aboutsummaryrefslogtreecommitdiff
path: root/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp')
-rw-r--r--src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp304
1 files changed, 148 insertions, 156 deletions
diff --git a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
index 982748810d..851e23bc84 100644
--- a/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
+++ b/src/runtime/CL/gemm/CLGEMMDefaultTypeValhall.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2020-2021 Arm Limited.
+ * Copyright (c) 2020-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -25,7 +25,8 @@
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "src/core/gpu/cl/kernels/gemm/ClGemmHelpers.h"
+
+#include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
#include <map>
#include <utility>
@@ -34,8 +35,7 @@ namespace arm_compute
{
namespace cl_gemm
{
-CLGEMMDefaultTypeValhall::CLGEMMDefaultTypeValhall(GPUTarget gpu)
- : ICLGEMMKernelSelection(gpu)
+CLGEMMDefaultTypeValhall::CLGEMMDefaultTypeValhall(GPUTarget gpu) : ICLGEMMKernelSelection(gpu)
{
}
@@ -44,189 +44,136 @@ CLGEMMKernelType CLGEMMDefaultTypeValhall::select_kernel(const CLGEMMKernelSelec
// _target could be used in the future to have a dedicated heuristic for each GPU IP
ARM_COMPUTE_UNUSED(_target);
- using FunctionExecutorPtr = CLGEMMKernelType (CLGEMMDefaultTypeValhall::*)(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
+ using FunctionExecutorPtr = CLGEMMKernelType (CLGEMMDefaultTypeValhall::*)(
+ unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant);
// Default configurations for Valhall architectures
- static std::map<DataType, FunctionExecutorPtr> gemm_default_configs =
- {
- { DataType::F32, &CLGEMMDefaultTypeValhall::default_f32 },
- { DataType::F16, &CLGEMMDefaultTypeValhall::default_f16 },
- { DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8 }
- };
+ static std::map<DataType, FunctionExecutorPtr> gemm_default_configs = {
+ {DataType::F32, &CLGEMMDefaultTypeValhall::default_f32},
+ {DataType::F16, &CLGEMMDefaultTypeValhall::default_f16},
+ {DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8}};
// Mali-G77 configurations
- static std::map<DataType, FunctionExecutorPtr> gemm_g77_configs =
- {
- { DataType::F32, &CLGEMMDefaultTypeValhall::default_f32 },
- { DataType::F16, &CLGEMMDefaultTypeValhall::g77_f16 },
- { DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8 }
- };
+ static std::map<DataType, FunctionExecutorPtr> gemm_g77_configs = {
+ {DataType::F32, &CLGEMMDefaultTypeValhall::default_f32},
+ {DataType::F16, &CLGEMMDefaultTypeValhall::g77_f16},
+ {DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8}};
// Mali-G78 configurations
- static std::map<DataType, FunctionExecutorPtr> gemm_g78_configs =
- {
- { DataType::F32, &CLGEMMDefaultTypeValhall::g78_f32 },
- { DataType::F16, &CLGEMMDefaultTypeValhall::g78_f16 },
- { DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8 },
- { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8 }
- };
+ static std::map<DataType, FunctionExecutorPtr> gemm_g78_configs = {
+ {DataType::F32, &CLGEMMDefaultTypeValhall::g78_f32},
+ {DataType::F16, &CLGEMMDefaultTypeValhall::g78_f16},
+ {DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8}};
+
+ // Mali-G710 and Mali-G610 configurations
+ static std::map<DataType, FunctionExecutorPtr> gemm_g710_configs = {
+ {DataType::F32, &CLGEMMDefaultTypeValhall::default_f32},
+ {DataType::F16, &CLGEMMDefaultTypeValhall::g710_f16},
+ {DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8}};
+
+ // Mali-G715 and Mali-G615 configurations
+ static std::map<DataType, FunctionExecutorPtr> gemm_g715_configs = {
+ {DataType::F32, &CLGEMMDefaultTypeValhall::g715_f32},
+ {DataType::F16, &CLGEMMDefaultTypeValhall::g715_f16},
+ {DataType::QASYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QASYMM8_SIGNED, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8, &CLGEMMDefaultTypeValhall::default_q8},
+ {DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultTypeValhall::default_q8}};
const DataType data_type = params.data_type;
- switch(_target)
+ switch (_target)
{
+ case GPUTarget::G710:
+ case GPUTarget::G610:
+ if (gemm_g710_configs.find(data_type) != gemm_g710_configs.end())
+ {
+ return (this->*gemm_g710_configs[data_type])(params.m, params.n, params.k, params.b,
+ params.is_rhs_constant);
+ }
+ ARM_COMPUTE_ERROR("Not supported data type");
+ case GPUTarget::G715:
+ case GPUTarget::G615:
+ if (gemm_g715_configs.find(data_type) != gemm_g715_configs.end())
+ {
+ return (this->*gemm_g715_configs[data_type])(params.m, params.n, params.k, params.b,
+ params.is_rhs_constant);
+ }
+ ARM_COMPUTE_ERROR("Not supported data type");
case GPUTarget::G78:
- if(gemm_g78_configs.find(data_type) != gemm_g78_configs.end())
