COMPMID-3939: Update GEMM heuristic Mali-G77
- Update heuristic for GEMM reshaped RHS only
- Fix left-over block size in CLGEMMMatrixMultiplyReshapedOlyRHSKernel
Change-Id: I34c738821ed2e4a537da4a15058eec164cb6b61f
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4305
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp
index 519e903..3f82dca 100644
--- a/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp
+++ b/src/core/CL/gemm/reshaped/CLGEMMReshapedKernelConfigurationValhall.cpp
@@ -90,13 +90,88 @@
ARM_COMPUTE_UNUSED(k);
ARM_COMPUTE_UNUSED(b);
- if(n <= 4)
+ 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(r_mk <= 0.11824845522642136)
{
- return configure_lhs_rhs_info(m, n, 4, 2, 8, 8, 2, true, true, true, false);
+ if(workload <= 880.0)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 4, 4, 1, 4, false, false, true, false, false);
+ }
+ else
+ {
+ if(r_nk <= 0.42521367967128754)
+ {
+ if(workload <= 1726.4000244140625)
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 2, false, false, true, false, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true);
+ }
+ }
+ else
+ {
+ if(workload <= 1241.6000366210938)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 4, 4, 1, 4, false, false, true, false, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 4, false, false, true, false, false);
+ }
+ }
+ }
}
else
{
- return configure_lhs_rhs_info(m, n, 4, 8, 4, 4, 2, true, true, true, false);
+ if(workload <= 11404.7998046875)
+ {
+ if(r_mk <= 1.0126488208770752)
+ {
+ if(r_mn <= 2.545312523841858)
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 2, 4, 4, 1, 4, false, false, true, false, false);
+ }
+ }
+ else
+ {
+ if(workload <= 2881.199951171875)
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 4, 2, false, false, true, false, true);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true);
+ }
+ }
+ }
+ else
+ {
+ if(r_nk <= 0.5765306055545807)
+ {
+ if(r_mn <= 6.010416746139526)
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, false, true, true, false, true);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true);
+ }
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 4, 4, 4, 2, 1, true, false, true, false, true);
+ }
+ }
}
}
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp
index f7939d2..e099167 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMReshapedOnlyRHSKernelConfigurationValhall.cpp
@@ -78,67 +78,108 @@
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationValhall::configure_G77_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
{
- ARM_COMPUTE_UNUSED(k);
-
- GEMMLHSMatrixInfo lhs_info_buf;
- GEMMRHSMatrixInfo rhs_info_buf;
- GEMMLHSMatrixInfo lhs_info_img;
- GEMMRHSMatrixInfo rhs_info_img;
-
- // Get lhs_info/rhs_info in case of OpenCL buffer
if(m == 1)
{
- const unsigned int h0 = std::max(n / 4, 1U);
- std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true);
- }
- else
- {
- if(m > 256)
+ const float r_mn = static_cast<float>(m) / static_cast<float>(n);
+ const float r_mk = static_cast<float>(m) / static_cast<float>(k);
+
+ if(r_mk <= 0.0064484127797186375)
{
- const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1));
- std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, v0, false, true, false, true);
+ if(r_mn <= 0.0028273810748942196)
+ {
+ GEMMLHSMatrixInfo lhs_info_buf;
+ GEMMRHSMatrixInfo rhs_info_buf;
+ GEMMLHSMatrixInfo lhs_info_img;
+ GEMMRHSMatrixInfo rhs_info_img;
+
+ const unsigned int h0 = std::max(n / 4, 1U);
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 8, 1, 16, false, true, false, false, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true, false);
+
+ return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
+ std::make_pair(lhs_info_buf, rhs_info_buf),
+ n, k, b, DataType::F32);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 8, false, true, false, false, false);
+ }
}
else
{
- const int v0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1));
- std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 4, 4, 1, v0, false, true, false, true);
+ if(r_mk <= 0.020312500186264515)
+ {
+ return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 4, false, true, false, false, false);
+ }
+ else
+ {
+ return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, true, false);
+ }
}
}
-
- // Get lhs_info/rhs_info in case of OpenCL image
- if(m == 1)
- {
- std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, 8, true, true, false, false, true);
- }
else
{
- if((m / 4) * (n / 4) > 4096)
+ 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);
+
+ if(workload <= 1999.2000122070312)
{
- const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1));
- std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, h0, false, true, false, false, true);
+ if(workload <= 747.1999816894531)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+ }
+ else
+ {
+ GEMMLHSMatrixInfo lhs_info_buf;
+ GEMMRHSMatrixInfo rhs_info_buf;
+ GEMMLHSMatrixInfo lhs_info_img;
+ GEMMRHSMatrixInfo rhs_info_img;
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, false, false, false, true, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+
+ return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
+ std::make_pair(lhs_info_buf, rhs_info_buf),
+ n, k, b, DataType::F32);
+ }
}
else
{
- const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(8)), static_cast<int>(1));
- std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 4, 1, h0, false, true, false, false, true);
+ if(r_mn <= 0.03348214365541935)
+ {
+ if(r_mk <= 0.028125000186264515)
+ {
+ return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+ }
+ else
+ {
+ GEMMLHSMatrixInfo lhs_info_buf;
+ GEMMRHSMatrixInfo rhs_info_buf;
+ GEMMLHSMatrixInfo lhs_info_img;
+ GEMMRHSMatrixInfo rhs_info_img;
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, false, false, false, true, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+
+ return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
+ std::make_pair(lhs_info_buf, rhs_info_buf),
+ n, k, b, DataType::F32);
+ }
+ }
+ else
+ {
+ GEMMLHSMatrixInfo lhs_info_buf;
+ GEMMRHSMatrixInfo rhs_info_buf;
+ GEMMLHSMatrixInfo lhs_info_img;
+ GEMMRHSMatrixInfo rhs_info_img;
+ std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, false, true);
+ std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 16, false, true, false, true, false);
+
+ return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
+ std::make_pair(lhs_info_buf, rhs_info_buf),
+ n, k, b, DataType::F32);
+ }
}
}
-
- const TensorInfo tensor_rhs_info(TensorShape(n, k, b), 1, DataType::F32);
- const TensorShape shape = compute_rhs_reshaped_shape(tensor_rhs_info, rhs_info_img);
- const TensorInfo tensor_reshaped_info(shape, 1, DataType::F32);
-
- // In case of small workloads, we use the OpenCL buffer rather than the OpenCL image2d
- const bool use_cl_image2d = ((m / lhs_info_img.m0) * (n / rhs_info_img.n0)) * b < 1024 ? false : true;
-
- if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, rhs_info_img)) && use_cl_image2d)
- {
- return std::make_pair(lhs_info_img, rhs_info_img);
- }
- else
- {
- return std::make_pair(lhs_info_buf, rhs_info_buf);
- }
}
std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMReshapedOnlyRHSKernelConfigurationValhall::configure_G77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
index 68f761b..d53aede 100644
--- a/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.cpp
@@ -247,14 +247,14 @@
const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1);
const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2);
- // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
- const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
- const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
-
// Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads.
