From db63b9c431264c9ef612e69a66b13a07b8f54786 Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Thu, 17 Jan 2019 09:47:04 +0000 Subject: COMPMID-1698: Implementing CLGEMMLowpMatrixMultiplyReshapedKernel Change-Id: Ia4db21b394a0b9235393202ce3c00b11cceb94ea Reviewed-on: https://review.mlplatform.org/568 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio --- .../CLGEMMReshapedConfigurationBifrost.cpp | 156 ++++++++++++--------- 1 file changed, 93 insertions(+), 63 deletions(-) (limited to 'src/runtime/CL/gemm_reshaped') diff --git a/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp b/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp index 079a52e61c..cd97849712 100644 --- a/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp +++ b/src/runtime/CL/gemm_reshaped/CLGEMMReshapedConfigurationBifrost.cpp @@ -32,18 +32,62 @@ namespace arm_compute { namespace cl_gemm { +namespace +{ +std::pair configure_gemm_reshaped(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, + bool lhs_interleave, bool rhs_interleave) +{ + GEMMLHSMatrixInfo lhs_info; + GEMMRHSMatrixInfo rhs_info; + + // Configure GEMMLHSMatrixInfo + lhs_info.m0 = m0; + lhs_info.k0 = k0; + lhs_info.v0 = ((m / (lhs_info.m0 * v0)) == 0) ? 1 : v0; + lhs_info.interleave = lhs_interleave; + lhs_info.transpose = false; + + // Configure GEMMRHSMatrixInfo + rhs_info.n0 = n0; + rhs_info.k0 = lhs_info.k0; + rhs_info.h0 = ((n / (rhs_info.n0 * h0)) == 0) ? 1 : h0; + rhs_info.interleave = rhs_interleave; + rhs_info.transpose = true; + + return std::make_pair(lhs_info, rhs_info); +} + +} // namespace + std::pair CLGEMMReshapedConfigurationBifrost::configure(unsigned int m, unsigned int n, unsigned int k, unsigned int b, DataType data_type) { - ARM_COMPUTE_ERROR_ON(data_type != DataType::F32); + ARM_COMPUTE_ERROR_ON(data_type != DataType::F32 && data_type != DataType::QASYMM8); ARM_COMPUTE_UNUSED(data_type); const GPUTarget gpu_target = CLScheduler::get().target(); + + using ConfigurationFunctionExecutorPtr = std::pair (CLGEMMReshapedConfigurationBifrost::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b); + + // Configurations for Mali-G76 + static std::map gemm_reshaped_configs_G76 = + { + { DataType::F32, &CLGEMMReshapedConfigurationBifrost::configure_G76_f32 }, + { DataType::QASYMM8, &CLGEMMReshapedConfigurationBifrost::configure_G76_u8 } + }; + + // Configurations for Mali-G7x + static std::map gemm_reshaped_configs_G7x = + { + { DataType::F32, &CLGEMMReshapedConfigurationBifrost::configure_G7x_f32 }, + { DataType::QASYMM8, &CLGEMMReshapedConfigurationBifrost::configure_G7x_u8 } + }; + switch(gpu_target) { case GPUTarget::G76: - return configure_G76_f32(m, n, k, b); + return (this->*gemm_reshaped_configs_G76[data_type])(m, n, k, b); default: - return configure_G7x_f32(m, n, k, b); + return (this->*gemm_reshaped_configs_G7x[data_type])(m, n, k, b); } } @@ -52,43 +96,43 @@ std::pair CLGEMMReshapedConfigurationBifro ARM_COMPUTE_UNUSED(k); ARM_COMPUTE_UNUSED(b); - GEMMLHSMatrixInfo lhs_info; - GEMMRHSMatrixInfo rhs_info; - if(n <= 4) { - // Configure GEMMLHSMatrixInfo - lhs_info.m0 = 4; - lhs_info.k0 = 8; - lhs_info.v0 = lhs_info.m0 * 16 < m ? 