|
|
|
|
|
|
|
|
|
|
|
#include <algorithm> |
|
#include <cmath> |
|
#include <cstring> |
|
#include <functional> |
|
#include <memory> |
|
#include <random> |
|
#include <vector> |
|
|
|
#include "bench/end2end.h" |
|
#include "bench/utils.h" |
|
#include <benchmark/benchmark.h> |
|
|
|
#include <xnnpack.h> |
|
#include <xnnpack/config.h> |
|
#include <xnnpack/dwconv.h> |
|
#include <xnnpack/microfnptr.h> |
|
#include <xnnpack/microparams-init.h> |
|
#include <xnnpack/models.h> |
|
|
|
|
|
static void DWConvEnd2EndBenchmark( |
|
benchmark::State& state, |
|
models::ExecutionPlanFactory model_factory, |
|
xnn_f16_dwconv_minmax_unipass_ukernel_fn dwconv_minmax, |
|
xnn_init_f16_minmax_params_fn init_params, |
|
uint8_t channel_tile, uint8_t primary_tile, |
|
benchmark::utils::IsaCheckFunction isa_check = nullptr) |
|
{ |
|
if (isa_check != nullptr && !isa_check(state)) { |
|
return; |
|
} |
|
if (xnn_initialize(nullptr ) != xnn_status_success) { |
|
state.SkipWithError("failed to initialize XNNPACK"); |
|
return; |
|
} |
|
|
|
struct xnn_dwconv_config* dwconv_config = xnn_init_f16_dwconv_config(); |
|
if (dwconv_config == nullptr) { |
|
state.SkipWithError("hardware does not support F16 DWCONV"); |
|
return; |
|
} |
|
|
|
|
|
struct xnn_dwconv_config saved_dwconv_params[XNN_MAX_F16_DWCONV_UKERNELS]; |
|
memcpy(saved_dwconv_params, dwconv_config, sizeof(saved_dwconv_params)); |
|
|
|
|
|
for (size_t i = 0; i < XNN_MAX_F16_DWCONV_UKERNELS; i++) { |
|
|
|
if (dwconv_config[i].primary_tile == primary_tile) { |
|
std::memset(&dwconv_config[i], 0, sizeof(dwconv_config[i])); |
|
|
|
|
|
dwconv_config[i].minmax.unipass = xnn_dwconv_unipass_ukernel_fn(dwconv_minmax); |
|
dwconv_config[i].channel_tile = channel_tile; |
|
dwconv_config[i].channel_subtile = channel_tile; |
|
dwconv_config[i].channel_round = 1; |
|
dwconv_config[i].primary_tile = primary_tile; |
|
dwconv_config[i].init.f16 = init_params; |
|
break; |
|
} |
|
} |
|
|
|
auto execution_plan = model_factory(nullptr); |
|
if (execution_plan.empty()) { |
|
state.SkipWithError("failed to create a model"); |
|
return; |
|
} |
|
|
|
for (auto _ : state) { |
|
for (const std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>& op : execution_plan) { |
|
xnn_status status = xnn_run_operator(op.get(), nullptr); |
|
if (status != xnn_status_success) { |
|
state.SkipWithError("failed to run a model"); |
|
return; |
|
} |
|
} |
|
} |
|
|
|
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
|
if (cpu_frequency != 0) { |
|
state.counters["cpufreq"] = cpu_frequency; |
|
} |
|
|
|
|
|
memcpy(dwconv_config, saved_dwconv_params, sizeof(saved_dwconv_params)); |
|
} |
|
|
|
static void DWConvEnd2EndBenchmark( |
|
benchmark::State& state, |
|
models::ExecutionPlanFactory model_factory, |
|
xnn_f16_dwconv_minmax_multipass_ukernel_fn dwconv_minmax, |
|
xnn_init_f16_minmax_params_fn init_params, |
|
uint8_t channel_tile, uint8_t channel_subtile, uint8_t channel_round, |
|
uint8_t primary_tile, uint8_t middle_tile, uint8_t last_tile, |
|
uint8_t primary_tile_to_replace, |
|
benchmark::utils::IsaCheckFunction isa_check = nullptr) |
|
{ |
|
if (isa_check != nullptr && !isa_check(state)) { |
|
return; |
|
} |
|
|
|
struct xnn_dwconv_config* dwconv_config = xnn_init_f16_dwconv_config(); |
|
if (dwconv_config == nullptr) { |
|
state.SkipWithError("failed to initialize f16 DWCONV config"); |
|
return; |
|
} |
|
|
|
|
|
struct xnn_dwconv_config saved_dwconv_params[XNN_MAX_F16_DWCONV_UKERNELS]; |
|
memcpy(saved_dwconv_params, dwconv_config, sizeof(saved_dwconv_params)); |
|
|
|
bool found = false; |
|
for (size_t i = 0; i < XNN_MAX_F16_DWCONV_UKERNELS; i++) { |
|
if (dwconv_config[i].primary_tile == primary_tile_to_replace) { |
|
found = true; |
|
} else if (dwconv_config[i].last_tile != 0) { |
|
|
|
found = true; |
|
} |
|
} |
|
|
|
if (!found) { |
|
state.SkipWithError("can't replace with multipass"); |
|
return; |
|
} |
|
|
|
|
|
for (size_t i = 0; i < XNN_MAX_F16_DWCONV_UKERNELS; i++) { |
|
|
|
if (dwconv_config[i].primary_tile == primary_tile_to_replace || |
|
dwconv_config[i].last_tile != 0) { |
|
|
|
|
|
|
|
std::memset(&dwconv_config[i], 0, sizeof(dwconv_config[i])); |
|
|
|
|
|
dwconv_config[i].minmax.multipass = xnn_dwconv_multipass_ukernel_fn(dwconv_minmax); |
|
dwconv_config[i].channel_tile = channel_tile; |
|
dwconv_config[i].channel_subtile = channel_subtile; |
|
dwconv_config[i].channel_round = channel_round; |
|
dwconv_config[i].primary_tile = primary_tile; |
|
dwconv_config[i].middle_tile = middle_tile; |
|
dwconv_config[i].last_tile = last_tile; |
|
dwconv_config[i].init.f16 = init_params; |
|
break; |
|
} |
|
} |
|
|
|
auto execution_plan = model_factory(nullptr); |
|
if (execution_plan.empty()) { |
|
state.SkipWithError("failed to create a model"); |
|
return; |
|
} |
|
|
|
for (auto _ : state) { |
|
for (const std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)>& op : execution_plan) { |
|
xnn_status status = xnn_run_operator(op.get(), nullptr); |
|
if (status != xnn_status_success) { |
|
state.SkipWithError("failed to run a model"); |
|
return; |
|
} |
|
} |
|
} |
|
|
|
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency(); |
|
if (cpu_frequency != 0) { |
|
state.counters["cpufreq"] = cpu_frequency; |
|
} |
|
|
|
memcpy(dwconv_config, saved_dwconv_params, sizeof(saved_dwconv_params)); |
|
} |
|
|
|
#if XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64) |
|
static void f16_dwconv_4p8c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p8c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_4p8c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p8c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_4p16c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p16c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_4p16c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p16c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_4p32c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p32c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_4p32c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_4p32c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 4, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
