// Copyright (c) Facebook, Inc. and its affiliates. // All rights reserved. // // Copyright 2019 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. // // Auto-generated file. Do not edit! // Specification: test/f16-gemm-minmax.yaml // Generator: tools/generate-gemm-test.py #include #include #include #include #include #include "xnnpack/allocator.h" #include "xnnpack/common.h" #include "xnnpack/gemm.h" #include "xnnpack/igemm.h" #include "xnnpack/isa-checks.h" #include "xnnpack/microparams-init.h" #include "xnnpack/pack.h" #include "xnnpack/packw.h" #include "xnnpack/ppmm.h" #include "xnnpack/requantization.h" #include "gemm-microkernel-tester.h" #include "next_prime.h" namespace { std::vector CreateTests1( size_t k_block, size_t adj_k_block, size_t mr, size_t nr, size_t kr, size_t sr, bool is_igemm, bool unsigned_inputs, std::function test_func, std::function isa_check = nullptr) { std::string kbs = std::to_string(k_block); std::string kb2s = std::to_string(k_block * 2); std::string akbs = std::to_string(adj_k_block); std::string nrs = std::to_string(nr); const GemmMicrokernelTester tester = GemmMicrokernelTester() .mr(mr).nr(nr).kr(kr).sr(sr).unsigned_inputs(unsigned_inputs); std::vector gemm_tests; gemm_tests.reserve(42); gemm_tests.push_back(GemmTestParams( "k_eq_" + kbs, tester.clone() .m(mr).n(nr).k(k_block) , test_func, isa_check)); gemm_tests.push_back(GemmTestParams( "strided_cn", tester.clone() .m(mr).n(nr).k(k_block) .cn_stride(xnnpack::NextPrime(nr + 1)) , test_func, isa_check)); if (!is_igemm) { gemm_tests.push_back(GemmTestParams( "k_eq_" + kbs + "_strided_a", tester.clone() .m(mr).n(nr).k(k_block) .a_stride(xnnpack::NextPrime(k_block + 1)) , test_func, isa_check)); } gemm_tests.push_back(GemmTestParams( "k_eq_" + kbs + "_subtile", tester.clone() .k(k_block).iterations(1) , test_func, isa_check) .loop_n(1, nr) .loop_m(1, mr)); gemm_tests.push_back(GemmTestParams( "k_eq_" + kbs + "_subtile_m", tester.clone() .n(nr).k(k_block).iterations(1) , test_func, isa_check) .loop_m(1, mr)); gemm_tests.push_back(GemmTestParams( "k_eq_" + kbs + "_subtile_n", tester.clone() .m(mr).k(k_block).iterations(1) , test_func, isa_check) .loop_n(1, nr)); if (k_block > 1) { gemm_tests.push_back(GemmTestParams( "k_lt_" + akbs, tester.clone() .m(mr).n(nr) , test_func, isa_check) .loop_k(1, adj_k_block - 1)); if (!is_igemm) { gemm_tests.push_back(GemmTestParams( "k_lt_" + akbs + "_strided_a", tester.clone() .m(mr).n(nr) .a_stride(xnnpack::NextPrime(adj_k_block + 1)) , test_func, isa_check) .loop_k(1, adj_k_block - 1)); } gemm_tests.push_back(GemmTestParams( "k_lt_" + akbs + "_subtile", tester.clone() .iterations(1) , test_func, isa_check) .loop_k(1, adj_k_block - 1) .loop_n(1, nr) .loop_m(1, mr)); } gemm_tests.push_back(GemmTestParams( "k_gt_" + akbs, tester.clone() .m(mr).n(nr) , test_func, isa_check) .loop_k(adj_k_block + 1, adj_k_block * 2 - 1, k_block)); if (is_igemm) { gemm_tests.push_back(GemmTestParams( "k_gt_" + akbs + "_strided_a", tester.clone() .m(mr).n(nr) .a_stride(xnnpack::NextPrime(adj_k_block * 2 + 1)) , test_func, isa_check) .loop_k(adj_k_block + 1, adj_k_block * 2 - 1, k_block)); } gemm_tests.push_back(GemmTestParams( "k_gt_" + akbs + "_subtile", tester.clone() .iterations(1) , test_func, isa_check) .loop_k(adj_k_block + 1, adj_k_block * 2 - 1, k_block) .loop_n(1, nr) .loop_m(1, mr)); if (k_block > 1) { gemm_tests.push_back(GemmTestParams( "k_div_" + kbs, tester.clone() .m(mr).n(nr) , test_func, isa_check) .loop_k(adj_k_block + k_block, k_block * 5, k_block)); if (is_igemm) { gemm_tests.push_back(GemmTestParams( "k_div_" + kbs + "_strided_a", tester.clone() .m(mr).n(nr) .a_stride(xnnpack::NextPrime(k_block * 3 + 1)) , test_func, isa_check) .loop_k(adj_k_block + k_block, k_block * 3, k_block)); } gemm_tests.push_back(GemmTestParams( "k_div_" + kbs + "_subtile", tester.clone() .iterations(1) , test_func, isa_check) .loop_k(adj_k_block + k_block, k_block * 5, k_block) .loop_n(1, nr) .loop_m(1, mr)); } gemm_tests.push_back(GemmTestParams( "n_gt_" + nrs, tester.clone() .m(mr) , test_func, isa_check) .loop_n(nr + 1, nr * 2 - 1) .loop_k(1, k_block * 3, k_block + 1)); gemm_tests.push_back(GemmTestParams( "n_gt_" + nrs + "_strided_cn", tester.clone() .m(mr) .cn_stride(xnnpack::NextPrime(nr + 1)) , test_func, isa_check) .loop_n(nr + 1, nr * 2 - 1) .loop_k(1, k_block * 3, k_block + 1)); if (!is_igemm) { gemm_tests.push_back(GemmTestParams( "n_gt_" + nrs + "_strided_a", tester.clone() .m(mr) .a_stride(xnnpack::NextPrime(k_block * 3 + 1)) , test_func, isa_check) .loop_n(nr + 1, nr * 2 - 1) .loop_k(1, k_block * 3, k_block)); } gemm_tests.push_back(GemmTestParams( "n_gt_" + nrs + "_subtile", tester.clone() .iterations(1) , test_func, isa_check) .loop_n(nr + 1, nr * 2 - 1) .loop_k(1, k_block * 3, k_block + 1) .loop_m(1, mr)); gemm_tests.push_back(GemmTestParams( "n_div_" + nrs, tester.clone() .m(mr) , test_func, isa_check) .loop_n(nr * 2, nr * 3, nr) .loop_k(1, k_block * 3, k_block + 1)); gemm_tests.push_back(GemmTestParams( "n_div_" + nrs + "_strided_cn", tester.clone() .m(mr) .cn_stride(xnnpack::NextPrime(nr + 1)) , test_func, isa_check) .loop_n(nr * 2, nr * 3, nr) .loop_k(1, k_block * 3, k_block + 1)); if (!is_igemm) { gemm_tests.push_back(GemmTestParams( "n_div_" + nrs + "_strided_a", tester.clone() .m(mr) .a_stride(xnnpack::NextPrime(k_block * 3 + 1)) , test_func, isa_check) .loop_n(nr * 2, nr * 3, nr) .loop_k(1, k_block * 3, k_block)); } gemm_tests.push_back(GemmTestParams( "n_div_" + nrs + "_subtile", tester.clone() .iterations(1) , test_func, isa_check) .loop_n(nr * 2, nr * 3, nr) .loop_k(1, k_block * 3, k_block + 1) .loop_m(1, mr)); if (is_igemm) { gemm_tests.push_back(GemmTestParams( "small_kernel", tester.clone() .m(mr).n(nr).ks(3) , test_func, isa_check) .loop_k(1, k_block * 3, k_block + 1)); gemm_tests.push_back(GemmTestParams( "small_kernel_subtile", tester.clone() .ks(3).iterations(1) , test_func, isa_check) .loop_k(1, k_block * 3, k_block + 1) .loop_n(1, nr) .loop_m(1, mr)); gemm_tests.push_back(GemmTestParams( "n_gt_" + nrs + "_small_kernel", tester.clone() .m(mr).ks(3) , test_func, isa_check) .loop_n(nr + 1, nr * 2 - 1) .loop_k(1, k_block * 3, k_block + 1)); gemm_tests.push_back(GemmTestParams( "n_div_" + nrs + "_small_kernel", tester.clone() .m(mr).ks(3) , test_func, isa_check) .loop_n(nr * 2, nr * 3, nr) .loop_k(1, k_block * 3, k_block + 1)); } gemm_tests.push_back(GemmTestParams( "strided_cm_subtile", tester.clone() .mr(mr).nr(nr).kr(kr).sr(sr) .cm_stride(xnnpack::NextPrime(nr + 1)) .iterations(1) , test_func, isa_check) .loop_k(1, k_block * 3, k_block + 1) .loop_n(1, nr) .loop_m(1, mr)); if (is_igemm) { gemm_tests.push_back(GemmTestParams( "a_offset", tester.clone() .m(mr).n(nr).ks(3) .a_offset(xnnpack::NextPrime(mr * k_block * 3 + 1)) , test_func, isa_check) .loop_k(1, k_block * 3, k_block + 1)); gemm_tests.push_back(GemmTestParams( "zero", tester.clone() .m(mr).n(nr).ks(3) .a_offset(xnnpack::NextPrime(mr * k_block * 3 + 1)) , test_func, isa_check) .loop_k(1, k_block * 3, k_block + 1) .loop_zi(0, mr - 1)); } gemm_tests.push_back(GemmTestParams( "qmin", tester.clone() .m(mr).n(nr).k(k_block).qmin(128) , test_func, isa_check)); gemm_tests.push_back(GemmTestParams( "qmax", tester.clone() .m(mr).n(nr).k(k_block).qmax(128) , test_func, isa_check)); gemm_tests.push_back(GemmTestParams( "strided_cm", tester.clone() .m(mr).n(nr).k(k_block) .cm_stride(xnnpack::NextPrime(nr + 1)) , test_func, isa_check)); return gemm_tests; } } // namespace #if XNN_ENABLE_ARM_FP16_VECTOR && XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X8__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x8__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X8__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x8__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X8__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x8__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_8X8__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/8, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_8x8__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X16__ASM_AARCH64_NEONFP16ARITH_LD32, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/2, /*adj_k_block=*/2, /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X16__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X16__ASM_AARCH64_NEONFP16ARITH_LD32, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/2, /*adj_k_block=*/2, /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld32, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X16__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__ASM_AARCH64_NEONFP16ARITH_CORTEX_A55, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__ASM_AARCH64_NEONFP16ARITH_CORTEX_A55R0, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__ASM_AARCH64_NEONFP16ARITH_CORTEX_A75, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a75, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__ASM_AARCH64_NEONFP16ARITH_LD32, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/2, /*adj_k_block=*/2, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_ld32, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__ASM_AARCH64_NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); #endif // XNN_ENABLE_ARM_FP16_VECTOR && XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY #if XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64) INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X8__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x8__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/8, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X16__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X16__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x16__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X16__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_8X16__NEONFP16ARITH_LD64, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/4, /*adj_k_block=*/4, /*mr=*/8, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_8x16__neonfp16arith_ld64, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__neon_ld4lane_u8); }, []() { TEST_REQUIRES_ARM_NEON_FP16_ARITH; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); #endif // XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64) #if XNN_ENABLE_AVX512FP16 && (XNN_ARCH_X86 || XNN_ARCH_X86_64) INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/1, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/4, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_5X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/5, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_5x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/6, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_7X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/7, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_7x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_8X32__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/8, /*nr=*/32, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_8x32__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x32__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/1, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/4, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_5X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/5, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_5x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/6, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_7X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/7, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_7x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_8X64__AVX512FP16_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/8, /*nr=*/64, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_8x64__avx512fp16_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x64__scalar_int_u4); }, []() { TEST_REQUIRES_X86_AVX512FP16; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); #endif // XNN_ENABLE_AVX512FP16 && (XNN_ARCH_X86 || XNN_ARCH_X86_64) #if XNN_ARCH_X86 || XNN_ARCH_X86_64 INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X8__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x8__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__avx2_u16); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X8__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x8__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__avx2_u16); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_5X8__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/5, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_5x8__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__avx2_u16); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_6X8__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_6x8__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__avx2_u16); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_7X8__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/7, /*nr=*/8, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_7x8__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x8__avx2_u16); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_1X16__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_1x16__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__avx2_u16_prfm); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_3X16__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/3, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_3x16__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__avx2_u16_prfm); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_4X16__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_4x16__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__avx2_u16_prfm); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); INSTANTIATE_TEST_SUITE_P( F16_GEMM_MINMAX_5X16__AVX2_BROADCAST, GemmTest, testing::ValuesIn(CreateTests1( /*k_block=*/1, /*adj_k_block=*/1, /*mr=*/5, /*nr=*/16, /*kr=*/1, /*sr=*/1, /*is_igemm=*/false, /*unsigned_inputs=*/false, [](GemmMicrokernelTester& tester) { tester.Test(xnn_f16_gemm_minmax_ukernel_5x16__avx2_broadcast, xnn_init_f16_minmax_scalar_params, xnn_x16_packw_gemm_goi_ukernel_x16__avx2_u16_prfm); }, []() { TEST_REQUIRES_X86_AVX2; })), [](const testing::TestParamInfo& info) { return info.param.test_name; }); #endif // XNN_ARCH_X86 || XNN_ARCH_X86_64