM_DEFAULT_UNROLL_M 1
  1440. #define SGEMM_DEFAULT_UNROLL_N 4
  1441. #define DGEMM_DEFAULT_UNROLL_N 4
  1442. #define QGEMM_DEFAULT_UNROLL_N 2
  1443. #define CGEMM_DEFAULT_UNROLL_N 2
  1444. #define ZGEMM_DEFAULT_UNROLL_N 2
  1445. #define XGEMM_DEFAULT_UNROLL_N 1
  1446. #else
  1447. #define SGEMM_DEFAULT_UNROLL_M 16
  1448. #define DGEMM_DEFAULT_UNROLL_M 16
  1449. #define QGEMM_DEFAULT_UNROLL_M 2
  1450. #define CGEMM_DEFAULT_UNROLL_M 8
  1451. #define ZGEMM_DEFAULT_UNROLL_M 4
  1452. #define XGEMM_DEFAULT_UNROLL_M 1
  1453. #define SGEMM_DEFAULT_UNROLL_N 4
  1454. #define DGEMM_DEFAULT_UNROLL_N 2
  1455. #define QGEMM_DEFAULT_UNROLL_N 2
  1456. #define CGEMM_DEFAULT_UNROLL_N 2
  1457. #define ZGEMM_DEFAULT_UNROLL_N 2
  1458. #define XGEMM_DEFAULT_UNROLL_N 1
  1459. #define SGEMM_DEFAULT_UNROLL_MN 32
  1460. #define DGEMM_DEFAULT_UNROLL_MN 32
  1461. #endif
  1462. #ifdef ARCH_X86
  1463. #define SGEMM_DEFAULT_P 512
  1464. #define SGEMM_DEFAULT_R sgemm_r
  1465. #define DGEMM_DEFAULT_P 512
  1466. #define DGEMM_DEFAULT_R dgemm_r
  1467. #define QGEMM_DEFAULT_P 504
  1468. #define QGEMM_DEFAULT_R qgemm_r
  1469. #define CGEMM_DEFAULT_P 128
  1470. #define CGEMM_DEFAULT_R 1024
  1471. #define ZGEMM_DEFAULT_P 512
  1472. #define ZGEMM_DEFAULT_R zgemm_r
  1473. #define XGEMM_DEFAULT_P 252
  1474. #define XGEMM_DEFAULT_R xgemm_r
  1475. #define SGEMM_DEFAULT_Q 256
  1476. #define DGEMM_DEFAULT_Q 256
  1477. #define QGEMM_DEFAULT_Q 128
  1478. #define CGEMM_DEFAULT_Q 256
  1479. #define ZGEMM_DEFAULT_Q 192
  1480. #define XGEMM_DEFAULT_Q 128
  1481. #else
  1482. #define SGEMM_DEFAULT_P 640
  1483. #define DGEMM_DEFAULT_P 192
  1484. #define CGEMM_DEFAULT_P 384
  1485. #define ZGEMM_DEFAULT_P 256
  1486. #define SGEMM_DEFAULT_Q 320
  1487. #define DGEMM_DEFAULT_Q 384
  1488. #define CGEMM_DEFAULT_Q 192
  1489. #define ZGEMM_DEFAULT_Q 128
  1490. #define SGEMM_DEFAULT_R sgemm_r
  1491. #define DGEMM_DEFAULT_R 8640
  1492. #define CGEMM_DEFAULT_R cgemm_r
  1493. #define ZGEMM_DEFAULT_R zgemm_r
  1494. #define QGEMM_DEFAULT_Q 128
  1495. #define QGEMM_DEFAULT_P 504
  1496. #define QGEMM_DEFAULT_R qgemm_r
  1497. #define XGEMM_DEFAULT_P 252
  1498. #define XGEMM_DEFAULT_R xgemm_r
  1499. #define XGEMM_DEFAULT_Q 128
  1500. #define CGEMM3M_DEFAULT_UNROLL_N 4
  1501. #define CGEMM3M_DEFAULT_UNROLL_M 8
  1502. #define ZGEMM3M_DEFAULT_UNROLL_N 4
  1503. #define ZGEMM3M_DEFAULT_UNROLL_M 4
  1504. #define CGEMM3M_DEFAULT_P 320
  1505. #define ZGEMM3M_DEFAULT_P 256
  1506. #define XGEMM3M_DEFAULT_P 112
  1507. #define CGEMM3M_DEFAULT_Q 320
  1508. #define ZGEMM3M_DEFAULT_Q 256
  1509. #define XGEMM3M_DEFAULT_Q 224
  1510. #define CGEMM3M_DEFAULT_R 12288
  1511. #define ZGEMM3M_DEFAULT_R 12288
  1512. #define XGEMM3M_DEFAULT_R 12288
  1513. #endif
  1514. #endif
  1515. #ifdef ATOM
  1516. #define SNUMOPT 2
  1517. #define DNUMOPT 1
  1518. #define GEMM_DEFAULT_OFFSET_A 64
  1519. #define GEMM_DEFAULT_OFFSET_B 0
  1520. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1521. #define SYMV_P 8
  1522. #ifdef ARCH_X86
  1523. #define SGEMM_DEFAULT_UNROLL_M 4
  1524. #define DGEMM_DEFAULT_UNROLL_M 2
  1525. #define QGEMM_DEFAULT_UNROLL_M 2
  1526. #define CGEMM_DEFAULT_UNROLL_M 2
  1527. #define ZGEMM_DEFAULT_UNROLL_M 1
  1528. #define XGEMM_DEFAULT_UNROLL_M 1
  1529. #else
  1530. #define SGEMM_DEFAULT_UNROLL_M 8
  1531. #define DGEMM_DEFAULT_UNROLL_M 4
  1532. #define QGEMM_DEFAULT_UNROLL_M 2
  1533. #define CGEMM_DEFAULT_UNROLL_M 4
  1534. #define ZGEMM_DEFAULT_UNROLL_M 2
  1535. #define XGEMM_DEFAULT_UNROLL_M 1
  1536. #endif
  1537. #define SGEMM_DEFAULT_UNROLL_N 4
  1538. #define DGEMM_DEFAULT_UNROLL_N 2
  1539. #define QGEMM_DEFAULT_UNROLL_N 2
  1540. #define CGEMM_DEFAULT_UNROLL_N 2
  1541. #define ZGEMM_DEFAULT_UNROLL_N 1
  1542. #define XGEMM_DEFAULT_UNROLL_N 1
  1543. #define SGEMM_DEFAULT_P sgemm_p
  1544. #define SGEMM_DEFAULT_R sgemm_r
  1545. #define DGEMM_DEFAULT_P dgemm_p
  1546. #define DGEMM_DEFAULT_R dgemm_r
  1547. #define QGEMM_DEFAULT_P qgemm_p
  1548. #define QGEMM_DEFAULT_R qgemm_r
  1549. #define CGEMM_DEFAULT_P cgemm_p
  1550. #define CGEMM_DEFAULT_R cgemm_r
  1551. #define ZGEMM_DEFAULT_P zgemm_p
  1552. #define ZGEMM_DEFAULT_R zgemm_r
  1553. #define XGEMM_DEFAULT_P xgemm_p
  1554. #define XGEMM_DEFAULT_R xgemm_r
  1555. #define SGEMM_DEFAULT_Q 256
  1556. #define DGEMM_DEFAULT_Q 256
  1557. #define QGEMM_DEFAULT_Q 256
  1558. #define CGEMM_DEFAULT_Q 256
  1559. #define ZGEMM_DEFAULT_Q 256
  1560. #define XGEMM_DEFAULT_Q 256
  1561. #endif
  1562. #ifdef ITANIUM2
  1563. #define SNUMOPT 4
  1564. #define DNUMOPT 4
  1565. #define GEMM_DEFAULT_OFFSET_A 0
  1566. #define GEMM_DEFAULT_OFFSET_B 128
  1567. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1568. #define SGEMM_DEFAULT_UNROLL_M 8
  1569. #define SGEMM_DEFAULT_UNROLL_N 8
  1570. #define DGEMM_DEFAULT_UNROLL_M 8
  1571. #define DGEMM_DEFAULT_UNROLL_N 8
  1572. #define QGEMM_DEFAULT_UNROLL_M 8
  1573. #define QGEMM_DEFAULT_UNROLL_N 8
  1574. #define CGEMM_DEFAULT_UNROLL_M 4
  1575. #define CGEMM_DEFAULT_UNROLL_N 4
  1576. #define ZGEMM_DEFAULT_UNROLL_M 4
  1577. #define ZGEMM_DEFAULT_UNROLL_N 4
  1578. #define XGEMM_DEFAULT_UNROLL_M 4
  1579. #define XGEMM_DEFAULT_UNROLL_N 4
  1580. #define SGEMM_DEFAULT_P sgemm_p
  1581. #define DGEMM_DEFAULT_P dgemm_p
  1582. #define QGEMM_DEFAULT_P qgemm_p
  1583. #define CGEMM_DEFAULT_P cgemm_p
  1584. #define ZGEMM_DEFAULT_P zgemm_p
  1585. #define XGEMM_DEFAULT_P xgemm_p
  1586. #define SGEMM_DEFAULT_Q 1024
  1587. #define DGEMM_DEFAULT_Q 1024
  1588. #define QGEMM_DEFAULT_Q 1024
  1589. #define CGEMM_DEFAULT_Q 1024
  1590. #define ZGEMM_DEFAULT_Q 1024
  1591. #define XGEMM_DEFAULT_Q 1024
  1592. #define SGEMM_DEFAULT_R sgemm_r
  1593. #define DGEMM_DEFAULT_R dgemm_r
  1594. #define QGEMM_DEFAULT_R qgemm_r
  1595. #define CGEMM_DEFAULT_R cgemm_r
