aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp89
1 files changed, 47 insertions, 42 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index 4de3d2b18a..dfacb687a8 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2020, 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -34,71 +34,75 @@
#include "kernels/a64_gemm_u8_4x4.hpp"
#include "kernels/a64_gemm_u8_8x12.hpp"
#include "kernels/a64_hybrid_u8u32_dot_6x16.hpp"
+#include "kernels/a64_hybrid_u8u32_mmla_6x16.hpp"
#include "kernels/a64_interleaved_u8u32_mmla_8x12.hpp"
#include "kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp"
#include "kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp"
#include "kernels/sve_hybrid_u8u32_dot_6x4VL.hpp"
+#include "kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp"
#include "kernels/sve_interleaved_u8u32_dot_8x3VL.hpp"
#include "kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp"
-#include "kernels/sve_smallK_hybrid_u8u32_dot_8x1VL.hpp"
namespace arm_gemm {
static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = {
-#ifdef __ARM_FEATURE_SVE
-#ifdef MMLA_INT8
-{
+#ifdef ARM_COMPUTE_ENABLE_SVE
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "sve_hybrid_u8u32_mmla_6x4VL",
+ [](const GemmArgs &args) { return args._ci->has_svei8mm(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>(args); }
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
"sve_interleaved_u8u32_mmla_8x3VL",
- [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>8); },
- [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
+ [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>(args); }
-},
-#endif
-{
- GemmMethod::GEMM_HYBRID,
- "smallK_hybrid_u8u32_dot_8x1VL",
- [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize<=64 && !args._indirect_input; },
- [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
- [](const GemmArgs &args) { return new GemmHybrid<cls_sve_smallK_hybrid_u8u32_dot_8x1VL, uint8_t, uint32_t>(args); }
-},
-{
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_HYBRID,
"sve_hybrid_u8u32_dot_6x4VL",
[](const GemmArgs &args) { return args._ci->has_sve(); },
- [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN && (((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8))); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>(args); }
-},
-{
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
"sve_interleaved_u8u32_dot_8x3VL",
[](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); },
- [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>(args); }
-},
-#endif
-#ifdef MMLA_INT8
-{
+),
+#endif // ARM_COMPUTE_ENABLE_SVE
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
"a64_interleaved_u8u32_mmla_8x12",
- [](const GemmArgs &args) { return (args._Ksize>8); },
- nullptr,
+ [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>(args); }
-},
-#endif
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "a64_hybrid_u8u32_mmla_6x16",
+ [](const GemmArgs &args) { return args._ci->has_i8mm(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>(args); }
+),
{
GemmMethod::GEMM_HYBRID,
"a64_smallK_hybrid_u8u32_dot_8x4",
- [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
- nullptr,
+ [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input && !args._accumulate; },
+ [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
[](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_8x4, uint8_t, uint32_t>(args); }
},
{
GemmMethod::GEMM_HYBRID,
"a64_smallK_hybrid_u8u32_dot_6x4",
- [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
- nullptr,
+ [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input && !args._accumulate; },
+ [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
[](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_6x4, uint8_t, uint32_t>(args); }
},
{
@@ -108,27 +112,27 @@ static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = {
[](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; },
[](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint8_t, uint32_t>(args); },
},
-{
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_HYBRID,
"a64_hybrid_u8u32_dot_6x16",
[](const GemmArgs &args) { return args._ci->has_dotprod(); },
- [](const GemmArgs &args) { return args._Nsize<=256 && args._Ksize>128; },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>(args); }
-},
-{
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
"a64_gemm_u8_8x12",
[](const GemmArgs &args) { return args._ci->has_dotprod(); },
- nullptr,
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>(args); }
-},
-{
+),
+GemmImplementation<uint8_t, uint32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
"a64_gemm_u8_4x4",
nullptr,
- nullptr,
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); },
[](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>(args); }
-},
+),
{
GemmMethod::DEFAULT,
"",
@@ -145,6 +149,7 @@ const GemmImplementation<uint8_t, uint32_t> *gemm_implementation_list<uint8_t, u
/* Explicitly instantiate the external functions for these types. */
template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<uint8_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
template KernelDescription get_gemm_method<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &);
template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t, Nothing> (const GemmArgs &args, const Nothing &);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 17:16:15 +0000