diff options
Diffstat (limited to 'src/runtime/heuristics/matmul_native')
6 files changed, 185 insertions, 292 deletions
diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp b/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp index 01102b3d60..b3c8d891dc 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.cpp @@ -28,30 +28,33 @@ #include "arm_compute/core/GPUTarget.h" #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/TensorInfo.h" -#include "src/gpu/cl/kernels/ClMatMulNativeKernel.h" -#include <utility> +#include "src/gpu/cl/kernels/ClMatMulNativeKernel.h" #include "src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.h" +#include <utility> + namespace arm_compute { namespace cl_matmul { -ClMatMulNativeDefaultConfigValhall::ClMatMulNativeDefaultConfigValhall(GPUTarget gpu) - : IClMatMulNativeKernelConfig(gpu) +ClMatMulNativeDefaultConfigValhall::ClMatMulNativeDefaultConfigValhall(GPUTarget gpu) : IClMatMulNativeKernelConfig(gpu) { } -MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure(const ITensorInfo *lhs, const ITensorInfo *rhs, const MatMulInfo &info) +MatMulKernelInfo +ClMatMulNativeDefaultConfigValhall::configure(const ITensorInfo *lhs, const ITensorInfo *rhs, const MatMulInfo &info) { - using ConfigurationFunctionExecutorPtr = MatMulKernelInfo (ClMatMulNativeDefaultConfigValhall::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo & info); + using ConfigurationFunctionExecutorPtr = MatMulKernelInfo (ClMatMulNativeDefaultConfigValhall::*)( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); - ClMatMulNativeConfigArray<ConfigurationFunctionExecutorPtr> configs_G710(&ClMatMulNativeDefaultConfigValhall::configure_G710_f32, - &ClMatMulNativeDefaultConfigValhall::configure_G710_f16, - &ClMatMulNativeDefaultConfigValhall::configure_G710_u8); + ClMatMulNativeConfigArray<ConfigurationFunctionExecutorPtr> configs_G710( + &ClMatMulNativeDefaultConfigValhall::configure_G710_f32, + &ClMatMulNativeDefaultConfigValhall::configure_G710_f16, + &ClMatMulNativeDefaultConfigValhall::configure_G710_u8); ConfigurationFunctionExecutorPtr func = nullptr; - switch(_target) + switch (_target) { case GPUTarget::G710: default: @@ -67,7 +70,7 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure(const ITensorInfo const bool is_batched = lhs_shape.num_dimensions() > 2; - if(is_batched == true) + if (is_batched == true) { lhs_shape.collapse_from(2); } @@ -81,103 +84,48 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure(const ITensorInfo return (this->*func)(m, n, k, b, rhs->lock_paddings(), info); } -MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) +MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f32( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) { - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = - { - { 3136, 64, 64, 36, 4, 4, 16, 1 }, - { 4096, 48, 32, 36, 4, 4, 4, 1 }, - { 688, 92, 68, 32, 2, 8, 4, 1 }, - { 24, 464, 412, 24, 2, 8, 4, 1 }, - { 112, 184, 144, 28, 4, 4, 16, 1 }, - { 5776, 64, 32, 36, 2, 4, 16, 1 }, - { 1568, 64, 40, 36, 2, 8, 8, 1 }, - { 2920, 64, 64, 24, 4, 4, 16, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_nt = - { - { 3136, 64, 64, 36, 4, 4, 8, 0 }, - { 4096, 48, 32, 36, 4, 4, 8, 0 }, - { 688, 92, 68, 32, 5, 4, 4, 0 }, - { 24, 464, 412, 24, 6, 2, 8, 0 }, - { 112, 184, 144, 28, 6, 4, 4, 0 }, - { 5776, 64, 32, 36, 5, 4, 4, 0 }, - { 1568, 64, 40, 36, 4, 4, 8, 0 }, - { 2920, 64, 64, 24, 4, 4, 8, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = - { - { 3136, 64, 64, 36, 4, 4, 4, 1 }, - { 4096, 48, 