  1596. #define ZGEMM_DEFAULT_R zgemm_r
  1597. #define XGEMM_DEFAULT_R xgemm_r
  1598. #define SYMV_P 16
  1599. #define GETRF_FACTOR 0.65
  1600. #endif
  1601. #if defined(EV4) || defined(EV5) || defined(EV6)
  1602. #ifdef EV4
  1603. #define SNUMOPT 1
  1604. #define DNUMOPT 1
  1605. #else
  1606. #define SNUMOPT 2
  1607. #define DNUMOPT 2
  1608. #endif
  1609. #define GEMM_DEFAULT_OFFSET_A 512
  1610. #define GEMM_DEFAULT_OFFSET_B 512
  1611. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1612. #define SGEMM_DEFAULT_UNROLL_M 4
  1613. #define SGEMM_DEFAULT_UNROLL_N 4
  1614. #define DGEMM_DEFAULT_UNROLL_M 4
  1615. #define DGEMM_DEFAULT_UNROLL_N 4
  1616. #define CGEMM_DEFAULT_UNROLL_M 2
  1617. #define CGEMM_DEFAULT_UNROLL_N 2
  1618. #define ZGEMM_DEFAULT_UNROLL_M 2
  1619. #define ZGEMM_DEFAULT_UNROLL_N 2
  1620. #define SYMV_P 8
  1621. #ifdef EV4
  1622. #define SGEMM_DEFAULT_P 32
  1623. #define SGEMM_DEFAULT_Q 112
  1624. #define SGEMM_DEFAULT_R 256
  1625. #define DGEMM_DEFAULT_P 32
  1626. #define DGEMM_DEFAULT_Q 56
  1627. #define DGEMM_DEFAULT_R 256
  1628. #define CGEMM_DEFAULT_P 32
  1629. #define CGEMM_DEFAULT_Q 64
  1630. #define CGEMM_DEFAULT_R 240
  1631. #define ZGEMM_DEFAULT_P 32
  1632. #define ZGEMM_DEFAULT_Q 32
  1633. #define ZGEMM_DEFAULT_R 240
  1634. #endif
  1635. #ifdef EV5
  1636. #define SGEMM_DEFAULT_P 64
  1637. #define SGEMM_DEFAULT_Q 256
  1638. #define DGEMM_DEFAULT_P 64
  1639. #define DGEMM_DEFAULT_Q 128
  1640. #define CGEMM_DEFAULT_P 64
  1641. #define CGEMM_DEFAULT_Q 128
  1642. #define ZGEMM_DEFAULT_P 64
  1643. #define ZGEMM_DEFAULT_Q 64
  1644. #endif
  1645. #ifdef EV6
  1646. #define SGEMM_DEFAULT_P 256
  1647. #define SGEMM_DEFAULT_Q 512
  1648. #define DGEMM_DEFAULT_P 256
  1649. #define DGEMM_DEFAULT_Q 256
  1650. #define CGEMM_DEFAULT_P 256
  1651. #define CGEMM_DEFAULT_Q 256
  1652. #define ZGEMM_DEFAULT_P 128
  1653. #define ZGEMM_DEFAULT_Q 256
  1654. #endif
  1655. #endif
  1656. #ifdef CELL
  1657. #define SNUMOPT 2
  1658. #define DNUMOPT 2
  1659. #define GEMM_DEFAULT_OFFSET_A 0
  1660. #define GEMM_DEFAULT_OFFSET_B 8192
  1661. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1662. #define SGEMM_DEFAULT_UNROLL_M 16
  1663. #define SGEMM_DEFAULT_UNROLL_N 4
  1664. #define DGEMM_DEFAULT_UNROLL_M 4
  1665. #define DGEMM_DEFAULT_UNROLL_N 4
  1666. #define CGEMM_DEFAULT_UNROLL_M 8
  1667. #define CGEMM_DEFAULT_UNROLL_N 2
  1668. #define ZGEMM_DEFAULT_UNROLL_M 2
  1669. #define ZGEMM_DEFAULT_UNROLL_N 2
  1670. #define SGEMM_DEFAULT_P 128
  1671. #define DGEMM_DEFAULT_P 128
  1672. #define CGEMM_DEFAULT_P 128
  1673. #define ZGEMM_DEFAULT_P 128
  1674. #define SGEMM_DEFAULT_Q 512
  1675. #define DGEMM_DEFAULT_Q 256
  1676. #define CGEMM_DEFAULT_Q 256
  1677. #define ZGEMM_DEFAULT_Q 128
  1678. #define SYMV_P 4
  1679. #endif
  1680. #ifdef PPCG4
  1681. #define GEMM_DEFAULT_OFFSET_A 0
  1682. #define GEMM_DEFAULT_OFFSET_B 1024
  1683. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1684. #define SGEMM_DEFAULT_UNROLL_M 16
  1685. #define SGEMM_DEFAULT_UNROLL_N 4
  1686. #define DGEMM_DEFAULT_UNROLL_M 4
  1687. #define DGEMM_DEFAULT_UNROLL_N 4
  1688. #define CGEMM_DEFAULT_UNROLL_M 2
  1689. #define CGEMM_DEFAULT_UNROLL_N 2
  1690. #define ZGEMM_DEFAULT_UNROLL_M 2
  1691. #define ZGEMM_DEFAULT_UNROLL_N 2
  1692. #define SGEMM_DEFAULT_P 256
  1693. #define DGEMM_DEFAULT_P 128
  1694. #define CGEMM_DEFAULT_P 128
  1695. #define ZGEMM_DEFAULT_P 64
  1696. #define SGEMM_DEFAULT_Q 256
  1697. #define DGEMM_DEFAULT_Q 256
  1698. #define CGEMM_DEFAULT_Q 256
  1699. #define ZGEMM_DEFAULT_Q 256
  1700. #define SYMV_P 4
  1701. #endif
  1702. #ifdef PPC970
  1703. #define SNUMOPT 4
  1704. #define DNUMOPT 4
  1705. #define GEMM_DEFAULT_OFFSET_A 2688
  1706. #define GEMM_DEFAULT_OFFSET_B 3072
  1707. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1708. #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
  1709. #define SGEMM_DEFAULT_UNROLL_M 4
  1710. #else
  1711. #define SGEMM_DEFAULT_UNROLL_M 16
  1712. #endif
  1713. #define SGEMM_DEFAULT_UNROLL_N 4
  1714. #define DGEMM_DEFAULT_UNROLL_M 4
  1715. #define DGEMM_DEFAULT_UNROLL_N 4
  1716. #if defined(__BYTE_ORDER__)&&(__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
  1717. #define CGEMM_DEFAULT_UNROLL_M 2
  1718. #else
  1719. #define CGEMM_DEFAULT_UNROLL_M 8
  1720. #endif
  1721. #define CGEMM_DEFAULT_UNROLL_N 2
  1722. #define ZGEMM_DEFAULT_UNROLL_M 2
  1723. #define ZGEMM_DEFAULT_UNROLL_N 2
  1724. #if defined(OS_LINUX) || defined(OS_DARWIN) || defined(OS_FREEBSD)
  1725. #if L2_SIZE == 1024976
  1726. #define SGEMM_DEFAULT_P 320
  1727. #define DGEMM_DEFAULT_P 256
  1728. #define CGEMM_DEFAULT_P 256
  1729. #define ZGEMM_DEFAULT_P 256
  1730. #else
  1731. #define SGEMM_DEFAULT_P 176
  1732. #define DGEMM_DEFAULT_P 176
  1733. #define CGEMM_DEFAULT_P 176
  1734. #define ZGEMM_DEFAULT_P 176
  1735. #endif
  1736. #endif
  1737. #define SGEMM_DEFAULT_Q 512
  1738. #define DGEMM_DEFAULT_Q 256
  1739. #define CGEMM_DEFAULT_Q 256
  1740. #define ZGEMM_DEFAULT_Q 128
  1741. #define SYMV_P 4
  1742. #endif
  1743. #ifdef PPC440
  1744. #define SNUMOPT 2
  1745. #define DNUMOPT 2
  1746. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1747. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1748. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1749. #define SGEMM_DEFAULT_UNROLL_M 4
  1750. #define SGEMM_DEFAULT_UNROLL_N 4
  1751. #define DGEMM_DEFAULT_UNROLL_M 4
  1752. #define DGEMM_DEFAULT_UNROLL_N 4
  1753. #define CGEMM_DEFAULT_UNROLL_M 2
  1754. #define CGEMM_DEFAULT_UNROLL_N 2
  1755. #define ZGEMM_DEFAULT_UNROLL_M 2
  1756. #define ZGEMM_DEFAULT_UNROLL_N 2
  1757. #define SGEMM_DEFAULT_P 512
  1758. #define DGEMM_DEFAULT_P 512
  1759. #define CGEMM_DEFAULT_P 512
  1760. #define ZGEMM_DEFAULT_P 512