static void f16_dwconv_9p8c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p8c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_9p8c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p8c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_9p16c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p16c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_9p16c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p16c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_9p32c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p32c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_9p32c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_9p32c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 9, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
static void f16_dwconv_25p8c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p8c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_25p8c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p8c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_25p16c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p16c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_25p16c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p16c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_25p32c__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p32c__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_25p32c__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p32c__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 25, benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
static void f16_dwconv_5f5m5l8c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l8c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_5f5m5l8c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l8c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_5f5m5l16c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l16c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_5f5m5l16c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l16c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_5f5m5l32c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l32c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_5f5m5l32c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l32c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
static void f16_dwconv_6f6m7l8c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l8c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_6f6m7l8c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l8c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_6f6m7l16c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l16c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_6f6m7l16c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l16c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_6f6m7l32c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l32c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_6f6m7l32c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l32c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
static void f16_dwconv_8f8m9l8c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l8c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_8f8m9l8c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l8c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
8, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_8f8m9l16c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l16c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_8f8m9l16c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l16c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
16, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_8f8m9l32c8s4r__neonfp16arith(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l32c8s4r__neonfp16arith, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
static void f16_dwconv_8f8m9l32c8s4r__neonfp16arith_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l32c8s4r__neonfp16arith_acc2, xnn_init_f16_minmax_fp16arith_params, |
|
32, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckNEONFP16ARITH); |
|
} |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_4p8c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_4p8c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_4p16c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_4p16c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_4p32c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_4p32c__neonfp16arith_acc2); |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_9p8c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_9p8c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_9p16c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_9p16c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_9p32c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_9p32c__neonfp16arith_acc2); |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_25p8c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p8c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p16c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p16c__neonfp16arith_acc2); |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p32c__neonfp16arith); |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p32c__neonfp16arith_acc2); |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l8c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l8c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l16c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l16c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l32c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l32c8s4r__neonfp16arith_acc2) |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l8c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l8c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l16c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l16c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l32c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l32c8s4r__neonfp16arith_acc2) |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l8c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l8c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l16c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l16c8s4r__neonfp16arith_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l32c8s4r__neonfp16arith) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l32c8s4r__neonfp16arith_acc2) |