32, 36, 2, 2, 16, 1 }, - { 688, 92, 68, 32, 4, 4, 4, 1 }, - { 24, 464, 412, 24, 6, 2, 8, 1 }, - { 112, 184, 144, 28, 4, 2, 16, 1 }, - { 5776, 64, 32, 36, 4, 4, 4, 1 }, - { 1568, 64, 40, 36, 4, 4, 8, 1 }, - { 2920, 64, 64, 24, 4, 4, 4, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_t = - { - { 3136, 64, 64, 36, 5, 4, 4, 0 }, - { 4096, 48, 32, 36, 5, 4, 4, 0 }, - { 688, 92, 68, 32, 5, 4, 4, 0 }, - { 24, 464, 412, 24, 6, 2, 4, 0 }, - { 112, 184, 144, 28, 5, 4, 4, 0 }, - { 5776, 64, 32, 36, 5, 4, 4, 0 }, - { 1568, 64, 40, 36, 5, 4, 4, 0 }, - { 2920, 64, 64, 24, 6, 2, 4, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = - { - { 3136, 64, 64, 36, 4, 4, 16, 1 }, - { 4096, 48, 32, 36, 4, 4, 4, 1 }, - { 688, 92, 68, 32, 2, 8, 4, 1 }, - { 24, 464, 412, 24, 2, 8, 4, 1 }, - { 112, 184, 144, 28, 4, 4, 16, 1 }, - { 5776, 64, 32, 36, 2, 8, 8, 1 }, - { 1568, 64, 40, 36, 4, 4, 8, 1 }, - { 2920, 64, 64, 24, 4, 4, 16, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_nt = - { - { 3136, 64, 64, 36, 4, 4, 4, 0 }, - { 4096, 48, 32, 36, 4, 4, 4, 0 }, - { 688, 92, 68, 32, 4, 4, 4, 0 }, - { 24, 464, 412, 24, 4, 4, 4, 0 }, - { 112, 184, 144, 28, 4, 4, 4, 0 }, - { 5776, 64, 32, 36, 4, 4, 8, 0 }, - { 1568, 64, 40, 36, 4, 4, 4, 0 }, - { 2920, 64, 64, 24, 4, 4, 4, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = - { - { 3136, 64, 64, 36, 4, 4, 4, 1 }, - { 4096, 48, 32, 36, 4, 4, 4, 1 }, - { 688, 92, 68, 32, 4, 4, 4, 1 }, - { 24, 464, 412, 24, 2, 2, 16, 1 }, - { 112, 184, 144, 28, 4, 4, 4, 1 }, - { 5776, 64, 32, 36, 4, 4, 4, 1 }, - { 1568, 64, 40, 36, 4, 4, 4, 1 }, - { 2920, 64, 64, 24, 4, 4, 4, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_t = - { - { 3136, 64, 64, 36, 4, 4, 4, 0 }, - { 4096, 48, 32, 36, 4, 4, 4, 0 }, - { 688, 92, 68, 32, 4, 4, 4, 0 }, - { 24, 464, 412, 24, 4, 2, 8, 0 }, - { 112, 184, 144, 28, 4, 4, 4, 0 }, - { 5776, 64, 32, 36, 4, 4, 4, 0 }, - { 1568, 64, 40, 36, 4, 4, 4, 0 }, - { 2920, 64, 64, 24, 4, 4, 4, 0 } - }; + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = { + {3136, 64, 64, 36, 4, 4, 16, 1}, {4096, 48, 32, 36, 4, 4, 4, 1}, {688, 92, 68, 32, 2, 8, 4, 1}, + {24, 464, 412, 24, 2, 8, 4, 1}, {112, 184, 144, 28, 4, 4, 16, 1}, {5776, 64, 32, 36, 2, 4, 16, 1}, + {1568, 64, 40, 36, 2, 8, 8, 1}, {2920, 64, 64, 24, 4, 4, 16, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_nt = { + {3136, 64, 64, 36, 4, 4, 8, 0}, {4096, 48, 32, 36, 4, 4, 8, 0}, {688, 92, 68, 32, 5, 4, 4, 0}, + {24, 464, 412, 24, 6, 2, 8, 0}, {112, 184, 144, 28, 6, 4, 4, 0}, {5776, 64, 32, 36, 5, 4, 4, 0}, + {1568, 64, 40, 36, 4, 4, 8, 0}, {2920, 64, 64, 24, 4, 4, 8, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = { + {3136, 64, 64, 36, 4, 4, 4, 1}, {4096, 48, 32, 36, 2, 2, 16, 1}, {688, 92, 68, 32, 4, 4, 4, 1}, + {24, 464, 412, 24, 6, 2, 8, 1}, {112, 184, 144, 28, 4, 2, 16, 1}, {5776, 64, 32, 36, 4, 4, 4, 1}, + {1568, 64, 40, 36, 4, 4, 8, 1}, {2920, 64, 64, 24, 4, 4, 4, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_t = { + {3136, 64, 64, 36, 5, 4, 4, 0}, {4096, 48, 32, 36, 5, 4, 4, 0}, {688, 92, 68, 32, 5, 4, 4, 0}, + {24, 464, 412, 24, 6, 2, 4, 0}, {112, 184, 144, 28, 5, 4, 4, 0}, {5776, 64, 32, 36, 5, 4, 4, 0}, + {1568, 64, 40, 36, 5, 4, 4, 0}, {2920, 64, 64, 24, 6, 2, 4, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = { + {3136, 64, 64, 36, 4, 4, 16, 1}, {4096, 48, 32, 36, 4, 4, 4, 1}, {688, 92, 68, 32, 2, 8, 4, 1}, + {24, 464, 412, 24, 2, 8, 4, 1}, {112, 184, 144, 28, 4, 4, 16, 1}, {5776, 64, 32, 36, 2, 8, 8, 1}, + {1568, 64, 40, 36, 4, 4, 8, 1}, {2920, 64, 64, 