| // Copyright 2021 Google LLC |
| // |
| // This source code is licensed under the BSD-style license found in the |
| // LICENSE file in the root directory of this source tree. |
| |
| #include <algorithm> |
| #include <cmath> |
| #include <functional> |
| #include <random> |
| #include <vector> |
| |
| #include <xnnpack.h> |
| |
| #include <benchmark/benchmark.h> |
| |
| #include "bench/end2end.h" |
| #include "bench/utils.h" |
| #include "models/models.h" |
| #include <xnnpack/dwconv.h> |
| #include <xnnpack/params.h> |
| #include <xnnpack/params-init.h> |
| |
| |
| static void DWConvEnd2EndBenchmark( |
| benchmark::State& state, |
| models::ExecutionPlanFactory model_factory, |
| xnn_qu8_dwconv_minmax_unipass_ukernel_function dwconv, |
| xnn_init_qu8_conv_minmax_params_fn init_params, |
| uint8_t channel_tile, uint8_t primary_tile, |
| benchmark::utils::IsaCheckFunction isa_check = nullptr) |
| { |
| if (isa_check && !isa_check(state)) { |
| return; |
| } |
| if (xnn_initialize(nullptr /* allocator */) != xnn_status_success) { |
| state.SkipWithError("failed to initialize XNNPACK"); |
| return; |
| } |
| |
| // Override microkernels chosen in xnn_initialize |
| for (size_t i = 0; i < XNN_MAX_QU8_DWCONV_UKERNELS; i++) { |
| // Replace only the microkernel the matching kernel size. |
| if (xnn_params.qu8.dwconv[i].primary_tile == primary_tile) { |
| // Note: do not directly assign to xnn_params.qu8.dwconv[i] because it breaks older gcc. |
| xnn_params.qu8.dwconv[i].minmax.unipass = xnn_dwconv_unipass_ukernel_function(dwconv); |
| xnn_params.qu8.dwconv[i].channel_tile = channel_tile; |
| xnn_params.qu8.dwconv[i].primary_tile = primary_tile; |
| xnn_params.qu8.dwconv[i].incremental_tile = 0; |
| xnn_params.qu8.dwconv[i].init.qu8 = 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; |
| } |
| } |
| |
| |
| #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| static void qu8_dwconv_up8x9__neon_mul8(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul8, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up16x9__neon_mul8(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up24x9__neon_mul8(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul8, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 24 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up32x9__neon_mul8(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 32 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up8x9__neon_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x9__neon_mul16, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up16x9__neon_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul16, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up24x9__neon_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 24 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| static void qu8_dwconv_up32x9__neon_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16, |
| xnn_init_qu8_conv_minmax_rndnu_neon_params, |
| 32 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckNEON); |
| } |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__neon_mul8); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__neon_mul8); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up24x9__neon_mul8); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up32x9__neon_mul8); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__neon_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__neon_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up24x9__neon_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up32x9__neon_mul16); |
| #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| |
| |
| #if XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| static void qu8_dwconv_up16x9__avx512skx_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx512skx_mul32, |
| xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX512SKX); |
| } |
| static void qu8_dwconv_up32x9__avx512skx_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32, |
| xnn_init_qu8_conv_minmax_fp32_avx512_params, |
| 32 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX512SKX); |
| } |
| static void qu8_dwconv_up8x9__avx2_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx2_mul32, |
| xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX2); |
| } |
| static void qu8_dwconv_up16x9__avx2_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32, |
| xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX2); |
| } |
| static void qu8_dwconv_up32x9__avx2_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32, |
| xnn_init_qu8_conv_minmax_fp32_avx2_params, |
| 32 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX2); |
| } |
| static void qu8_dwconv_up8x9__avx_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX); |
| } |
| static void qu8_dwconv_up16x9__avx_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX); |
| } |
| static void qu8_dwconv_up8x9__avx_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__avx_mul32, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX); |
| } |
| static void qu8_dwconv_up16x9__avx_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul32, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckAVX); |
| } |
| static void qu8_dwconv_up8x9__sse41_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckSSE41); |
| } |
| static void qu8_dwconv_up16x9__sse41_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckSSE41); |
| } |
| static void qu8_dwconv_up8x9__sse41_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul32, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 8 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckSSE41); |
| } |
| static void qu8_dwconv_up16x9__sse41_mul32(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse41_mul32, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 16 /* channel tile */, 9 /* primary tile */, benchmark::utils::CheckSSE41); |
| } |
| static void qu8_dwconv_up8x9__sse2_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 8 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up16x9__sse2_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__sse2_mul16, |
| xnn_init_qu8_conv_minmax_fp32_sse2_params, |
| 16 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__avx512skx_mul32); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up32x9__avx512skx_mul32); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__avx2_mul32); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__avx2_mul32); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up32x9__avx2_mul32); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__avx_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__avx_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__avx_mul32); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__avx_mul32); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__sse41_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__sse41_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__sse41_mul32); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__sse41_mul32); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__sse2_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__sse2_mul16); |
| #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64 |
| |
| |
| #if XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| static void qu8_dwconv_up8x9__wasmsimd_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16, |
| xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 8 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up16x9__wasmsimd_mul16(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__wasmsimd_mul16, |
| xnn_init_qu8_conv_minmax_fp32_wasmsimd_params, |
| 16 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up8x9__wasmsimd_mul16); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up16x9__wasmsimd_mul16); |
| #endif // XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| |
| |
| #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| static void qu8_dwconv_up1x9__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__wasm_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up2x9__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 2 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up4x9__wasm_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__wasm_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 4 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up1x9__wasm_fmagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up2x9__wasm_fmagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up4x9__wasm_fmagic); |
| #endif // XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD |
| |
| |
| static void qu8_dwconv_up1x9__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 1 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up2x9__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 2 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up4x9__scalar_fmagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_fmagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params, |
| 4 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| static void qu8_dwconv_up1x9__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_imagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 1 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up2x9__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 2 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up4x9__scalar_imagic(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_imagic, |
| xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params, |
| 4 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| static void qu8_dwconv_up1x9__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_lrintf, |
| xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 1 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up2x9__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf, |
| xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 2 /* channel tile */, 9 /* primary tile */); |
| } |
| static void qu8_dwconv_up4x9__scalar_lrintf(benchmark::State& state, models::ExecutionPlanFactory model) { |
| DWConvEnd2EndBenchmark(state, model, |
| xnn_qu8_dwconv_minmax_fp32_ukernel_up4x9__scalar_lrintf, |
| xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params, |
| 4 /* channel tile */, 9 /* primary tile */); |
| } |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up1x9__scalar_fmagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up2x9__scalar_fmagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up4x9__scalar_fmagic); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up1x9__scalar_imagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up2x9__scalar_imagic); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up4x9__scalar_imagic); |
| |
| BENCHMARK_QU8_END2END(qu8_dwconv_up1x9__scalar_lrintf); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up2x9__scalar_lrintf); |
| BENCHMARK_QU8_END2END(qu8_dwconv_up4x9__scalar_lrintf); |
| |
| |
| #ifndef XNNPACK_BENCHMARK_NO_MAIN |
| BENCHMARK_MAIN(); |
| #endif |