  1761. #define SGEMM_DEFAULT_Q 1024
  1762. #define DGEMM_DEFAULT_Q 512
  1763. #define CGEMM_DEFAULT_Q 512
  1764. #define ZGEMM_DEFAULT_Q 256
  1765. #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
  1766. #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
  1767. #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
  1768. #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
  1769. #define SYMV_P 4
  1770. #endif
  1771. #ifdef PPC440FP2
  1772. #define SNUMOPT 4
  1773. #define DNUMOPT 4
  1774. #define GEMM_DEFAULT_OFFSET_A (32 * 0)
  1775. #define GEMM_DEFAULT_OFFSET_B (32 * 0)
  1776. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1777. #define SGEMM_DEFAULT_UNROLL_M 8
  1778. #define SGEMM_DEFAULT_UNROLL_N 4
  1779. #define DGEMM_DEFAULT_UNROLL_M 8
  1780. #define DGEMM_DEFAULT_UNROLL_N 4
  1781. #define CGEMM_DEFAULT_UNROLL_M 4
  1782. #define CGEMM_DEFAULT_UNROLL_N 2
  1783. #define ZGEMM_DEFAULT_UNROLL_M 4
  1784. #define ZGEMM_DEFAULT_UNROLL_N 2
  1785. #define SGEMM_DEFAULT_P 128
  1786. #define DGEMM_DEFAULT_P 128
  1787. #define CGEMM_DEFAULT_P 128
  1788. #define ZGEMM_DEFAULT_P 128
  1789. #if 1
  1790. #define SGEMM_DEFAULT_Q 4096
  1791. #define DGEMM_DEFAULT_Q 3072
  1792. #define CGEMM_DEFAULT_Q 2048
  1793. #define ZGEMM_DEFAULT_Q 1024
  1794. #else
  1795. #define SGEMM_DEFAULT_Q 512
  1796. #define DGEMM_DEFAULT_Q 256
  1797. #define CGEMM_DEFAULT_Q 256
  1798. #define ZGEMM_DEFAULT_Q 128
  1799. #endif
  1800. #define SYMV_P 4
  1801. #endif
  1802. #if defined(POWER3) || defined(POWER4) || defined(POWER5)
  1803. #define GEMM_DEFAULT_OFFSET_A 0
  1804. #define GEMM_DEFAULT_OFFSET_B 2048
  1805. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1806. #define SGEMM_DEFAULT_UNROLL_M 4
  1807. #define SGEMM_DEFAULT_UNROLL_N 4
  1808. #define DGEMM_DEFAULT_UNROLL_M 4
  1809. #define DGEMM_DEFAULT_UNROLL_N 4
  1810. #define CGEMM_DEFAULT_UNROLL_M 2
  1811. #define CGEMM_DEFAULT_UNROLL_N 2
  1812. #define ZGEMM_DEFAULT_UNROLL_M 2
  1813. #define ZGEMM_DEFAULT_UNROLL_N 2
  1814. #ifdef POWER3
  1815. #define SNUMOPT 4
  1816. #define DNUMOPT 4
  1817. #define SGEMM_DEFAULT_P 256
  1818. #define SGEMM_DEFAULT_Q 432
  1819. #define SGEMM_DEFAULT_R 1012
  1820. #define DGEMM_DEFAULT_P 256
  1821. #define DGEMM_DEFAULT_Q 216
  1822. #define DGEMM_DEFAULT_R 1012
  1823. #define ZGEMM_DEFAULT_P 256
  1824. #define ZGEMM_DEFAULT_Q 104
  1825. #define ZGEMM_DEFAULT_R 1012
  1826. #endif
  1827. #if defined(POWER4)
  1828. #ifdef ALLOC_HUGETLB
  1829. #define SGEMM_DEFAULT_P 184
  1830. #define DGEMM_DEFAULT_P 184
  1831. #define CGEMM_DEFAULT_P 184
  1832. #define ZGEMM_DEFAULT_P 184
  1833. #else
  1834. #define SGEMM_DEFAULT_P 144
  1835. #define DGEMM_DEFAULT_P 144
  1836. #define CGEMM_DEFAULT_P 144
  1837. #define ZGEMM_DEFAULT_P 144
  1838. #endif
  1839. #endif
  1840. #if defined(POWER5)
  1841. #ifdef ALLOC_HUGETLB
  1842. #define SGEMM_DEFAULT_P 512
  1843. #define DGEMM_DEFAULT_P 256
  1844. #define CGEMM_DEFAULT_P 256
  1845. #define ZGEMM_DEFAULT_P 128
  1846. #else
  1847. #define SGEMM_DEFAULT_P 320
  1848. #define DGEMM_DEFAULT_P 160
  1849. #define CGEMM_DEFAULT_P 160
  1850. #define ZGEMM_DEFAULT_P 80
  1851. #endif
  1852. #define SGEMM_DEFAULT_Q 256
  1853. #define CGEMM_DEFAULT_Q 256
  1854. #define DGEMM_DEFAULT_Q 256
  1855. #define ZGEMM_DEFAULT_Q 256
  1856. #endif
  1857. #define SYMV_P 8
  1858. #endif
  1859. #if defined(POWER6)
  1860. #define SNUMOPT 4
  1861. #define DNUMOPT 4
  1862. #define GEMM_DEFAULT_OFFSET_A 384
  1863. #define GEMM_DEFAULT_OFFSET_B 1024
  1864. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  1865. #define SGEMM_DEFAULT_UNROLL_M 4
  1866. #define SGEMM_DEFAULT_UNROLL_N 4
  1867. #define DGEMM_DEFAULT_UNROLL_M 4
  1868. #define DGEMM_DEFAULT_UNROLL_N 4
  1869. #define CGEMM_DEFAULT_UNROLL_M 2
  1870. #define CGEMM_DEFAULT_UNROLL_N 4
  1871. #define ZGEMM_DEFAULT_UNROLL_M 2
  1872. #define ZGEMM_DEFAULT_UNROLL_N 4
  1873. #define SGEMM_DEFAULT_P 992
  1874. #define DGEMM_DEFAULT_P 480
  1875. #define CGEMM_DEFAULT_P 488
  1876. #define ZGEMM_DEFAULT_P 248
  1877. #define SGEMM_DEFAULT_Q 504
  1878. #define DGEMM_DEFAULT_Q 504
  1879. #define CGEMM_DEFAULT_Q 400
  1880. #define ZGEMM_DEFAULT_Q 400
  1881. #define SYMV_P 8
  1882. #endif
  1883. #if defined(POWER8)
  1884. #define SNUMOPT 16
  1885. #define DNUMOPT 8
  1886. #define GEMM_DEFAULT_OFFSET_A 0
  1887. #define GEMM_DEFAULT_OFFSET_B 65536
  1888. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1889. #if defined(__32BIT__)
  1890. #warning using BINARY32==POWER6
  1891. #define SGEMM_DEFAULT_UNROLL_M 4
  1892. #define SGEMM_DEFAULT_UNROLL_N 4
  1893. #define DGEMM_DEFAULT_UNROLL_M 4
  1894. #define DGEMM_DEFAULT_UNROLL_N 4
  1895. #define CGEMM_DEFAULT_UNROLL_M 2
  1896. #define CGEMM_DEFAULT_UNROLL_N 4
  1897. #define ZGEMM_DEFAULT_UNROLL_M 2
  1898. #define ZGEMM_DEFAULT_UNROLL_N 4
  1899. #else
  1900. #define SGEMM_DEFAULT_UNROLL_M 16
  1901. #define SGEMM_DEFAULT_UNROLL_N 8
  1902. #define DGEMM_DEFAULT_UNROLL_M 16
  1903. #define DGEMM_DEFAULT_UNROLL_N 4
  1904. #define CGEMM_DEFAULT_UNROLL_M 8
  1905. #define CGEMM_DEFAULT_UNROLL_N 4
  1906. #define ZGEMM_DEFAULT_UNROLL_M 8
  1907. #define ZGEMM_DEFAULT_UNROLL_N 2
  1908. #endif
  1909. #define SGEMM_DEFAULT_P 1280UL
  1910. #define DGEMM_DEFAULT_P 640UL
  1911. #define CGEMM_DEFAULT_P 640UL
  1912. #define ZGEMM_DEFAULT_P 320UL
  1913. #define SGEMM_DEFAULT_Q 640UL
  1914. #define DGEMM_DEFAULT_Q 720UL
  1915. #define CGEMM_DEFAULT_Q 640UL
  1916. #define ZGEMM_DEFAULT_Q 640UL
  1917. #if 0
  1918. #define SGEMM_DEFAULT_R SGEMM_DEFAULT_P
  1919. #define DGEMM_DEFAULT_R DGEMM_DEFAULT_P
  1920. #define CGEMM_DEFAULT_R CGEMM_DEFAULT_P
  1921. #define ZGEMM_DEFAULT_R ZGEMM_DEFAULT_P
  1922. #endif
  1923. #define SGEMM_DEFAULT_R 4096