24, 4, 4, 16, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_nt = { + {3136, 64, 64, 36, 4, 4, 4, 0}, {4096, 48, 32, 36, 4, 4, 4, 0}, {688, 92, 68, 32, 4, 4, 4, 0}, + {24, 464, 412, 24, 4, 4, 4, 0}, {112, 184, 144, 28, 4, 4, 4, 0}, {5776, 64, 32, 36, 4, 4, 8, 0}, + {1568, 64, 40, 36, 4, 4, 4, 0}, {2920, 64, 64, 24, 4, 4, 4, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = { + {3136, 64, 64, 36, 4, 4, 4, 1}, {4096, 48, 32, 36, 4, 4, 4, 1}, {688, 92, 68, 32, 4, 4, 4, 1}, + {24, 464, 412, 24, 2, 2, 16, 1}, {112, 184, 144, 28, 4, 4, 4, 1}, {5776, 64, 32, 36, 4, 4, 4, 1}, + {1568, 64, 40, 36, 4, 4, 4, 1}, {2920, 64, 64, 24, 4, 4, 4, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_t = { + {3136, 64, 64, 36, 4, 4, 4, 0}, {4096, 48, 32, 36, 4, 4, 4, 0}, {688, 92, 68, 32, 4, 4, 4, 0}, + {24, 464, 412, 24, 4, 2, 8, 0}, {112, 184, 144, 28, 4, 4, 4, 0}, {5776, 64, 32, 36, 4, 4, 4, 0}, + {1568, 64, 40, 36, 4, 4, 4, 0}, {2920, 64, 64, 24, 4, 4, 4, 0}}; const bool adj_lhs = info.adj_lhs(); const bool adj_rhs = info.adj_rhs(); @@ -185,17 +133,17 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f32(unsigned const MatMulNativeConfigsMatrix *configs_best_to_use = nullptr; const MatMulNativeConfigsMatrix *configs_fallback_to_use = nullptr; - if((adj_lhs == false) && (adj_rhs == false)) + if ((adj_lhs == false) && (adj_rhs == false)) { configs_best_to_use = &configs_mnkb_best_nt_nt; configs_fallback_to_use = &configs_mnkb_fallback_nt_nt; } - else if((adj_lhs == false) && (adj_rhs == true)) + else if ((adj_lhs == false) && (adj_rhs == true)) { configs_best_to_use = &configs_mnkb_best_nt_t; configs_fallback_to_use = &configs_mnkb_fallback_nt_t; } - else if((adj_lhs == true) && (adj_rhs == false)) + else if ((adj_lhs == true) && (adj_rhs == false)) { configs_best_to_use = &configs_mnkb_best_t_nt; configs_fallback_to_use = &configs_mnkb_fallback_t_nt; @@ -209,108 +157,51 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f32(unsigned MatMulKernelInfo desc0 = find_info(*configs_best_to_use, adj_lhs, adj_rhs, m, n, k, b); MatMulKernelInfo desc1 = find_info(*configs_fallback_to_use, adj_lhs, adj_rhs, m, n, k, b); - return select_info(desc0, - desc1, - m, n, k, b, DataType::F32, rhs_lock_padding); + return select_info(desc0, desc1, m, n, k, b, DataType::F32, rhs_lock_padding); } -MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) +MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f16( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) { - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = - { - { 3136, 64, 64, 36, 4, 4, 16, 1 }, - { 4096, 48, 32, 36, 4, 4, 8, 1 }, - { 688, 92, 68, 32, 4, 4, 16, 1 }, - { 24, 464, 412, 24, 4, 4, 4, 1 }, - { 112, 184, 144, 28, 4, 4, 16, 1 }, - { 5776, 64, 32, 36, 4, 4, 8, 1 }, - { 1568, 64, 40, 36, 4, 4, 8, 1 }, - { 2920, 64, 64, 24, 4, 4, 16, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_nt = - { - { 3136, 64, 64, 36, 6, 4, 8, 0 }, - { 4096, 48, 32, 36, 6, 4, 8, 0 }, - { 688, 92, 68, 32, 6, 4, 8, 0 }, - { 24, 464, 412, 24, 4, 4, 8, 0 }, - { 112, 184, 144, 28, 6, 4, 8, 0 }, - { 5776, 64, 32, 36, 6, 4, 8, 0 }, - { 1568, 64, 40, 36, 6, 4, 8, 0 }, - { 2920, 64, 64, 24, 6, 4, 8, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = - { - { 3136, 64, 64, 36, 6, 4, 8, 1 }, - { 4096, 48, 32, 36, 6, 4, 8, 1 }, - { 688, 92, 68, 32, 4, 4, 4, 1 }, - { 24, 464, 412, 24, 6, 2, 4, 1 }, - { 112, 184, 144, 28, 4, 2, 16, 1 }, - { 5776, 64, 32, 36, 6, 4, 8, 1 }, - { 1568, 64, 40, 36, 6, 4, 8, 1 }, - { 2920, 64, 64, 24, 6, 4, 8, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_t = - { - { 3136, 64, 64, 36, 6, 2, 16, 0 }, - { 4096, 48, 32, 36, 5, 4, 8, 0 }, - { 688, 92, 68, 32, 6, 2, 16, 0 }, - { 24, 464, 412, 24, 6, 2, 16, 0 }, - { 112, 184, 144, 28, 6, 2, 16, 0 }, - { 5776, 64, 32, 36, 5, 4, 8, 0 }, - { 1568, 64, 40, 36, 5, 4, 8, 0 }, - { 2920, 64, 64, 24, 6, 2, 16, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = - { - { 3136, 64, 64, 36, 4, 4, 16, 1 }, - { 4096, 48, 32, 36, 4, 4, 4, 1 }, - { 688, 92, 68, 32, 4, 4, 4, 1 }, - { 24, 464, 412, 24, 4, 4, 4, 1 }, - { 112, 184, 144, 28, 4, 4, 4, 1 }, - { 5776, 64, 32, 36, 4, 4, 4, 1 }, - { 1568, 64, 40, 36, 4, 4, 4, 1 }, - { 2920, 64, 64, 24, 4, 4, 4, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_nt = - { - { 3136, 64, 64, 36, 4, 4, 4, 0 }, - { 4096, 48, 32, 36, 4, 4, 4, 0 }, - { 688, 92, 68, 32, 4, 4, 4, 0 }, - { 24, 464, 412, 24, 4, 4, 4, 0 }, - { 112, 184, 144, 28, 4, 4, 4, 0 }, - { 5776, 64, 32, 36, 4, 4, 4, 0 }, - { 1568, 64, 40, 36, 4, 4, 4, 0 }, - { 2920, 64, 64, 24, 4, 4, 4, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = - { - { 3136, 64, 64, 36, 4, 4, 16, 1 }, - { 4096, 48, 32, 36, 4, 4, 8, 1 }, - { 688, 92, 68, 32, 4, 4, 4, 1 }, - { 24, 464, 412, 24, 4, 2, 8, 1 }, - { 112, 184, 144, 28, 4, 2, 16, 1 }, - { 5776, 64, 32, 36, 4, 4, 16, 1 }, - { 1568, 64, 40, 36, 4, 4, 8, 1 }, - { 2920, 64, 64, 24, 4, 4, 16, 1 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_t = - { - { 3136, 64, 64, 36, 4, 4, 8, 0 }, - { 4096, 48, 32, 36, 4, 4, 8, 0 }, - { 688, 92, 68, 32, 4, 4, 8, 0 }, - { 24, 464, 412, 24, 4, 4, 8, 0 }, - { 112, 184, 144, 28, 4, 4, 8, 0 }, - { 5776, 64, 32, 36, 4, 4, 8, 0 }, - { 1568, 64, 40, 36, 4, 4, 8, 0 }, - { 2920, 64, 64, 24, 4, 4, 8, 0 } - }; + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = { + {3136, 64, 64, 36, 4, 4, 16, 1}, {4096, 48, 32, 36, 4, 4, 8, 1}, {688, 92, 68, 32, 4, 4, 16, 1}, + {24, 464, 412, 24, 4, 4, 4, 1}, {112, 184, 144, 28, 4, 4, 16, 1}, {5776, 64, 32, 36, 4, 4, 8, 1}, + {1568, 64, 40, 36, 4, 4, 8, 1}, {2920, 64, 64, 24, 4, 4, 16, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_nt = { + {3136, 64, 64, 36, 6, 4, 8, 0}, {4096, 48, 32, 36, 6, 4, 8, 0}, {688, 92, 68, 32, 6, 4, 8, 0}, + {24, 464, 412, 24, 4, 4, 8, 0}, {112, 184, 144, 28, 6, 4, 8, 0}, {5776, 64, 32, 36, 6, 4, 8, 0}, + {1568, 64, 40, 36, 6, 4, 8, 0}, {2920, 64, 64, 24, 6, 4, 8, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = { + {3136, 64, 64, 36, 6, 4, 8, 1}, {4096, 48, 32, 36, 6, 4, 8, 1}, {688, 92, 68, 32, 4, 4, 4, 1}, + {24, 464, 412, 24, 6, 2, 4, 1}, {112, 184, 144, 28, 4, 2, 16, 1}, {5776, 64, 32, 36, 6, 4, 8, 1}, + {1568, 64, 40, 36, 6, 4, 8, 1}, {2920, 64, 64, 24, 6, 4, 8, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_nt_t = { + {3136, 64, 64, 36, 6, 2, 16, 0}, {4096, 48, 32, 36, 5, 4, 8, 0}, {688, 92, 68, 32, 6, 2, 16, 0}, + {24, 464, 412, 24, 6, 2, 16, 0}, {112, 184, 144, 28, 6, 2, 16, 0}, {5776, 64, 32, 36, 5, 4, 8, 0}, + {1568, 64, 40, 36, 5, 4, 8, 0}, {2920, 64, 64, 24, 6, 2, 16, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = { + {3136, 64, 64, 36, 4, 4, 16, 1}, {4096, 48, 32, 36, 4, 4, 4, 1}, {688, 92, 68, 32, 4, 4, 4, 1}, + {24, 464, 412, 24, 4, 4, 4, 1}, {112, 184, 144, 28, 4, 4, 4, 1}, {5776, 64, 32, 36, 4, 4, 4, 1}, + {1568, 64, 40, 36, 4, 4, 4, 1}, {2920, 64, 64, 24, 4, 4, 4, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_nt = { + {3136, 64, 64, 36, 4, 4, 4, 0}, {4096, 48, 32, 36, 4, 4, 4, 0}, {688, 92, 68, 32, 4, 4, 4, 0}, + {24, 464, 412, 24, 4, 4, 4, 0}, {112, 184, 144, 28, 4, 4, 4, 