2 : 16; - lhs_info.interleave = true; - lhs_info.transpose = false; - - // Configure GEMMRHSMatrixInfo - rhs_info.n0 = 2; - rhs_info.k0 = lhs_info.k0; - rhs_info.h0 = rhs_info.n0 * 16 < n ? 2 : 16; - rhs_info.interleave = false; - rhs_info.transpose = true; + return configure_gemm_reshaped(m, n, 4, 2, 8, 16, 16, true, false); } else { - // Configure GEMMLHSMatrixInfo - lhs_info.m0 = 5; - lhs_info.k0 = 4; - lhs_info.v0 = lhs_info.m0 * 2 < m ? 1 : 2; - lhs_info.interleave = false; - lhs_info.transpose = false; - - // Configure GEMMRHSMatrixInfo - rhs_info.n0 = 4; - rhs_info.k0 = lhs_info.k0; - rhs_info.h0 = rhs_info.n0 * 16 < n ? 2 : 16; - rhs_info.interleave = true; - rhs_info.transpose = true; + return configure_gemm_reshaped(m, n, 5, 4, 4, 2, 16, false, true); } +} - return std::make_pair(lhs_info, rhs_info); +std::pair CLGEMMReshapedConfigurationBifrost::configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + ARM_COMPUTE_UNUSED(k); + ARM_COMPUTE_UNUSED(b); + + if(dot8_supported(CLKernelLibrary::get().get_device())) + { + if(n <= 4) + { + return configure_gemm_reshaped(m, n, 4, 2, 16, 2, 2, true, false); + } + else + { + return configure_gemm_reshaped(m, n, 4, 4, 16, 2, 2, true, false); + } + } + else + { + if(n <= 4) + { + return configure_gemm_reshaped(m, n, 4, 2, 8, 2, 2, true, false); + } + else + { + return configure_gemm_reshaped(m, n, 6, 4, 4, 2, 2, true, true); + } + } } std::pair CLGEMMReshapedConfigurationBifrost::configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b) @@ -96,43 +140,29 @@ std::pair CLGEMMReshapedConfigurationBifro ARM_COMPUTE_UNUSED(k); ARM_COMPUTE_UNUSED(b); - GEMMLHSMatrixInfo lhs_info; - GEMMRHSMatrixInfo rhs_info; - if(n <= 4) { - // Configure GEMMLHSMatrixInfo - lhs_info.m0 = 4; - lhs_info.k0 = 8; - lhs_info.v0 = lhs_info.m0 * 16 < m ? 2 : 16; - lhs_info.interleave = true; - lhs_info.transpose = false; - - // Configure GEMMRHSMatrixInfo - rhs_info.n0 = 2; - rhs_info.k0 = lhs_info.k0; - rhs_info.h0 = rhs_info.n0 * 16 < n ? 2 : 16; - rhs_info.interleave = false; - rhs_info.transpose = true; + return configure_gemm_reshaped(m, n, 4, 2, 8, 16, 16, true, false); } else { - // Configure GEMMLHSMatrixInfo - lhs_info.m0 = 4; - lhs_info.k0 = 2; - lhs_info.v0 = lhs_info.m0 * 8 < m ? 2 : 8; - lhs_info.interleave = false; - lhs_info.transpose = false; - - // Configure GEMMRHSMatrixInfo - rhs_info.n0 = 4; - rhs_info.k0 = lhs_info.k0; - rhs_info.h0 = rhs_info.n0 * 16 < n ? 2 : 16; - rhs_info.interleave = false; - rhs_info.transpose = true; + return configure_gemm_reshaped(m, n, 4, 4, 2, 8, 16, false, false); } +} - return std::make_pair(lhs_info, rhs_info); +std::pair CLGEMMReshapedConfigurationBifrost::configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b) +{ + ARM_COMPUTE_UNUSED(k); + ARM_COMPUTE_UNUSED(b); + + if(n <= 4) + { + return configure_gemm_reshaped(m, n, 4, 2, 16, 4, 1, false, false); + } + else + { + return configure_gemm_reshaped(m, n, 4, 4, 16, 2, 2, false, true); + } } } // namespace cl_gemm } // namespace arm_compute \ No newline at end of file -- cgit v1.2.1