From 0d548048a60035349b90b903924fbf38c74796b9 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Thu, 3 Oct 2019 15:12:09 +0100 Subject: COMPMID-2572: Update the heuristic in CLGEMM for FP16 Change-Id: Ia7f248d72bf2690a8a4dae3f2e2afc983109bd6e Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/2035 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Giorgio Arena --- ...MMReshapedOnlyRHSKernelConfigurationBifrost.cpp | 112 +++++++++++++++++---- 1 file changed, 95 insertions(+), 17 deletions(-) (limited to 'src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp') diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp index 9f3fc3aae7..5955bac384 100644 --- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp +++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationBifrost.cpp @@ -42,9 +42,6 @@ CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::CLGEMMReshapedOnlyRHSKernelConf std::pair CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type) { - ARM_COMPUTE_ERROR_ON(data_type != DataType::F32 && data_type != DataType::QASYMM8); - ARM_COMPUTE_UNUSED(data_type); - using ConfigurationFunctionExecutorPtr = std::pair (CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b); @@ -52,6 +49,7 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi static std::map gemm_configs_G51 = { { DataType::F32, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G51_f32 }, + { DataType::F16, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G51_f16 }, { DataType::QASYMM8, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G51_u8 } }; @@ -59,6 +57,7 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi static std::map gemm_configs_G76 = { { DataType::F32, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G76_f32 }, + { DataType::F16, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G76_f16 }, { DataType::QASYMM8, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G76_u8 } }; @@ -66,17 +65,39 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi static std::map gemm_configs_G7x = { { DataType::F32, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_f32 }, + { DataType::F16, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_f16 }, { DataType::QASYMM8, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_u8 } }; switch(_target) { case GPUTarget::G76: - return (this->*gemm_configs_G76[data_type])(m, n, k, b); + if (gemm_configs_G76.find(data_type) != gemm_configs_G76.end()) + { + return (this->*gemm_configs_G76[data_type])(m, n, k, b); + } + else + { + ARM_COMPUTE_ERROR("Not supported data type"); + } case GPUTarget::G51: - return (this->*gemm_configs_G51[data_type])(m, n, k, b); + if (gemm_configs_G51.find(data_type) != gemm_configs_G51.end()) + { + return (this->*gemm_configs_G51[data_type])(m, n, k, b); + } + else + { + ARM_COMPUTE_ERROR("Not supported data type"); + } default: - return (this->*gemm_configs_G7x[data_type])(m, n, k, b); + if (gemm_configs_G7x.find(data_type) != gemm_configs_G7x.end()) + { + return (this->*gemm_configs_G7x[data_type])(m, n, k, b); + } + else + { + ARM_COMPUTE_ERROR("Not supported data type"); + } } } @@ -89,12 +110,12 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi { if(n > 2048) { - const unsigned int h0 = std::max(n / 4, static_cast(1)); + const unsigned int h0 = std::max(n / 4, 1U); return configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true); } else { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); } } @@ -111,7 +132,7 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi if(m == 1) { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); } else @@ -128,7 +149,7 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi if(m == 1) { const unsigned int n0 = n < 1280? 2 : 4; - const unsigned int h0 = std::max(n / n0, static_cast(1)); + const unsigned int h0 = std::max(n / n0, 1U); return configure_lhs_rhs_info(m, n, 1, n0, 4, 1, h0, false, true, false, true); } else @@ -137,6 +158,63 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi } } +std::pair CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + ARM_COMPUTE_UNUSED(k); + ARM_COMPUTE_UNUSED(b); + + if(m == 1) + { + if(n > 2048) + { + const unsigned int h0 = std::max(n / 4, 1U); + return configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true); + } + else + { + const unsigned int h0 = std::max(n / 2, 1U); + return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); + } + } + else + { + return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 4, false, true, false, true); + } +} + +std::pair CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G76_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + ARM_COMPUTE_UNUSED(k); + ARM_COMPUTE_UNUSED(b); + + if(m == 1) + { + const unsigned int h0 = std::max(n / 2, 1U); + return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true); + } + else + { + return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, true); + } +} + +std::pair CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G51_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + ARM_COMPUTE_UNUSED(k); + ARM_COMPUTE_UNUSED(b); + + if(m == 1) + { + const unsigned int n0 = n < 1280? 2 : 4; + const unsigned int h0 = std::max(n / n0, 1U); + return configure_lhs_rhs_info(m, n, 1, n0, 8, 1, h0, false, true, false, true); + } + else + { + return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, true); + } +} + std::pair CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b) { ARM_COMPUTE_UNUSED(k); @@ -146,12 +224,12 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi { if(m == 1) { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, false, true, false, true); } else { - const unsigned int h0 = std::max(n / 4, static_cast(1)); + const unsigned int h0 = std::max(n / 4, 1U); return configure_lhs_rhs_info(m, n, 4, 4, 16, 1, h0, false, true, false, true); } } @@ -159,12 +237,12 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi { if(m == 1) { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, h0, false, true, false, true); } else { - const unsigned int h0 = std::max(n / 4, static_cast(1)); + const unsigned int h0 = std::max(n / 4, 1U); return configure_lhs_rhs_info(m, n, 2, 2, 16, 1, h0, false, true, false, true); } } @@ -177,7 +255,7 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi if(m == 1) { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, false, true, false, true); } else @@ -193,12 +271,12 @@ std::pair CLGEMMReshapedOnlyRHSKernelConfi if(m == 1) { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, h0, false, true, false, true); } else { - const unsigned int h0 = std::max(n / 2, static_cast(1)); + const unsigned int h0 = std::max(n / 2, 1U); return configure_lhs_rhs_info(m, n, 4, 2, 16, 1, h0, false, true, false, true); } } -- cgit v1.2.1