+ if (gemm_g78_configs.find(data_type) != gemm_g78_configs.end())
{
- return (this->*gemm_g78_configs[data_type])(params.m, params.n, params.k, params.b, params.is_rhs_constant);
+ return (this->*gemm_g78_configs[data_type])(params.m, params.n, params.k, params.b,
+ params.is_rhs_constant);
}
ARM_COMPUTE_ERROR("Not supported data type");
case GPUTarget::G77:
- if(gemm_g77_configs.find(data_type) != gemm_g77_configs.end())
+ 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);
+ 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())
+ 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);
+ 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");
}
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::default_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType CLGEMMDefaultTypeValhall::default_f32(
+ unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(m, n, k, b);
- return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE_V1;
+ return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::default_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType CLGEMMDefaultTypeValhall::default_f16(
+ unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(m, n, k, b);
- return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE_V1;
+ return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::g77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::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;
- }
+ ARM_COMPUTE_UNUSED(m, n, k, b);
- if(m == 1)
- {
- return CLGEMMKernelType::RESHAPED_ONLY_RHS;
- }
+ return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
+}
- 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;
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::g710_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+{
+ ARM_COMPUTE_UNUSED(m, n, k, b);
- 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;
- }
- }
+ return is_rhs_constant ? CLGEMMKernelType::RESHAPED_ONLY_RHS : CLGEMMKernelType::NATIVE;
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::default_q8(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType CLGEMMDefaultTypeValhall::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);
- if(is_rhs_constant)
+ if (is_rhs_constant)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
@@ -236,47 +183,48 @@ CLGEMMKernelType CLGEMMDefaultTypeValhall::default_q8(unsigned int m, unsigned i
}
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::g78_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::g78_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(b);
- if(!is_rhs_constant)
+ if (!is_rhs_constant)
{
- return CLGEMMKernelType::NATIVE_V1;
+ return CLGEMMKernelType::NATIVE;
}
- if(m == 1)
+ if (m == 1)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
- if(n <= 272.0000f)
+ if (n <= 272.0000f)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
else
{
- if(k <= 471.0000f)
+ if (k <= 471.0000f)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
else
{
- if(m <= 72.5000f)
+ if (m <= 72.5000f)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
else
{
- if(m <= 90.5000f)
+ if (m <= 90.5000f)
{
return CLGEMMKernelType::RESHAPED;
}
else
{
- if(k <= 2448.0000f)
+ if (k <= 2448.0000f)
{
- if(n <= 756.0000f)
+ if (n <= 756.0000f)
{
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
@@ -295,16 +243,60 @@ CLGEMMKernelType CLGEMMDefaultTypeValhall::g78_f32(unsigned int m, unsigned int
}
}
-CLGEMMKernelType CLGEMMDefaultTypeValhall::g78_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::g78_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
{
ARM_COMPUTE_UNUSED(m, n, k, b);
- if(!is_rhs_constant)
+ if (!is_rhs_constant)
{
- return CLGEMMKernelType::NATIVE_V1;
+ return CLGEMMKernelType::NATIVE;
}
return CLGEMMKernelType::RESHAPED_ONLY_RHS;
}
+
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::g715_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+{
+ if (!is_rhs_constant)
+ {
+ return default_f32(m, n, k, b, is_rhs_constant);
+ }
+
+ unsigned int best_m0;
+ unsigned int best_n0;
+
+ if (opencl::kernels::gemm::is_mmul_kernel_preferred(m, n, k, b, DataType::F32, best_m0, best_n0))
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL;
+ }
+ else
+ {
+ return default_f32(m, n, k, b, is_rhs_constant);
+ }
+}
+
+CLGEMMKernelType
+CLGEMMDefaultTypeValhall::g715_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b, bool is_rhs_constant)
+{
+ if (!is_rhs_constant)
+ {
+ return g78_f16(m, n, k, b, is_rhs_constant);
+ }
+
+ unsigned int best_m0;
+ unsigned int best_n0;
+
+ if (opencl::kernels::gemm::is_mmul_kernel_preferred(m, n, k, b, DataType::F16, best_m0, best_n0))
+ {
+ return CLGEMMKernelType::RESHAPED_ONLY_RHS_MMUL;
+ }
+ else
+ {
+ return g78_f16(m, n, k, b, is_rhs_constant);
+ }
+}
+
} // namespace cl_gemm
} // namespace arm_compute