0}, {5776, 64, 32, 36, 4, 4, 4, 0}, + {1568, 64, 40, 36, 4, 4, 4, 0}, {2920, 64, 64, 24, 4, 4, 4, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = { + {3136, 64, 64, 36, 4, 4, 16, 1}, {4096, 48, 32, 36, 4, 4, 8, 1}, {688, 92, 68, 32, 4, 4, 4, 1}, + {24, 464, 412, 24, 4, 2, 8, 1}, {112, 184, 144, 28, 4, 2, 16, 1}, {5776, 64, 32, 36, 4, 4, 16, 1}, + {1568, 64, 40, 36, 4, 4, 8, 1}, {2920, 64, 64, 24, 4, 4, 16, 1}}; + + const MatMulNativeConfigsMatrix configs_mnkb_fallback_t_t = { + {3136, 64, 64, 36, 4, 4, 8, 0}, {4096, 48, 32, 36, 4, 4, 8, 0}, {688, 92, 68, 32, 4, 4, 8, 0}, + {24, 464, 412, 24, 4, 4, 8, 0}, {112, 184, 144, 28, 4, 4, 8, 0}, {5776, 64, 32, 36, 4, 4, 8, 0}, + {1568, 64, 40, 36, 4, 4, 8, 0}, {2920, 64, 64, 24, 4, 4, 8, 0}}; const bool adj_lhs = info.adj_lhs(); const bool adj_rhs = info.adj_rhs(); @@ -318,17 +209,17 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f16(unsigned const MatMulNativeConfigsMatrix *configs_best_to_use = nullptr; const MatMulNativeConfigsMatrix *configs_fallback_to_use = nullptr; - if((adj_lhs == false) && (adj_rhs == false)) + if ((adj_lhs == false) && (adj_rhs == false)) { configs_best_to_use = &configs_mnkb_best_nt_nt; configs_fallback_to_use = &configs_mnkb_fallback_nt_nt; } - else if((adj_lhs == false) && (adj_rhs == true)) + else if ((adj_lhs == false) && (adj_rhs == true)) { configs_best_to_use = &configs_mnkb_best_nt_t; configs_fallback_to_use = &configs_mnkb_fallback_nt_t; } - else if((adj_lhs == true) && (adj_rhs == false)) + else if ((adj_lhs == true) && (adj_rhs == false)) { configs_best_to_use = &configs_mnkb_best_t_nt; configs_fallback_to_use = &configs_mnkb_fallback_t_nt; @@ -342,75 +233,46 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_f16(unsigned MatMulKernelInfo desc0 = find_info(*configs_best_to_use, adj_lhs, adj_rhs, m, n, k, b); MatMulKernelInfo desc1 = find_info(*configs_fallback_to_use, adj_lhs, adj_rhs, m, n, k, b); - return select_info(desc0, - desc1, - m, n, k, b, DataType::F16, rhs_lock_padding); + return select_info(desc0, desc1, m, n, k, b, DataType::F16, rhs_lock_padding); } -MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) +MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_u8( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info) { ARM_COMPUTE_UNUSED(rhs_lock_padding); - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = - { - { 3136, 64, 64, 36, 6, 4, 4, 0 }, - { 4096, 48, 32, 36, 6, 4, 4, 0 }, - { 688, 92, 68, 32, 2, 8, 4, 0 }, - { 24, 464, 412, 24, 4, 4, 4, 0 }, - { 112, 184, 144, 28, 6, 4, 4, 0 }, - { 5776, 64, 32, 36, 6, 4, 4, 0 }, - { 1568, 64, 40, 36, 6, 4, 4, 0 }, - { 2920, 64, 64, 24, 5, 4, 4, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = - { - { 3136, 64, 64, 36, 4, 4, 16, 0 }, - { 4096, 48, 32, 36, 4, 4, 16, 0 }, - { 688, 92, 68, 32, 4, 4, 16, 0 }, - { 24, 464, 412, 24, 6, 2, 16, 0 }, - { 112, 184, 144, 28, 4, 4, 16, 0 }, - { 5776, 64, 32, 36, 4, 4, 16, 0 }, - { 1568, 64, 40, 36, 6, 4, 4, 0 }, - { 2920, 64, 64, 24, 4, 4, 16, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = - { - { 3136, 64, 64, 36, 4, 4, 8, 0 }, - { 4096, 48, 32, 36, 4, 4, 8, 0 }, - { 688, 92, 68, 32, 4, 4, 4, 0 }, - { 24, 464, 412, 24, 4, 4, 4, 0 }, - { 112, 184, 144, 28, 4, 4, 8, 0 }, - { 5776, 64, 32, 36, 4, 4, 8, 0 }, - { 1568, 64, 40, 36, 4, 4, 8, 0 }, - { 2920, 64, 64, 24, 4, 4, 8, 0 } - }; - - const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = - { - { 3136, 64, 64, 36, 4, 2, 16, 0 }, - { 4096, 48, 32, 36, 4, 4, 4, 0 }, - { 688, 92, 68, 32, 4, 4, 8, 0 }, - { 24, 464, 