|
|
|
#endif |
|
|
|
#if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
|
static void f16_dwconv_25p8c__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p8c__fma3, xnn_init_f16_minmax_avx_params, |
|
8, 25, benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_25p8c__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p8c__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
8, 25, benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_25p16c__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p16c__fma3, xnn_init_f16_minmax_avx_params, |
|
16, 25, benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_25p16c__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p16c__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
16, 25, benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_25p32c__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p32c__fma3, xnn_init_f16_minmax_avx_params, |
|
32, 25, benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_25p32c__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_25p32c__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
32, 25, benchmark::utils::CheckFMA3); |
|
} |
|
|
|
static void f16_dwconv_5f5m5l8c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l8c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_5f5m5l8c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l8c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_5f5m5l16c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l16c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_5f5m5l16c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l16c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_5f5m5l32c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l32c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_5f5m5l32c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_5f5m5l32c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
5, 5, 5, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
|
|
static void f16_dwconv_6f6m7l8c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l8c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_6f6m7l8c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l8c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_6f6m7l16c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l16c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_6f6m7l16c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l16c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_6f6m7l32c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l32c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_6f6m7l32c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_6f6m7l32c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
6, 6, 7, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
|
|
static void f16_dwconv_8f8m9l8c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l8c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_8f8m9l8c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l8c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
8, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_8f8m9l16c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l16c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_8f8m9l16c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l16c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
16, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_8f8m9l32c8s4r__fma3(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l32c8s4r__fma3, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
static void f16_dwconv_8f8m9l32c8s4r__fma3_acc2(benchmark::State& state, models::ExecutionPlanFactory model) { |
|
DWConvEnd2EndBenchmark( |
|
state, model, |
|
xnn_f16_dwconv_minmax_ukernel_8f8m9l32c8s4r__fma3_acc2, xnn_init_f16_minmax_avx_params, |
|
32, 8, 4, |
|
8, 8, 9, |
|
25, |
|
benchmark::utils::CheckFMA3); |
|
} |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_25p8c__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p8c__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p16c__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p16c__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p32c__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_25p32c__fma3_acc2) |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l8c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l8c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l16c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l16c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l32c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_5f5m5l32c8s4r__fma3_acc2) |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l8c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l8c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l16c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l16c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l32c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_6f6m7l32c8s4r__fma3_acc2) |
|
|
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l8c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l8c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l16c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l16c8s4r__fma3_acc2) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l32c8s4r__fma3) |
|
BENCHMARK_FP16_END2END(f16_dwconv_8f8m9l32c8s4r__fma3_acc2) |
|
|
|
#endif |
|
|
|
#ifndef XNNPACK_BENCHMARK_NO_MAIN |
|
BENCHMARK_MAIN(); |
|
#endif |
|
|