  1924. #define DGEMM_DEFAULT_R 4096
  1925. #define CGEMM_DEFAULT_R 4096
  1926. #define ZGEMM_DEFAULT_R 4096
  1927. #define SYMV_P 8
  1928. #endif
  1929. #if defined(POWER9)
  1930. #define SNUMOPT 16
  1931. #define DNUMOPT 8
  1932. #define GEMM_DEFAULT_OFFSET_A 0
  1933. #define GEMM_DEFAULT_OFFSET_B 65536
  1934. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1935. #define SGEMM_DEFAULT_UNROLL_M 16
  1936. #define SGEMM_DEFAULT_UNROLL_N 8
  1937. #define DGEMM_DEFAULT_UNROLL_M 16
  1938. #define DGEMM_DEFAULT_UNROLL_N 4
  1939. #define CGEMM_DEFAULT_UNROLL_M 8
  1940. #define CGEMM_DEFAULT_UNROLL_N 4
  1941. #define ZGEMM_DEFAULT_UNROLL_M 8
  1942. #define ZGEMM_DEFAULT_UNROLL_N 2
  1943. #define SGEMM_DEFAULT_P 832
  1944. #define DGEMM_DEFAULT_P 128
  1945. #define CGEMM_DEFAULT_P 512
  1946. #define ZGEMM_DEFAULT_P 256
  1947. #define SGEMM_DEFAULT_Q 1026
  1948. #define DGEMM_DEFAULT_Q 384
  1949. #define CGEMM_DEFAULT_Q 1026
  1950. #define ZGEMM_DEFAULT_Q 1026
  1951. #define SGEMM_DEFAULT_R 4096
  1952. #define DGEMM_DEFAULT_R 4096
  1953. #define CGEMM_DEFAULT_R 4096
  1954. #define ZGEMM_DEFAULT_R 4096
  1955. #define SYMV_P 8
  1956. #endif
  1957. #if defined(POWER10)
  1958. #define SNUMOPT 16
  1959. #define DNUMOPT 8
  1960. #define GEMM_DEFAULT_OFFSET_A 0
  1961. #define GEMM_DEFAULT_OFFSET_B 65536
  1962. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  1963. #define SGEMM_DEFAULT_UNROLL_M 16
  1964. #define SGEMM_DEFAULT_UNROLL_N 8
  1965. #define DGEMM_DEFAULT_UNROLL_M 8
  1966. #define DGEMM_DEFAULT_UNROLL_N 8
  1967. #define CGEMM_DEFAULT_UNROLL_M 8
  1968. #define CGEMM_DEFAULT_UNROLL_N 4
  1969. #define ZGEMM_DEFAULT_UNROLL_M 8
  1970. #define ZGEMM_DEFAULT_UNROLL_N 2
  1971. #define SGEMM_DEFAULT_P 832
  1972. #define DGEMM_DEFAULT_P 320
  1973. #define CGEMM_DEFAULT_P 512
  1974. #define ZGEMM_DEFAULT_P 256
  1975. #define SGEMM_DEFAULT_Q 1026
  1976. #define DGEMM_DEFAULT_Q 960
  1977. #define CGEMM_DEFAULT_Q 1026
  1978. #define ZGEMM_DEFAULT_Q 1026
  1979. #define SGEMM_DEFAULT_R 4096
  1980. #define DGEMM_DEFAULT_R 4096
  1981. #define CGEMM_DEFAULT_R 4096
  1982. #define ZGEMM_DEFAULT_R 4096
  1983. #define SYMV_P 8
  1984. #undef SBGEMM_DEFAULT_UNROLL_N
  1985. #undef SBGEMM_DEFAULT_UNROLL_M
  1986. #undef SBGEMM_DEFAULT_P
  1987. #undef SBGEMM_DEFAULT_R
  1988. #undef SBGEMM_DEFAULT_Q
  1989. #define SBGEMM_DEFAULT_UNROLL_M 16
  1990. #define SBGEMM_DEFAULT_UNROLL_N 8
  1991. #define SBGEMM_DEFAULT_P 832
  1992. #define SBGEMM_DEFAULT_Q 1026
  1993. #define SBGEMM_DEFAULT_R 4096
  1994. #endif
  1995. #if defined(SPARC) && defined(V7)
  1996. #define SNUMOPT 4
  1997. #define DNUMOPT 4
  1998. #define GEMM_DEFAULT_OFFSET_A 0
  1999. #define GEMM_DEFAULT_OFFSET_B 2048
  2000. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2001. #define SGEMM_DEFAULT_UNROLL_M 2
  2002. #define SGEMM_DEFAULT_UNROLL_N 8
  2003. #define DGEMM_DEFAULT_UNROLL_M 2
  2004. #define DGEMM_DEFAULT_UNROLL_N 8
  2005. #define CGEMM_DEFAULT_UNROLL_M 1
  2006. #define CGEMM_DEFAULT_UNROLL_N 4
  2007. #define ZGEMM_DEFAULT_UNROLL_M 1
  2008. #define ZGEMM_DEFAULT_UNROLL_N 4
  2009. #define SGEMM_DEFAULT_P 256
  2010. #define DGEMM_DEFAULT_P 256
  2011. #define CGEMM_DEFAULT_P 256
  2012. #define ZGEMM_DEFAULT_P 256
  2013. #define SGEMM_DEFAULT_Q 512
  2014. #define DGEMM_DEFAULT_Q 256
  2015. #define CGEMM_DEFAULT_Q 256
  2016. #define ZGEMM_DEFAULT_Q 128
  2017. #define SYMV_P 8
  2018. #define GEMM_THREAD gemm_thread_mn
  2019. #endif
  2020. #if (defined(SPARC) && defined(V9)) || defined(__sparc_v9__)
  2021. #define SNUMOPT 2
  2022. #define DNUMOPT 2
  2023. #define GEMM_DEFAULT_OFFSET_A 0
  2024. #define GEMM_DEFAULT_OFFSET_B 2048
  2025. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2026. #define SGEMM_DEFAULT_UNROLL_M 4
  2027. #define SGEMM_DEFAULT_UNROLL_N 4
  2028. #define DGEMM_DEFAULT_UNROLL_M 4
  2029. #define DGEMM_DEFAULT_UNROLL_N 4
  2030. #define CGEMM_DEFAULT_UNROLL_M 2
  2031. #define CGEMM_DEFAULT_UNROLL_N 2
  2032. #define ZGEMM_DEFAULT_UNROLL_M 2
  2033. #define ZGEMM_DEFAULT_UNROLL_N 2
  2034. #define SGEMM_DEFAULT_P 512
  2035. #define DGEMM_DEFAULT_P 512
  2036. #define CGEMM_DEFAULT_P 512
  2037. #define ZGEMM_DEFAULT_P 512
  2038. #define SGEMM_DEFAULT_Q 1024
  2039. #define DGEMM_DEFAULT_Q 512
  2040. #define CGEMM_DEFAULT_Q 512
  2041. #define ZGEMM_DEFAULT_Q 256
  2042. #define SYMV_P 8
  2043. #endif
  2044. #ifdef SICORTEX
  2045. #define SNUMOPT 2
  2046. #define DNUMOPT 2
  2047. #define GEMM_DEFAULT_OFFSET_A 0
  2048. #define GEMM_DEFAULT_OFFSET_B 0
  2049. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2050. #define SGEMM_DEFAULT_UNROLL_M 2
  2051. #define SGEMM_DEFAULT_UNROLL_N 8
  2052. #define DGEMM_DEFAULT_UNROLL_M 2
  2053. #define DGEMM_DEFAULT_UNROLL_N 8
  2054. #define CGEMM_DEFAULT_UNROLL_M 1
  2055. #define CGEMM_DEFAULT_UNROLL_N 4
  2056. #define ZGEMM_DEFAULT_UNROLL_M 1
  2057. #define ZGEMM_DEFAULT_UNROLL_N 4
  2058. #define SGEMM_DEFAULT_P 108
  2059. #define DGEMM_DEFAULT_P 112
  2060. #define CGEMM_DEFAULT_P 108
  2061. #define ZGEMM_DEFAULT_P 112
  2062. #define SGEMM_DEFAULT_Q 288
  2063. #define DGEMM_DEFAULT_Q 144
  2064. #define CGEMM_DEFAULT_Q 144
  2065. #define ZGEMM_DEFAULT_Q 72
  2066. #define SGEMM_DEFAULT_R 2000
  2067. #define DGEMM_DEFAULT_R 2000
  2068. #define CGEMM_DEFAULT_R 2000
  2069. #define ZGEMM_DEFAULT_R 2000
  2070. #define SYMV_P 16
  2071. #endif
  2072. #if defined(LOONGSON3R4)
  2073. #define SNUMOPT 2
  2074. #define DNUMOPT 2
  2075. #define GEMM_DEFAULT_OFFSET_A 0
  2076. #define GEMM_DEFAULT_OFFSET_B 0
  2077. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2078. #ifdef HAVE_MSA
  2079. #define SGEMM_DEFAULT_UNROLL_M 8
  2080. #define SGEMM_DEFAULT_UNROLL_N 8
  2081. #define DGEMM_DEFAULT_UNROLL_M 8