412, 24, 4, 2, 16, 0 }, - { 112, 184, 144, 28, 4, 2, 16, 0 }, - { 5776, 64, 32, 36, 4, 4, 4, 0 }, - { 1568, 64, 40, 36, 4, 4, 8, 0 }, - { 2920, 64, 64, 24, 4, 2, 16, 0 } - }; + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_nt = { + {3136, 64, 64, 36, 6, 4, 4, 0}, {4096, 48, 32, 36, 6, 4, 4, 0}, {688, 92, 68, 32, 2, 8, 4, 0}, + {24, 464, 412, 24, 4, 4, 4, 0}, {112, 184, 144, 28, 6, 4, 4, 0}, {5776, 64, 32, 36, 6, 4, 4, 0}, + {1568, 64, 40, 36, 6, 4, 4, 0}, {2920, 64, 64, 24, 5, 4, 4, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_nt_t = { + {3136, 64, 64, 36, 4, 4, 16, 0}, {4096, 48, 32, 36, 4, 4, 16, 0}, {688, 92, 68, 32, 4, 4, 16, 0}, + {24, 464, 412, 24, 6, 2, 16, 0}, {112, 184, 144, 28, 4, 4, 16, 0}, {5776, 64, 32, 36, 4, 4, 16, 0}, + {1568, 64, 40, 36, 6, 4, 4, 0}, {2920, 64, 64, 24, 4, 4, 16, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_nt = { + {3136, 64, 64, 36, 4, 4, 8, 0}, {4096, 48, 32, 36, 4, 4, 8, 0}, {688, 92, 68, 32, 4, 4, 4, 0}, + {24, 464, 412, 24, 4, 4, 4, 0}, {112, 184, 144, 28, 4, 4, 8, 0}, {5776, 64, 32, 36, 4, 4, 8, 0}, + {1568, 64, 40, 36, 4, 4, 8, 0}, {2920, 64, 64, 24, 4, 4, 8, 0}}; + + const MatMulNativeConfigsMatrix configs_mnkb_best_t_t = { + {3136, 64, 64, 36, 4, 2, 16, 0}, {4096, 48, 32, 36, 4, 4, 4, 0}, {688, 92, 68, 32, 4, 4, 8, 0}, + {24, 464, 412, 24, 4, 2, 16, 0}, {112, 184, 144, 28, 4, 2, 16, 0}, {5776, 64, 32, 36, 4, 4, 4, 0}, + {1568, 64, 40, 36, 4, 4, 8, 0}, {2920, 64, 64, 24, 4, 2, 16, 0}}; const bool adj_lhs = info.adj_lhs(); const bool adj_rhs = info.adj_rhs(); - if((adj_lhs == false) && (adj_rhs == false)) + if ((adj_lhs == false) && (adj_rhs == false)) { return find_info(configs_mnkb_best_nt_nt, adj_lhs, adj_rhs, m, n, k, b); } - else if((adj_lhs == false) && (adj_rhs == true)) + else if ((adj_lhs == false) && (adj_rhs == true)) { return find_info(configs_mnkb_best_nt_t, adj_lhs, adj_rhs, m, n, k, b); } - else if((adj_lhs == true) && (adj_rhs == false)) + else if ((adj_lhs == true) && (adj_rhs == false)) { return find_info(configs_mnkb_best_t_nt, adj_lhs, adj_rhs, m, n, k, b); } @@ -419,5 +281,5 @@ MatMulKernelInfo ClMatMulNativeDefaultConfigValhall::configure_G710_u8(unsigned return find_info(configs_mnkb_best_t_t, adj_lhs, adj_rhs, m, n, k, b); } } -} // namespace opencl +} // namespace cl_matmul } // namespace arm_compute diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.h b/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.h index fe167d18dd..6b39db6a3f 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.h +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeDefaultConfigValhall.h @@ -44,10 +44,13 @@ public: MatMulKernelInfo configure(const ITensorInfo *lhs, const ITensorInfo *rhs, const MatMulInfo &info) override; private: - MatMulKernelInfo configure_G710_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); - MatMulKernelInfo configure_G710_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); - MatMulKernelInfo configure_G710_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); + MatMulKernelInfo configure_G710_f32( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); + MatMulKernelInfo configure_G710_f16( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); + MatMulKernelInfo configure_G710_u8( + unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool rhs_lock_padding, const MatMulInfo &info); }; -} // namespace opencl +} // namespace cl_matmul } // namespace arm_compute #endif /* SRC_RUNTIME_HEURISTICS_MATMUL_NATIVE_CLMATMULNATIVEDEFAULTCONFIGVALHALL */ diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp index 1e06e84d4d..89cad30214 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.cpp @@ -26,6 +26,7 @@ #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/TensorShape.h" + #include "src/gpu/cl/kernels/ClMatMulNativeKernel.h" #include <limits> @@ -37,22 +38,32 @@ namespace cl_matmul { MatMulKernelInfo select_info(const MatMulKernelInfo &info0, const MatMulKernelInfo &info1, - unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type, bool rhs_lock_padding) + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b, + DataType data_type, + bool rhs_lock_padding) { - ARM_COMPUTE_ERROR_ON_MSG(info1.export_rhs_to_cl_image == true, "The fallback MatMul configuration cannot have export_to_cl_image = true"); - ARM_COMPUTE_ERROR_ON_MSG(info0.adj_lhs != info1.adj_lhs, "The MatMul configurations must have the same adj_lhs value"); - ARM_COMPUTE_ERROR_ON_MSG(info0.adj_rhs != info1.adj_rhs, "The MatMul configurations must have the same adj_rhs value"); + ARM_COMPUTE_ERROR_ON_MSG(info1.export_rhs_to_cl_image == true, + "The fallback MatMul configuration cannot have export_to_cl_image = true"); + ARM_COMPUTE_ERROR_ON_MSG(info0.adj_lhs != info1.adj_lhs, + "The MatMul configurations must have the same adj_lhs value"); + ARM_COMPUTE_ERROR_ON_MSG(info0.adj_rhs != info1.adj_rhs, + "The MatMul configurations must have the same adj_rhs value"); const bool adj_lhs = info0.adj_lhs; const bool adj_rhs = info0.adj_rhs; - TensorInfo lhs_info = !adj_lhs ? TensorInfo(TensorShape(k, m, b), 1, data_type) : TensorInfo(TensorShape(m, k, b), 1, data_type); - TensorInfo rhs_info = !adj_rhs ? TensorInfo(TensorShape(n, k, b), 1, data_type) : TensorInfo(TensorShape(k, n, b), 1, data_type); + TensorInfo lhs_info = + !adj_lhs ? TensorInfo(TensorShape(k, m, b), 1, data_type) : TensorInfo(TensorShape(m, k, b), 1, data_type); + TensorInfo rhs_info = + !adj_rhs ? TensorInfo(TensorShape(n, k, b), 1, data_type) : TensorInfo(TensorShape(k, n, b), 1, data_type); TensorInfo dst_info; - if(rhs_lock_padding == false) + if (rhs_lock_padding == false) { - if(bool(opencl::kernels::ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &dst_info, info0))) + if (bool(opencl::kernels::ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &dst_info, info0))) { return info0; } @@ -67,7 +78,13 @@ MatMulKernelInfo select_info(const MatMulKernelInfo &info0, } } -MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lhs, bool adj_rhs, unsigned int m, unsigned int n, unsigned int k, unsigned int b) +MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, + bool adj_lhs, + bool adj_rhs, + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b) { size_t min_acc = std::numeric_limits<size_t>::max(); size_t min_idx = 0; @@ -76,12 +93,13 @@ MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lh const size_t num_rows = configs.size(); const size_t num_cols = configs[0].size(); - ARM_COMPUTE_ERROR_ON_MSG(num_cols != 8U, "The entry should have 8 integer values representing: M, N, K, B, M0, N0. K0, IMG_RHS"); + ARM_COMPUTE_ERROR_ON_MSG(num_cols != 8U, + "The entry should have 8 integer values representing: M, N, K, B, M0, N0. K0, IMG_RHS"); ARM_COMPUTE_UNUSED(num_cols); // Find nearest GeMM workload // Note: the workload does not depend on the K dimension - for(size_t y = 0; y < num_rows; ++y) + for (size_t y = 0; y < num_rows; ++y) { size_t mc0 = static_cast<size_t>(configs[y][0]); size_t nc0 = static_cast<size_t>(configs[y][1]); @@ -94,7 +112,7 @@ MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lh