// NOTE: This might have implications on heuristics and performance
const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
+ // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
+ const unsigned int partial_store_m0 = internal_m % internal_m0;
+ const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
+
// Create build options
CLBuildOptions build_opts;
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
diff --git a/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.cpp b/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.cpp
index acae0e7..da41859 100644
--- a/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.cpp
+++ b/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.cpp
@@ -46,8 +46,8 @@
using FunctionExecutorPtr = CLGEMMKernelType (CLGEMMKernelSelectionValhall::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
- // Configurations for Valhall architectures
- static std::map<DataType, FunctionExecutorPtr> gemm_configs =
+ // Default configurations for Valhall architectures
+ static std::map<DataType, FunctionExecutorPtr> gemm_default_configs =
{
{ DataType::F32, &CLGEMMKernelSelectionValhall::default_f32 },
{ DataType::F16, &CLGEMMKernelSelectionValhall::default_f16 },
@@ -57,14 +57,34 @@
{ DataType::QSYMM8_PER_CHANNEL, &CLGEMMKernelSelectionValhall::default_q8 }
};
+ // Mali-G77 configurations
+ static std::map<DataType, FunctionExecutorPtr> gemm_g77_configs =
+ {
+ { DataType::F32, &CLGEMMKernelSelectionValhall::default_f32 },
+ { DataType::F16, &CLGEMMKernelSelectionValhall::g77_f16 },
+ { DataType::QASYMM8, &CLGEMMKernelSelectionValhall::default_q8 },
+ { DataType::QASYMM8_SIGNED, &CLGEMMKernelSelectionValhall::default_q8 },
+ { DataType::QSYMM8, &CLGEMMKernelSelectionValhall::default_q8 },
+ { DataType::QSYMM8_PER_CHANNEL, &CLGEMMKernelSelectionValhall::default_q8 }
+ };
+
const DataType data_type = params.data_type;
- if(gemm_configs.find(data_type) != gemm_configs.end())
+ switch(_target)
{
- return (this->*gemm_configs[data_type])(params.m, params.n, params.k, params.b, params.is_rhs_constant);
+ case GPUTarget::G77:
+ if(gemm_g77_configs.find(data_type) != gemm_g77_configs.end())
+ {
+ return (this->*gemm_g77_configs[data_type])(params.m, params.n, params.k, params.b, params.is_rhs_constant);
+ }
+ ARM_COMPUTE_ERROR("Not supported data type");
+ default:
+ if(gemm_default_configs.find(data_type) != gemm_default_configs.end())
+ {
+ return (this->*gemm_default_configs[data_type])(params.m, params.n, params.k, params.b, params.is_rhs_constant);
+ }
+ ARM_COMPUTE_ERROR("Not supported data type");
}
-
- ARM_COMPUTE_ERROR("Not supported data type");
}
CLGEMMKernelType CLGEMMKernelSelectionValhall::default_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
@@ -81,6 +101,110 @@
return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE_V1;
}
+CLGEMMKernelType CLGEMMKernelSelectionValhall::g77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+{
+ 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 workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f;
+
+ if(r_mk <= 0.6817956566810608)
+ {
+ if(workload <= 801.6000061035156)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ if(r_mn <= 0.0839829258620739)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ if(r_mk <= 0.24917218834161758)
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ else
+ {
+ if(workload <= 2551.75)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ if(workload <= 5061.574951171875)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ }
+ }
+ }
+ }
+ else
+ {
+ if(r_mk <= 4.849947690963745)
+ {
+ if(workload <= 17618.4501953125)
+ {
+ if(workload <= 5224.699951171875)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ if(r_nk <= 0.7933054566383362)
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+ }
+ else
+ {
+ if(workload <= 20275.2001953125)
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ else
+ {
+ if(r_mk <= 3.07421875)
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED;
+ }
+ }
+ }
+ }
+ else
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS;
+ }
+ }
+}
+
CLGEMMKernelType CLGEMMKernelSelectionValhall::default_q8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(m, n, k, b);
diff --git a/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.h b/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.h
index cbea9ea..82e46f6 100644
--- a/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.h
+++ b/src/runtime/CL/gemm/CLGEMMKernelSelectionValhall.h
@@ -47,6 +47,7 @@
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);
CLGEMMKernelType default_q8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
+ CLGEMMKernelType g77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
};
} // namespace cl_gemm
} // namespace arm_compute