  2082. #define DGEMM_DEFAULT_UNROLL_N 4
  2083. #define CGEMM_DEFAULT_UNROLL_M 8
  2084. #define CGEMM_DEFAULT_UNROLL_N 4
  2085. #define ZGEMM_DEFAULT_UNROLL_M 4
  2086. #define ZGEMM_DEFAULT_UNROLL_N 4
  2087. #else
  2088. #define SGEMM_DEFAULT_UNROLL_M 8
  2089. #define SGEMM_DEFAULT_UNROLL_N 4
  2090. #define DGEMM_DEFAULT_UNROLL_M 4
  2091. #define DGEMM_DEFAULT_UNROLL_N 4
  2092. #define CGEMM_DEFAULT_UNROLL_M 4
  2093. #define CGEMM_DEFAULT_UNROLL_N 2
  2094. #define ZGEMM_DEFAULT_UNROLL_M 2
  2095. #define ZGEMM_DEFAULT_UNROLL_N 2
  2096. #endif
  2097. #define SGEMM_DEFAULT_P 64
  2098. #define DGEMM_DEFAULT_P 44
  2099. #define CGEMM_DEFAULT_P 64
  2100. #define ZGEMM_DEFAULT_P 32
  2101. #define SGEMM_DEFAULT_Q 192
  2102. #define DGEMM_DEFAULT_Q 92
  2103. #define CGEMM_DEFAULT_Q 128
  2104. #define ZGEMM_DEFAULT_Q 80
  2105. #define SGEMM_DEFAULT_R 640
  2106. #define DGEMM_DEFAULT_R dgemm_r
  2107. #define CGEMM_DEFAULT_R 640
  2108. #define ZGEMM_DEFAULT_R 640
  2109. #define GEMM_OFFSET_A1 0x10000
  2110. #define GEMM_OFFSET_B1 0x100000
  2111. #define SYMV_P 16
  2112. #endif
  2113. #if defined(LOONGSON3R3)
  2114. ////Copy from SICORTEX
  2115. #define SNUMOPT 2
  2116. #define DNUMOPT 2
  2117. #define GEMM_DEFAULT_OFFSET_A 0
  2118. #define GEMM_DEFAULT_OFFSET_B 0
  2119. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2120. #define SGEMM_DEFAULT_UNROLL_M 8
  2121. #define SGEMM_DEFAULT_UNROLL_N 4
  2122. #define DGEMM_DEFAULT_UNROLL_M 4
  2123. #define DGEMM_DEFAULT_UNROLL_N 4
  2124. #define CGEMM_DEFAULT_UNROLL_M 4
  2125. #define CGEMM_DEFAULT_UNROLL_N 2
  2126. #define ZGEMM_DEFAULT_UNROLL_M 2
  2127. #define ZGEMM_DEFAULT_UNROLL_N 2
  2128. #define SGEMM_DEFAULT_P 64
  2129. #define DGEMM_DEFAULT_P 44
  2130. #define CGEMM_DEFAULT_P 64
  2131. #define ZGEMM_DEFAULT_P 32
  2132. #define SGEMM_DEFAULT_Q 192
  2133. #define DGEMM_DEFAULT_Q 92
  2134. #define CGEMM_DEFAULT_Q 128
  2135. #define ZGEMM_DEFAULT_Q 80
  2136. #define SGEMM_DEFAULT_R 640
  2137. #define DGEMM_DEFAULT_R dgemm_r
  2138. #define CGEMM_DEFAULT_R 640
  2139. #define ZGEMM_DEFAULT_R 640
  2140. #define GEMM_OFFSET_A1 0x10000
  2141. #define GEMM_OFFSET_B1 0x100000
  2142. #define SYMV_P 16
  2143. #endif
  2144. #if defined(P5600) || defined(MIPS1004K) || defined(MIPS24K) || defined(I6400) || defined(P6600) || defined(I6500)
  2145. #define SNUMOPT 2
  2146. #define DNUMOPT 2
  2147. #define GEMM_DEFAULT_OFFSET_A 0
  2148. #define GEMM_DEFAULT_OFFSET_B 0
  2149. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2150. #ifdef HAVE_MSA
  2151. #define SGEMM_DEFAULT_UNROLL_M 8
  2152. #define SGEMM_DEFAULT_UNROLL_N 8
  2153. #define DGEMM_DEFAULT_UNROLL_M 8
  2154. #define DGEMM_DEFAULT_UNROLL_N 4
  2155. #define CGEMM_DEFAULT_UNROLL_M 8
  2156. #define CGEMM_DEFAULT_UNROLL_N 4
  2157. #define ZGEMM_DEFAULT_UNROLL_M 4
  2158. #define ZGEMM_DEFAULT_UNROLL_N 4
  2159. #else
  2160. #define SGEMM_DEFAULT_UNROLL_M 2
  2161. #define SGEMM_DEFAULT_UNROLL_N 2
  2162. #define DGEMM_DEFAULT_UNROLL_M 2
  2163. #define DGEMM_DEFAULT_UNROLL_N 2
  2164. #define CGEMM_DEFAULT_UNROLL_M 2
  2165. #define CGEMM_DEFAULT_UNROLL_N 2
  2166. #define ZGEMM_DEFAULT_UNROLL_M 2
  2167. #define ZGEMM_DEFAULT_UNROLL_N 2
  2168. #endif
  2169. #define SGEMM_DEFAULT_P 128
  2170. #define DGEMM_DEFAULT_P 128
  2171. #define CGEMM_DEFAULT_P 96
  2172. #define ZGEMM_DEFAULT_P 64
  2173. #define SGEMM_DEFAULT_Q 240
  2174. #define DGEMM_DEFAULT_Q 120
  2175. #define CGEMM_DEFAULT_Q 120
  2176. #define ZGEMM_DEFAULT_Q 120
  2177. #define SGEMM_DEFAULT_R 12288
  2178. #define DGEMM_DEFAULT_R 8192
  2179. #define CGEMM_DEFAULT_R 4096
  2180. #define ZGEMM_DEFAULT_R 4096
  2181. #define SYMV_P 16
  2182. #endif
  2183. #ifdef RISCV64_GENERIC
  2184. #define GEMM_DEFAULT_OFFSET_A 0
  2185. #define GEMM_DEFAULT_OFFSET_B 0
  2186. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2187. #define SGEMM_DEFAULT_UNROLL_M 2
  2188. #define SGEMM_DEFAULT_UNROLL_N 2
  2189. #define DGEMM_DEFAULT_UNROLL_M 2
  2190. #define DGEMM_DEFAULT_UNROLL_N 2
  2191. #define CGEMM_DEFAULT_UNROLL_M 2
  2192. #define CGEMM_DEFAULT_UNROLL_N 2
  2193. #define ZGEMM_DEFAULT_UNROLL_M 2
  2194. #define ZGEMM_DEFAULT_UNROLL_N 2
  2195. #define SGEMM_DEFAULT_P 128
  2196. #define DGEMM_DEFAULT_P 128
  2197. #define CGEMM_DEFAULT_P 96
  2198. #define ZGEMM_DEFAULT_P 64
  2199. #define SGEMM_DEFAULT_Q 240
  2200. #define DGEMM_DEFAULT_Q 120
  2201. #define CGEMM_DEFAULT_Q 120
  2202. #define ZGEMM_DEFAULT_Q 120
  2203. #define SGEMM_DEFAULT_R 12288
  2204. #define DGEMM_DEFAULT_R 8192
  2205. #define CGEMM_DEFAULT_R 4096
  2206. #define ZGEMM_DEFAULT_R 4096
  2207. #define SYMV_P 16
  2208. #define GEMM_DEFAULT_OFFSET_A 0
  2209. #define GEMM_DEFAULT_OFFSET_B 0
  2210. #endif
  2211. #ifdef C910V
  2212. #define GEMM_DEFAULT_OFFSET_A 0
  2213. #define GEMM_DEFAULT_OFFSET_B 0
  2214. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2215. #define SGEMM_DEFAULT_UNROLL_M 16
  2216. #define SGEMM_DEFAULT_UNROLL_N 4
  2217. #define DGEMM_DEFAULT_UNROLL_M 8
  2218. #define DGEMM_DEFAULT_UNROLL_N 4
  2219. #define CGEMM_DEFAULT_UNROLL_M 2
  2220. #define CGEMM_DEFAULT_UNROLL_N 2
  2221. #define ZGEMM_DEFAULT_UNROLL_M 2
  2222. #define ZGEMM_DEFAULT_UNROLL_N 2
  2223. #define SGEMM_DEFAULT_P 160
  2224. #define DGEMM_DEFAULT_P 160
  2225. #define CGEMM_DEFAULT_P 96
  2226. #define ZGEMM_DEFAULT_P 64
  2227. #define SGEMM_DEFAULT_Q 240
  2228. #define DGEMM_DEFAULT_Q 128
  2229. #define CGEMM_DEFAULT_Q 120
  2230. #define ZGEMM_DEFAULT_Q 120
  2231. #define SGEMM_DEFAULT_R 12288
  2232. #define DGEMM_DEFAULT_R 8192
  2233. #define CGEMM_DEFAULT_R 4096
  2234. #define ZGEMM_DEFAULT_R 4096
  2235. #define SYMV_P 16
  2236. #define GEMM_DEFAULT_OFFSET_A 0
  2237. #define GEMM_DEFAULT_OFFSET_B 0
  2238. #endif
  2239. #ifdef ARMV7
