hailin
/
sglang_v0.5.2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
sglang_v0.5.2/pytorch_2.8.0/third_party/XNNPACK/test/qd8-f16-qc8w-gemm-minmax.cc

771 lines
26 KiB
C++
Raw Blame History

// 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/qd8-f16-qc8w-gemm-minmax.yaml
// Generator: tools/generate-gemm-test.py
#include <cstddef>
#include <functional>
#include <string>
#include <vector>
#include <gtest/gtest.h>
#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<GemmTestParams> 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<void(GemmMicrokernelTester& tester)> test_func,
std::function<void()> 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<GemmTestParams> 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_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_16X64C4__AVX512AMX, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/64,
/*adj_k_block=*/64,
/*mr=*/16, /*nr=*/64, /*kr=*/4, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_16x64c4__avx512amx,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX512AMX;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_AVX512AMX && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
#if XNN_ENABLE_ARM_I8MM && XNN_ARCH_ARM64
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_1X8C8__NEONI8MM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/1, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_1x8c8__neoni8mm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_I8MM;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_2X32C8__NEONI8MM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/2, /*nr=*/32, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_2x32c8__neoni8mm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_I8MM;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X32C8__NEONI8MM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/4, /*nr=*/32, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x32c8__neoni8mm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_I8MM;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_6X16C8__NEONI8MM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/6, /*nr=*/16, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_6x16c8__neoni8mm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_I8MM;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_8X32C8__NEONI8MM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/8, /*nr=*/32, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_8x32c8__neoni8mm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_I8MM;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_ARM_I8MM && XNN_ARCH_ARM64
#if XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X8C4__NEONDOTFP16ARITH, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/4,
/*adj_k_block=*/4,
/*mr=*/4, /*nr=*/8, /*kr=*/4, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x8c4__neondotfp16arith,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_DOT_FP16_ARITH;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
#if XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X16C4__ASM_AARCH64_NEONDOTFP16ARITH_CORTEX_A55, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/4, /*nr=*/16, /*kr=*/4, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondotfp16arith_cortex_a55,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_DOT_FP16_ARITH;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X16C4__ASM_AARCH64_NEONDOTFP16ARITH_LD128, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/4, /*nr=*/16, /*kr=*/4, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x16c4__asm_aarch64_neondotfp16arith_ld128,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_DOT_FP16_ARITH;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
#if XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_5X8C4__NEONDOTFP16ARITH, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/4,
/*adj_k_block=*/4,
/*mr=*/5, /*nr=*/8, /*kr=*/4, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_5x8c4__neondotfp16arith,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_ARM_NEON_DOT_FP16_ARITH;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_ARM_DOTPROD && XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
#if XNN_ENABLE_AVX256VNNI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_8X8C8__AVX256VNNI, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/8, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_8x8c8__avx256vnni,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_10X8C8__AVX256VNNI, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/10, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_10x8c8__avx256vnni,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_1X8C8__AVX256VNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/1, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_1x8c8__avx256vnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_5X8C8__AVX256VNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_5x8c8__avx256vnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_7X8C8__AVX256VNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/7, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_7x8c8__avx256vnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_8X8C8__AVX256VNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/8, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_8x8c8__avx256vnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_9X8C8__AVX256VNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/9, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_9x8c8__avx256vnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX256VNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_AVX256VNNI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
#if XNN_ENABLE_AVXVNNI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_2X8C8__AVXVNNI, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/2, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_2x8c8__avxvnni,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVXVNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_5X8C8__AVXVNNI, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_5x8c8__avxvnni,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVXVNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X8C8__AVXVNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/4, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x8c8__avxvnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVXVNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_5X8C8__AVXVNNI_PRFM, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/16,
/*adj_k_block=*/16,
/*mr=*/5, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/true,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_5x8c8__avxvnni_prfm,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVXVNNI;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ENABLE_AVXVNNI && (XNN_ARCH_X86 || XNN_ARCH_X86_64)
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
INSTANTIATE_TEST_SUITE_P(
QD8_F16_QC8W_GEMM_MINMAX_4X8C8__AVX2, GemmTest,
testing::ValuesIn(CreateTests1(
/*k_block=*/8,
/*adj_k_block=*/8,
/*mr=*/4, /*nr=*/8, /*kr=*/8, /*sr=*/1,
/*is_igemm=*/false,
/*unsigned_inputs=*/false,
[](GemmMicrokernelTester& tester) {
tester.Test(xnn_qd8_f16_qc8w_gemm_minmax_ukernel_4x8c8__avx2,
xnn_init_f16_minmax_scalar_params,
xnn_pack_qs8_gemm_goi_w);
},
[]() {
TEST_REQUIRES_X86_AVX2;
})),
[](const testing::TestParamInfo<GemmTest::ParamType>& info) {
return info.param.test_name;
});
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
Reference in New Issue View Git Blame Copy Permalink