acc += (k - kc0) * (k - kc0); acc += (b - bc0) * (b - bc0); acc = std::sqrt(acc); - if(acc < min_acc) + if (acc < min_acc) { min_acc = acc; min_idx = y; diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.h b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.h index 3881617558..a114fffa68 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.h +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeHelpers.h @@ -52,7 +52,12 @@ using MatMulNativeConfigsMatrix = std::vector<std::vector<int32_t>>; */ MatMulKernelInfo select_info(const MatMulKernelInfo &info0, const MatMulKernelInfo &info1, - unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type, bool rhs_lock_padding); + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b, + DataType data_type, + bool rhs_lock_padding); /** Find the preferred configurations for the MatMul Native kernel using the MatMulNativeConfigsMatrix provided by the user * @@ -66,7 +71,13 @@ MatMulKernelInfo select_info(const MatMulKernelInfo &info0, * * @return @ref MatMulKernelInfo */ -MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, bool adj_lhs, bool adj_rhs, unsigned int m, unsigned int n, unsigned int k, unsigned int b); +MatMulKernelInfo find_info(const MatMulNativeConfigsMatrix &configs, + bool adj_lhs, + bool adj_rhs, + unsigned int m, + unsigned int n, + unsigned int k, + unsigned int b); } // namespace cl_matmul } // namespace arm_compute #endif /* SRC_RUNTIME_HEURISTICS_MATMUL_NATIVE_CLMATMULNATIVEHELPERS */ diff --git a/src/runtime/heuristics/matmul_native/ClMatMulNativeKernelConfig.h b/src/runtime/heuristics/matmul_native/ClMatMulNativeKernelConfig.h index a2dbfc7dd5..b10018a6d2 100644 --- a/src/runtime/heuristics/matmul_native/ClMatMulNativeKernelConfig.h +++ b/src/runtime/heuristics/matmul_native/ClMatMulNativeKernelConfig.h @@ -45,7 +45,7 @@ public: */ static std::unique_ptr<IClMatMulNativeKernelConfig> create(GPUTarget gpu) { - switch(get_arch_from_target(gpu)) + switch (get_arch_from_target(gpu)) { case GPUTarget::MIDGARD: case GPUTarget::BIFROST: @@ -56,6 +56,6 @@ public: } } }; -} // namespace opencl +} // namespace cl_matmul } // namespace arm_compute #endif /* SRC_RUNTIME_HEURISTICS_MATMUL_NATIVE_CLMATMULNATIVEKERNELCONFIG */ diff --git a/src/runtime/heuristics/matmul_native/IClMatMulNativeKernelConfig.h b/src/runtime/heuristics/matmul_native/IClMatMulNativeKernelConfig.h index 4f548bd01d..b9b091100c 100644 --- a/src/runtime/heuristics/matmul_native/IClMatMulNativeKernelConfig.h +++ b/src/runtime/heuristics/matmul_native/IClMatMulNativeKernelConfig.h @@ -28,6 +28,7 @@ #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" #include "arm_compute/function_info/MatMulInfo.h" + #include "src/core/common/Macros.h" namespace arm_compute @@ -53,8 +54,7 @@ public: * @param[in] func_int8 Function to call for matmul native Int8 (QASYMM8, QASYMM8_SIGNED, QSYMM8_PER_CHANNEL) * */ - ClMatMulNativeConfigArray(T func_f32, T func_f16, T func_int8) - : _configs{ func_f32, func_f16, func_int8 } + ClMatMulNativeConfigArray(T func_f32, T func_f16, T func_int8) : _configs{func_f32, func_f16, func_int8} { } @@ -66,7 +66,7 @@ public: */ T get_function(DataType data_type) { - switch(data_type) + switch (data_type) { case DataType::F32: return _configs.at(DT_F32); @@ -93,8 +93,7 @@ public: * * @param[in] arch GPU target */ - IClMatMulNativeKernelConfig(GPUTarget arch) - : _target(arch) + IClMatMulNativeKernelConfig(GPUTarget arch) : _target(arch) { } ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(IClMatMulNativeKernelConfig); |