  2240. #define SNUMOPT 2
  2241. #define DNUMOPT 2
  2242. #define GEMM_DEFAULT_OFFSET_A 0
  2243. #define GEMM_DEFAULT_OFFSET_B 0
  2244. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2245. #define SGEMM_DEFAULT_UNROLL_M 4
  2246. #define SGEMM_DEFAULT_UNROLL_N 4
  2247. #define DGEMM_DEFAULT_UNROLL_M 4
  2248. #define DGEMM_DEFAULT_UNROLL_N 4
  2249. #define CGEMM_DEFAULT_UNROLL_M 2
  2250. #define CGEMM_DEFAULT_UNROLL_N 2
  2251. #define ZGEMM_DEFAULT_UNROLL_M 2
  2252. #define ZGEMM_DEFAULT_UNROLL_N 2
  2253. #define SGEMM_DEFAULT_P 128
  2254. #define DGEMM_DEFAULT_P 128
  2255. #define CGEMM_DEFAULT_P 96
  2256. #define ZGEMM_DEFAULT_P 64
  2257. #define SGEMM_DEFAULT_Q 240
  2258. #define DGEMM_DEFAULT_Q 120
  2259. #define CGEMM_DEFAULT_Q 120
  2260. #define ZGEMM_DEFAULT_Q 120
  2261. #define SGEMM_DEFAULT_R 12288
  2262. #define DGEMM_DEFAULT_R 8192
  2263. #define CGEMM_DEFAULT_R 4096
  2264. #define ZGEMM_DEFAULT_R 4096
  2265. #define SYMV_P 16
  2266. #endif
  2267. #if defined(ARMV6)
  2268. #define SNUMOPT 2
  2269. #define DNUMOPT 2
  2270. #define GEMM_DEFAULT_OFFSET_A 0
  2271. #define GEMM_DEFAULT_OFFSET_B 0
  2272. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2273. #define SGEMM_DEFAULT_UNROLL_M 4
  2274. #define SGEMM_DEFAULT_UNROLL_N 2
  2275. #define DGEMM_DEFAULT_UNROLL_M 4
  2276. #define DGEMM_DEFAULT_UNROLL_N 2
  2277. #define CGEMM_DEFAULT_UNROLL_M 2
  2278. #define CGEMM_DEFAULT_UNROLL_N 2
  2279. #define ZGEMM_DEFAULT_UNROLL_M 2
  2280. #define ZGEMM_DEFAULT_UNROLL_N 2
  2281. #define SGEMM_DEFAULT_P 128
  2282. #define DGEMM_DEFAULT_P 128
  2283. #define CGEMM_DEFAULT_P 96
  2284. #define ZGEMM_DEFAULT_P 64
  2285. #define SGEMM_DEFAULT_Q 240
  2286. #define DGEMM_DEFAULT_Q 120
  2287. #define CGEMM_DEFAULT_Q 120
  2288. #define ZGEMM_DEFAULT_Q 120
  2289. #define SGEMM_DEFAULT_R 12288
  2290. #define DGEMM_DEFAULT_R 8192
  2291. #define CGEMM_DEFAULT_R 4096
  2292. #define ZGEMM_DEFAULT_R 4096
  2293. #define SYMV_P 16
  2294. #endif
  2295. /* Common ARMv8 parameters */
  2296. #if defined(ARMV8)
  2297. #define SNUMOPT 2
  2298. #define DNUMOPT 2
  2299. #define GEMM_DEFAULT_OFFSET_A 0
  2300. #define GEMM_DEFAULT_OFFSET_B 0
  2301. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2302. #define SYMV_P 16
  2303. #if defined(CORTEXA57) || \
  2304. defined(CORTEXA72) || defined(CORTEXA73) || \
  2305. defined(FALKOR) || defined(TSV110) || defined(EMAG8180)
  2306. #define SGEMM_DEFAULT_UNROLL_M 16
  2307. #define SGEMM_DEFAULT_UNROLL_N 4
  2308. #define DGEMM_DEFAULT_UNROLL_M 8
  2309. #define DGEMM_DEFAULT_UNROLL_N 4
  2310. #define CGEMM_DEFAULT_UNROLL_M 8
  2311. #define CGEMM_DEFAULT_UNROLL_N 4
  2312. #define ZGEMM_DEFAULT_UNROLL_M 4
  2313. #define ZGEMM_DEFAULT_UNROLL_N 4
  2314. /*FIXME: this should be using the cache size, but there is currently no easy way to
  2315. query that on ARM. So if getarch counted more than 8 cores we simply assume the host
  2316. is a big desktop or server with abundant cache rather than a phone or embedded device */
  2317. #if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180)
  2318. #define SGEMM_DEFAULT_P 512
  2319. #define DGEMM_DEFAULT_P 256
  2320. #define CGEMM_DEFAULT_P 256
  2321. #define ZGEMM_DEFAULT_P 128
  2322. #define SGEMM_DEFAULT_Q 1024
  2323. #define DGEMM_DEFAULT_Q 512
  2324. #define CGEMM_DEFAULT_Q 512
  2325. #define ZGEMM_DEFAULT_Q 512
  2326. #else
  2327. #define SGEMM_DEFAULT_P 128
  2328. #define DGEMM_DEFAULT_P 160
  2329. #define CGEMM_DEFAULT_P 128
  2330. #define ZGEMM_DEFAULT_P 128
  2331. #define SGEMM_DEFAULT_Q 352
  2332. #define DGEMM_DEFAULT_Q 128
  2333. #define CGEMM_DEFAULT_Q 224
  2334. #define ZGEMM_DEFAULT_Q 112
  2335. #endif
  2336. #define SGEMM_DEFAULT_R 4096
  2337. #define DGEMM_DEFAULT_R 4096
  2338. #define CGEMM_DEFAULT_R 4096
  2339. #define ZGEMM_DEFAULT_R 2048
  2340. #elif defined(CORTEXA53)
  2341. #define SGEMM_DEFAULT_UNROLL_M 8
  2342. #define SGEMM_DEFAULT_UNROLL_N 8
  2343. #define DGEMM_DEFAULT_UNROLL_M 8
  2344. #define DGEMM_DEFAULT_UNROLL_N 4
  2345. #define CGEMM_DEFAULT_UNROLL_M 8
  2346. #define CGEMM_DEFAULT_UNROLL_N 4
  2347. #define ZGEMM_DEFAULT_UNROLL_M 4
  2348. #define ZGEMM_DEFAULT_UNROLL_N 4
  2349. #define SGEMM_DEFAULT_P 256
  2350. #define DGEMM_DEFAULT_P 160
  2351. #define CGEMM_DEFAULT_P 128
  2352. #define ZGEMM_DEFAULT_P 128
  2353. #define SGEMM_DEFAULT_Q 256
  2354. #define DGEMM_DEFAULT_Q 128
  2355. #define CGEMM_DEFAULT_Q 224
  2356. #define ZGEMM_DEFAULT_Q 112
  2357. #define SGEMM_DEFAULT_R 4096
  2358. #define DGEMM_DEFAULT_R 4096
  2359. #define CGEMM_DEFAULT_R 4096
  2360. #define ZGEMM_DEFAULT_R 2048
  2361. #elif defined(THUNDERX)
  2362. #define SGEMM_DEFAULT_UNROLL_M 4
  2363. #define SGEMM_DEFAULT_UNROLL_N 4
  2364. #define DGEMM_DEFAULT_UNROLL_M 2
  2365. #define DGEMM_DEFAULT_UNROLL_N 2
  2366. #define CGEMM_DEFAULT_UNROLL_M 2
  2367. #define CGEMM_DEFAULT_UNROLL_N 2
  2368. #define ZGEMM_DEFAULT_UNROLL_M 2
  2369. #define ZGEMM_DEFAULT_UNROLL_N 2
  2370. #define SGEMM_DEFAULT_P 128
  2371. #define DGEMM_DEFAULT_P 128
  2372. #define CGEMM_DEFAULT_P 96
  2373. #define ZGEMM_DEFAULT_P 64
  2374. #define SGEMM_DEFAULT_Q 240
  2375. #define DGEMM_DEFAULT_Q 120
  2376. #define CGEMM_DEFAULT_Q 120
  2377. #define ZGEMM_DEFAULT_Q 120
  2378. #define SGEMM_DEFAULT_R 12288
  2379. #define DGEMM_DEFAULT_R 8192
  2380. #define CGEMM_DEFAULT_R 4096
  2381. #define ZGEMM_DEFAULT_R 4096
  2382. #elif defined(THUNDERX2T99)
  2383. #define SGEMM_DEFAULT_UNROLL_M 16
  2384. #define SGEMM_DEFAULT_UNROLL_N 4
  2385. #define DGEMM_DEFAULT_UNROLL_M 8
  2386. #define DGEMM_DEFAULT_UNROLL_N 4
  2387. #define CGEMM_DEFAULT_UNROLL_M 8
  2388. #define CGEMM_DEFAULT_UNROLL_N 4
  2389. #define ZGEMM_DEFAULT_UNROLL_M 4
  2390. #define ZGEMM_DEFAULT_UNROLL_N 4
  2391. #define SGEMM_DEFAULT_P 128
