static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G52 =
{
{ DataType::F32, &CLGEMMReshapedKernelConfigurationBifrost::configure_G52_f32 },
- { DataType::F16, &CLGEMMReshapedKernelConfigurationBifrost::configure_G7x_f16 },
+ { DataType::F16, &CLGEMMReshapedKernelConfigurationBifrost::configure_G52_f16 },
{ DataType::QASYMM8, &CLGEMMReshapedKernelConfigurationBifrost::configure_G7x_u8 },
{ DataType::QSYMM8, &CLGEMMReshapedKernelConfigurationBifrost::configure_G7x_u8 },
{ DataType::QASYMM8_SIGNED, &CLGEMMReshapedKernelConfigurationBifrost::configure_G7x_u8 },
}
}
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedKernelConfigurationBifrost::configure_G52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+ ARM_COMPUTE_UNUSED(k);
+
+ const float workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f;
+
+ if(workload <= 232.8000f)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 4, 4, 4, 4, true, true, true, false, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 4, true, true, true, false, false);
+ }
+}
+
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedKernelConfigurationBifrost::configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
{
ARM_COMPUTE_UNUSED(k);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G52_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G52 =
{
{ DataType::F32, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G52_f32 },
- { DataType::F16, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_f16 },
+ { DataType::F16, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G52_f16 },
{ DataType::QASYMM8, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_u8 },
{ DataType::QSYMM8, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_u8 },
{ DataType::QASYMM8_SIGNED, &CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G7x_u8 },
GEMMRHSMatrixInfo rhs_info_img;
const bool is_workload_big = ((m * n * b) / 16) >= 2048;
- // Get lhs_info/rhs_info in case of OpenCL buffer
+
if(m == 1)
{
- if(n <= 204.0)
+ if(n >= 8192)
{
- return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 16, false, true, false, true, false);
+ const unsigned int h0 = std::max(n / 4, 1U);
+ return configure_lhs_rhs_info(m, n, 1, 4, 8, 1, h0, false, true, false, true, false);
}
else
{
- return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 32, false, true, false, true, false);
+ const unsigned int h0 = std::max(n / 2, 1U);
+ if(n <= 204)
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, false, true, false, true, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true, false);
+ }
}
}
else
}
}
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+ const float r_mn = static_cast<float>(m) / static_cast<float>(n);
+ const float workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f;
+ const float r_mk = static_cast<float>(m) / static_cast<float>(k);
+ const float r_nk = static_cast<float>(n) / static_cast<float>(k);
+
+ if(m == 1)
+ {
+ if(r_mk <= 0.0026f)
+ {
+ if(r_nk <= 0.4664f)
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 32, false, true, false, true, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, false, true);
+ }
+ }
+ else
+ {
+ if(r_mk <= 0.0148f)
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 32, false, true, false, true, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, false, true);
+ }
+ }
+ }
+ else
+ {
+ if(workload <= 362.6000f)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 2, 8, 1, 16, false, false, false, true, false);
+ }
+ else
+ {
+ if(r_mn <= 22.6067f)
+ {
+ if(workload <= 708.8000f)
+ {
+ return configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, false, true);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 5, 8, 2, 1, 16, false, false, false, false, false);
+ }
+ }
+ else
+ {
+ if(r_nk <= 0.0917f)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 2, 8, 1, 16, false, false, false, true, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 2, false, false, false, false, true);
+ }
+ }
+ }
+ }
+}
+
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationBifrost::configure_G76_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
{
ARM_COMPUTE_UNUSED(k);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G52_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G51_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G76_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G51_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G7x_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
static std::map<DataType, FunctionExecutorPtr> gemm_g52_configs =
{
{ DataType::F32, &CLGEMMKernelSelectionBifrost::g52_f32 },
- { DataType::F16, &CLGEMMKernelSelectionBifrost::default_f16 },
+ { DataType::F16, &CLGEMMKernelSelectionBifrost::g52_f16 },
{ DataType::QASYMM8, &CLGEMMKernelSelectionBifrost::default_q8 },
{ DataType::QASYMM8_SIGNED, &CLGEMMKernelSelectionBifrost::default_q8 },
{ DataType::QSYMM8, &CLGEMMKernelSelectionBifrost::default_q8 },
}
}
+CLGEMMKernelType CLGEMMKernelSelectionBifrost::g52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+{
+ ARM_COMPUTE_UNUSED(b);
+
+ if (!is_rhs_constant)
+ {
+ return CLGEMMKernelType::NATIVE_V1;
+ }
+
+ if (m == 1)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+
+ const float r_mn = static_cast<float>(m) / static_cast<float>(n);
+ const float r_mk = static_cast<float>(m) / static_cast<float>(k);
+ const float r_nk = static_cast<float>(n) / static_cast<float>(k);
+ const float r_mnk = static_cast<float>(m) / (static_cast<float>(n) * static_cast<float>(k));
+
+ if(r_mn <= 22.9200f)
+ {
+ if(r_mk <= 0.0157f)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ if(r_mnk <= 7809.3750f)
+ {
+ if(r_mnk <= 101.7937f)
+ {
+ if(r_mn <= 0.4594f)
+ {
+ if(r_mk <= 0.0557f)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+ else
+ {
+ if(r_nk <= 0.4396f)
+ {
+ if(r_mn <= 1.5182f)
+ {
+ if(r_mnk <= 1709.9167f)
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+ else
+ {
+ if(r_mnk <= 1330.6000f)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ }
+ else
+ {
+ if(r_mn <= 2.5896f)
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ else
+ {
+ if(r_mnk <= 326.6667f)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ }
+ }
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+ }
+ else
+ {
+ if(r_mn <= 86.7578f)
+ {
+ if(r_mnk <= 11231.6406f)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+}
+
CLGEMMKernelType CLGEMMKernelSelectionBifrost::g71_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(b);
CLGEMMKernelType g52_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
CLGEMMKernelType g76_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
CLGEMMKernelType g76_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
+ CLGEMMKernelType g52_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
CLGEMMKernelType g71_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
CLGEMMKernelType default_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
CLGEMMKernelType default_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);