  2392. #define DGEMM_DEFAULT_P 160
  2393. #define CGEMM_DEFAULT_P 128
  2394. #define ZGEMM_DEFAULT_P 128
  2395. #define SGEMM_DEFAULT_Q 352
  2396. #define DGEMM_DEFAULT_Q 128
  2397. #define CGEMM_DEFAULT_Q 224
  2398. #define ZGEMM_DEFAULT_Q 112
  2399. #define SGEMM_DEFAULT_R 4096
  2400. #define DGEMM_DEFAULT_R 4096
  2401. #define CGEMM_DEFAULT_R 4096
  2402. #define ZGEMM_DEFAULT_R 4096
  2403. #elif defined(THUNDERX3T110)
  2404. #define SGEMM_DEFAULT_UNROLL_M 16
  2405. #define SGEMM_DEFAULT_UNROLL_N 4
  2406. #define DGEMM_DEFAULT_UNROLL_M 8
  2407. #define DGEMM_DEFAULT_UNROLL_N 4
  2408. #define CGEMM_DEFAULT_UNROLL_M 8
  2409. #define CGEMM_DEFAULT_UNROLL_N 4
  2410. #define ZGEMM_DEFAULT_UNROLL_M 4
  2411. #define ZGEMM_DEFAULT_UNROLL_N 4
  2412. #define SGEMM_DEFAULT_P 128
  2413. #define DGEMM_DEFAULT_P 320
  2414. #define CGEMM_DEFAULT_P 128
  2415. #define ZGEMM_DEFAULT_P 128
  2416. #define SGEMM_DEFAULT_Q 352
  2417. #define DGEMM_DEFAULT_Q 128
  2418. #define CGEMM_DEFAULT_Q 224
  2419. #define ZGEMM_DEFAULT_Q 112
  2420. #define SGEMM_DEFAULT_R 4096
  2421. #define DGEMM_DEFAULT_R 4096
  2422. #define CGEMM_DEFAULT_R 4096
  2423. #define ZGEMM_DEFAULT_R 4096
  2424. #elif defined(NEOVERSEN1)
  2425. #define SGEMM_DEFAULT_UNROLL_M 16
  2426. #define SGEMM_DEFAULT_UNROLL_N 4
  2427. #define DGEMM_DEFAULT_UNROLL_M 8
  2428. #define DGEMM_DEFAULT_UNROLL_N 4
  2429. #define CGEMM_DEFAULT_UNROLL_M 8
  2430. #define CGEMM_DEFAULT_UNROLL_N 4
  2431. #define ZGEMM_DEFAULT_UNROLL_M 4
  2432. #define ZGEMM_DEFAULT_UNROLL_N 4
  2433. #define SGEMM_DEFAULT_P 128
  2434. #define DGEMM_DEFAULT_P 160
  2435. #define CGEMM_DEFAULT_P 128
  2436. #define ZGEMM_DEFAULT_P 128
  2437. #define SGEMM_DEFAULT_Q 352
  2438. #define DGEMM_DEFAULT_Q 128
  2439. #define CGEMM_DEFAULT_Q 224
  2440. #define ZGEMM_DEFAULT_Q 112
  2441. #define SGEMM_DEFAULT_R 4096
  2442. #define DGEMM_DEFAULT_R 4096
  2443. #define CGEMM_DEFAULT_R 4096
  2444. #define ZGEMM_DEFAULT_R 4096
  2445. #else /* Other/undetected ARMv8 cores */
  2446. #define SGEMM_DEFAULT_UNROLL_M 16
  2447. #define SGEMM_DEFAULT_UNROLL_N 4
  2448. #define DGEMM_DEFAULT_UNROLL_M 8
  2449. #define DGEMM_DEFAULT_UNROLL_N 4
  2450. #define CGEMM_DEFAULT_UNROLL_M 8
  2451. #define CGEMM_DEFAULT_UNROLL_N 4
  2452. #define ZGEMM_DEFAULT_UNROLL_M 4
  2453. #define ZGEMM_DEFAULT_UNROLL_N 4
  2454. #define SGEMM_DEFAULT_P 128
  2455. #define DGEMM_DEFAULT_P 160
  2456. #define CGEMM_DEFAULT_P 128
  2457. #define ZGEMM_DEFAULT_P 128
  2458. #define SGEMM_DEFAULT_Q 352
  2459. #define DGEMM_DEFAULT_Q 128
  2460. #define CGEMM_DEFAULT_Q 224
  2461. #define ZGEMM_DEFAULT_Q 112
  2462. #define SGEMM_DEFAULT_R 4096
  2463. #define DGEMM_DEFAULT_R 4096
  2464. #define CGEMM_DEFAULT_R 4096
  2465. #define ZGEMM_DEFAULT_R 4096
  2466. #endif /* Cores */
  2467. #endif /* ARMv8 */
  2468. #if defined(ARMV5)
  2469. #define SNUMOPT 2
  2470. #define DNUMOPT 2
  2471. #define GEMM_DEFAULT_OFFSET_A 0
  2472. #define GEMM_DEFAULT_OFFSET_B 0
  2473. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2474. #define SGEMM_DEFAULT_UNROLL_M 2
  2475. #define SGEMM_DEFAULT_UNROLL_N 2
  2476. #define DGEMM_DEFAULT_UNROLL_M 2
  2477. #define DGEMM_DEFAULT_UNROLL_N 2
  2478. #define CGEMM_DEFAULT_UNROLL_M 2
  2479. #define CGEMM_DEFAULT_UNROLL_N 2
  2480. #define ZGEMM_DEFAULT_UNROLL_M 2
  2481. #define ZGEMM_DEFAULT_UNROLL_N 2
  2482. #define SGEMM_DEFAULT_P 128
  2483. #define DGEMM_DEFAULT_P 128
  2484. #define CGEMM_DEFAULT_P 96
  2485. #define ZGEMM_DEFAULT_P 64
  2486. #define SGEMM_DEFAULT_Q 240
  2487. #define DGEMM_DEFAULT_Q 120
  2488. #define CGEMM_DEFAULT_Q 120
  2489. #define ZGEMM_DEFAULT_Q 120
  2490. #define SGEMM_DEFAULT_R 12288
  2491. #define DGEMM_DEFAULT_R 8192
  2492. #define CGEMM_DEFAULT_R 4096
  2493. #define ZGEMM_DEFAULT_R 4096
  2494. #define SYMV_P 16
  2495. #endif
  2496. #ifdef CORTEXA9
  2497. #define SNUMOPT 2
  2498. #define DNUMOPT 2
  2499. #define GEMM_DEFAULT_OFFSET_A 0
  2500. #define GEMM_DEFAULT_OFFSET_B 0
  2501. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2502. #define SGEMM_DEFAULT_UNROLL_M 4
  2503. #define SGEMM_DEFAULT_UNROLL_N 4
  2504. #define DGEMM_DEFAULT_UNROLL_M 4
  2505. #define DGEMM_DEFAULT_UNROLL_N 4
  2506. #define CGEMM_DEFAULT_UNROLL_M 2
  2507. #define CGEMM_DEFAULT_UNROLL_N 2
  2508. #define ZGEMM_DEFAULT_UNROLL_M 2
  2509. #define ZGEMM_DEFAULT_UNROLL_N 2
  2510. #define SGEMM_DEFAULT_P 128
  2511. #define DGEMM_DEFAULT_P 128
  2512. #define CGEMM_DEFAULT_P 96
  2513. #define ZGEMM_DEFAULT_P 64
  2514. #define SGEMM_DEFAULT_Q 240
  2515. #define DGEMM_DEFAULT_Q 120
  2516. #define CGEMM_DEFAULT_Q 120
  2517. #define ZGEMM_DEFAULT_Q 120
  2518. #define SGEMM_DEFAULT_R 12288
  2519. #define DGEMM_DEFAULT_R 8192
  2520. #define CGEMM_DEFAULT_R 4096
  2521. #define ZGEMM_DEFAULT_R 4096
  2522. #define SYMV_P 16
  2523. #endif
  2524. #ifdef CORTEXA15
  2525. #define SNUMOPT 2
  2526. #define DNUMOPT 2
  2527. #define GEMM_DEFAULT_OFFSET_A 0
  2528. #define GEMM_DEFAULT_OFFSET_B 0
  2529. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2530. #define SGEMM_DEFAULT_UNROLL_M 4
  2531. #define SGEMM_DEFAULT_UNROLL_N 4
  2532. #define DGEMM_DEFAULT_UNROLL_M 4
  2533. #define DGEMM_DEFAULT_UNROLL_N 4
  2534. #define CGEMM_DEFAULT_UNROLL_M 2
  2535. #define CGEMM_DEFAULT_UNROLL_N 2
  2536. #define ZGEMM_DEFAULT_UNROLL_M 2
  2537. #define ZGEMM_DEFAULT_UNROLL_N 2
  2538. #define SGEMM_DEFAULT_P 128
  2539. #define DGEMM_DEFAULT_P 128
  2540. #define CGEMM_DEFAULT_P 96
  2541. #define ZGEMM_DEFAULT_P 64
  2542. #define SGEMM_DEFAULT_Q 240
  2543. #define DGEMM_DEFAULT_Q 120
  2544. #define CGEMM_DEFAULT_Q 120
  2545. #define ZGEMM_DEFAULT_Q 120
  2546. #define SGEMM_DEFAULT_R 12288
  2547. #define DGEMM_DEFAULT_R 8192
  2548. #define CGEMM_DEFAULT_R 4096
  2549. #define ZGEMM_DEFAULT_R 4096
  2550. #define SYMV_P 16
  2551. #endif
  2552. #if defined(ZARCH_GENERIC)
  2553. #define SNUMOPT 2
  2554. #define DNUMOPT 2
  2555. #define GEMM_DEFAULT_OFFSET_A 0
  2556. #define GEMM_DEFAULT_OFFSET_B 0
  2557. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2558. #define SGEMM_DEFAULT_UNROLL_M 2
  2559. #define SGEMM_DEFAULT_UNROLL_N 2
  2560. #define DGEMM_DEFAULT_UNROLL_M 2
  2561. #define DGEMM_DEFAULT_UNROLL_N 2
  2562. #define CGEMM_DEFAULT_UNROLL_M 2
  2563. #define CGEMM_DEFAULT_UNROLL_N 2
  2564. #define ZGEMM_DEFAULT_UNROLL_M 2
  2565. #define ZGEMM_DEFAULT_UNROLL_N 2
  2566. #define SGEMM_DEFAULT_P 128
  2567. #define DGEMM_DEFAULT_P 128
  2568. #define CGEMM_DEFAULT_P 96
  2569. #define ZGEMM_DEFAULT_P 64
  2570. #define SGEMM_DEFAULT_Q 240
  2571. #define DGEMM_DEFAULT_Q 120
  2572. #define CGEMM_DEFAULT_Q 120
  2573. #define ZGEMM_DEFAULT_Q 120
  2574. #define SGEMM_DEFAULT_R 12288
  2575. #define DGEMM_DEFAULT_R 8192
  2576. #define CGEMM_DEFAULT_R 4096
  2577. #define ZGEMM_DEFAULT_R 4096
  2578. #define SYMV_P 16
  2579. #endif
  2580. #if defined(Z13)
  2581. #define SNUMOPT 2
  2582. #define DNUMOPT 2
  2583. #define GEMM_DEFAULT_OFFSET_A 0
  2584. #define GEMM_DEFAULT_OFFSET_B 0
  2585. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2586. #define SGEMM_DEFAULT_UNROLL_M 8
  2587. #define SGEMM_DEFAULT_UNROLL_N 4
  2588. #define DGEMM_DEFAULT_UNROLL_M 8
  2589. #define DGEMM_DEFAULT_UNROLL_N 4
  2590. #define CGEMM_DEFAULT_UNROLL_M 4
  2591. #define CGEMM_DEFAULT_UNROLL_N 4
  2592. #define ZGEMM_DEFAULT_UNROLL_M 4
  2593. #define ZGEMM_DEFAULT_UNROLL_N 4
  2594. #define SGEMM_DEFAULT_P 456
  2595. #define DGEMM_DEFAULT_P 320
  2596. #define CGEMM_DEFAULT_P 480
  2597. #define ZGEMM_DEFAULT_P 224
  2598. #define SGEMM_DEFAULT_Q 488
  2599. #define DGEMM_DEFAULT_Q 384
  2600. #define CGEMM_DEFAULT_Q 128
  2601. #define ZGEMM_DEFAULT_Q 352
  2602. #define SGEMM_DEFAULT_R 8192
  2603. #define DGEMM_DEFAULT_R 4096
  2604. #define CGEMM_DEFAULT_R 4096
  2605. #define ZGEMM_DEFAULT_R 2048
  2606. #define SYMV_P 16
  2607. #endif
  2608. #if defined(Z14)
  2609. #define SNUMOPT 2
  2610. #define DNUMOPT 2
  2611. #define GEMM_DEFAULT_OFFSET_A 0
  2612. #define GEMM_DEFAULT_OFFSET_B 0
  2613. #define GEMM_DEFAULT_ALIGN 0x03fffUL
  2614. #define SGEMM_DEFAULT_UNROLL_M 16
  2615. #define SGEMM_DEFAULT_UNROLL_N 4
  2616. #define DGEMM_DEFAULT_UNROLL_M 8
  2617. #define DGEMM_DEFAULT_UNROLL_N 4
  2618. #define CGEMM_DEFAULT_UNROLL_M 4
  2619. #define CGEMM_DEFAULT_UNROLL_N 4
  2620. #define ZGEMM_DEFAULT_UNROLL_M 4
  2621. #define ZGEMM_DEFAULT_UNROLL_N 4
  2622. #define SGEMM_DEFAULT_P 480
  2623. #define DGEMM_DEFAULT_P 320
  2624. #define CGEMM_DEFAULT_P 480
  2625. #define ZGEMM_DEFAULT_P 224
  2626. #define SGEMM_DEFAULT_Q 512
  2627. #define DGEMM_DEFAULT_Q 384
  2628. #define CGEMM_DEFAULT_Q 128
  2629. #define ZGEMM_DEFAULT_Q 352
  2630. #define SGEMM_DEFAULT_R 8192
  2631. #define DGEMM_DEFAULT_R 4096
  2632. #define CGEMM_DEFAULT_R 4096
  2633. #define ZGEMM_DEFAULT_R 2048
  2634. #define SYMV_P 16
  2635. #endif
  2636. #ifdef GENERIC
  2637. #define SNUMOPT 2
  2638. #define DNUMOPT 2
  2639. #define GEMM_DEFAULT_OFFSET_A 0
  2640. #define GEMM_DEFAULT_OFFSET_B 0
  2641. #define GEMM_DEFAULT_ALIGN 0x0ffffUL
  2642. #define SGEMM_DEFAULT_UNROLL_N 2
  2643. #define DGEMM_DEFAULT_UNROLL_N 2
  2644. #define QGEMM_DEFAULT_UNROLL_N 2
  2645. #define CGEMM_DEFAULT_UNROLL_N 2
  2646. #define ZGEMM_DEFAULT_UNROLL_N 2
  2647. #define XGEMM_DEFAULT_UNROLL_N 1
  2648. #ifdef ARCH_X86
  2649. #define SGEMM_DEFAULT_UNROLL_M 2
  2650. #define DGEMM_DEFAULT_UNROLL_M 2
  2651. #define QGEMM_DEFAULT_UNROLL_M 2
  2652. #define CGEMM_DEFAULT_UNROLL_M 2
  2653. #define ZGEMM_DEFAULT_UNROLL_M 2
  2654. #define XGEMM_DEFAULT_UNROLL_M 1
  2655. #else
  2656. #define SGEMM_DEFAULT_UNROLL_M 2
  2657. #define DGEMM_DEFAULT_UNROLL_M 2
  2658. #define QGEMM_DEFAULT_UNROLL_M 2
  2659. #define CGEMM_DEFAULT_UNROLL_M 2
  2660. #define ZGEMM_DEFAULT_UNROLL_M 2
  2661. #define XGEMM_DEFAULT_UNROLL_M 1
  2662. #endif
  2663. #define SGEMM_DEFAULT_P sgemm_p
  2664. #define DGEMM_DEFAULT_P dgemm_p
  2665. #define QGEMM_DEFAULT_P qgemm_p
  2666. #define CGEMM_DEFAULT_P cgemm_p
  2667. #define ZGEMM_DEFAULT_P zgemm_p
  2668. #define XGEMM_DEFAULT_P xgemm_p
  2669. #define SGEMM_DEFAULT_R sgemm_r
  2670. #define DGEMM_DEFAULT_R dgemm_r
  2671. #define QGEMM_DEFAULT_R qgemm_r
  2672. #define CGEMM_DEFAULT_R cgemm_r
  2673. #define ZGEMM_DEFAULT_R zgemm_r
  2674. #define XGEMM_DEFAULT_R xgemm_r
  2675. #define SGEMM_DEFAULT_Q 128
  2676. #define DGEMM_DEFAULT_Q 128
  2677. #define QGEMM_DEFAULT_Q 128
  2678. #define CGEMM_DEFAULT_Q 128
  2679. #define ZGEMM_DEFAULT_Q 128
  2680. #define XGEMM_DEFAULT_Q 128
  2681. #define SYMV_P 16
  2682. #endif
  2683. #ifndef QGEMM_DEFAULT_UNROLL_M
  2684. #define QGEMM_DEFAULT_UNROLL_M 2
  2685. #endif
  2686. #ifndef QGEMM_DEFAULT_UNROLL_N
  2687. #define QGEMM_DEFAULT_UNROLL_N 2
  2688. #endif
  2689. #ifndef XGEMM_DEFAULT_UNROLL_M
  2690. #define XGEMM_DEFAULT_UNROLL_M 2
  2691. #endif
  2692. #ifndef XGEMM_DEFAULT_UNROLL_N
  2693. #define XGEMM_DEFAULT_UNROLL_N 2
  2694. #endif
  2695. #ifndef HAVE_SSE2
  2696. #define SHUFPD_0 shufps $0x44,
  2697. #define SHUFPD_1 shufps $0x4e,
  2698. #define SHUFPD_2 shufps $0xe4,
  2699. #define SHUFPD_3 shufps $0xee,
  2700. #endif
  2701. #ifndef SHUFPD_0
  2702. #define SHUFPD_0 shufpd $0,
  2703. #endif
  2704. #ifndef SHUFPD_1
  2705. #define SHUFPD_1 shufpd $1,
  2706. #endif
  2707. #ifndef SHUFPD_2
  2708. #define SHUFPD_2 shufpd $2,
  2709. #endif
  2710. #ifndef SHUFPD_3
  2711. #define SHUFPD_3 shufpd $3,
  2712. #endif
  2713. #ifndef SHUFPS_39
  2714. #define SHUFPS_39 shufps $0x39,
  2715. #endif
  2716. #endif