aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_int8.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/gemm_int8.cpp187
1 files changed, 114 insertions, 73 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index f7d8f65aea..fedda3a47a 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2019 ARM Limited.
+ * Copyright (c) 2017-2020, 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -26,109 +26,149 @@
#include "arm_gemm.hpp"
#include "gemm_common.hpp"
#include "gemm_hybrid.hpp"
+#include "gemm_hybrid_indirect.hpp"
#include "gemm_implementation.hpp"
#include "gemm_interleaved.hpp"
-#include "gemm_native.hpp"
-#include "kernels/a64_gemm_s16_12x8.hpp"
-#include "kernels/a64_gemm_s8_12x8.hpp"
+#include "kernels/a64_gemm_s16_8x12.hpp"
+#include "kernels/a64_gemm_s8_8x12.hpp"
#include "kernels/a64_gemm_s8_4x4.hpp"
-#include "kernels/a64_hybrid_s8s32_dot_16x4.hpp"
-#include "kernels/a64_interleaved_s8s32_mmla_12x8.hpp"
-#include "kernels/a64_smallK_hybrid_s8s32_dot_4x6.hpp"
-#include "kernels/a64_smallK_hybrid_s8s32_dot_4x8.hpp"
-#include "kernels/sve_hybrid_s8s32_dot_4VLx4.hpp"
-#include "kernels/sve_interleaved_s8s32_dot_3VLx8.hpp"
-#include "kernels/sve_interleaved_s8s32_mmla_3VLx8.hpp"
-#include "kernels/sve_native_s8s32_dot_4VLx4.hpp"
-#include "kernels/sve_smallK_hybrid_s8s32_dot_1VLx8.hpp"
+#include "kernels/a64_hybrid_s8s32_dot_6x16.hpp"
+#include "kernels/a64_hybrid_s8s32_mmla_6x16.hpp"
+#include "kernels/a64_interleaved_s8s32_mmla_8x12.hpp"
+#include "kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp"
+#include "kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp"
+
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SME2
+
+#include "kernels/sve_hybrid_s8s32_dot_6x4VL.hpp"
+#include "kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp"
+#include "kernels/sve_interleaved_s8s32_dot_8x3VL.hpp"
+#include "kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SVE
namespace arm_gemm {
static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
-#ifdef __ARM_FEATURE_SVE
-#ifdef V8P6
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+// SME kernels
{
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_s8s32_mmla_3VLx8",
- [](const GemmArgs &args) { return (args._Ksize>8); },
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_s8s32_mmla_3VLx8, int8_t, int32_t>(args); }
-},
-#endif
-{
- GemmMethod::GEMM_HYBRID,
- "smallK_hybrid_s8s32_dot_1VLx8",
- [](const GemmArgs &args) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; },
- nullptr,
- [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_1VLx8, int8_t, int32_t>(args); }
-},
-{
- GemmMethod::GEMM_HYBRID,
- "hybrid_s8s32_dot_4VLx4",
- [](const GemmArgs &args) { return args._Ksize>=16 && !args._trA && args._pretransposed_hint; },
- [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_s8s32_dot_4VLx4, int8_t, int32_t>(args); }
-},
-{
- GemmMethod::GEMM_NATIVE,
- "native_s8s32_dot_4VLx4",
- [](const GemmArgs &args) { return (args._Ksize>=16 && !args._trA && !args._trB); },
- [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
- [](const GemmArgs &args) { return new GemmNative<native_s8s32_dot_4VLx4, int8_t, int32_t>(args); }
+ "sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
+ [](const GemmArgs &args) { const auto VL = sme::get_vector_length<int32_t>();
+ return args._Nsize >= 8*VL || args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
+ [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL, int8_t, int32_t>(args); }
},
{
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_s8s32_dot_3VLx8",
- [](const GemmArgs &args) { return (args._Ksize>4); },
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_s8s32_dot_3VLx8, int8_t, int32_t>(args); }
+ "sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
+ [](const GemmArgs &args) { const auto VL = sme::get_vector_length<int32_t>();
+ return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
+ [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL, int8_t, int32_t>(args); }
},
-#endif
-#ifdef V8P6
{
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_s8s32_mmla_12x8",
- [](const GemmArgs &args) { return (args._Ksize>8); },
+ "sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_s8s32_mmla_12x8, int8_t, int32_t>(args); }
+ [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL, int8_t, int32_t>(args); }
},
-#endif
-{
+#endif // ARM_COMPUTE_ENABLE_SME2
+GemmImplementation<int8_t, int32_t>::with_estimate(
GemmMethod::GEMM_HYBRID,
- "smallK_hybrid_s8s32_dot_4x8",
- [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; },
- nullptr,
- [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_4x8, int8_t, int32_t>(args); }
-},
+ "sve_hybrid_s8s32_mmla_6x4VL",
+ [](const GemmArgs &args) { return args._ci->has_svei8mm(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "sve_interleaved_s8s32_mmla_8x3VL",
+ [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "sve_hybrid_s8s32_dot_6x4VL",
+ [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize>=16; },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "sve_interleaved_s8s32_dot_8x3VL",
+ [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_SVE
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_interleaved_s8s32_mmla_8x12",
+ [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "a64_hybrid_s8s32_mmla_6x16",
+ [](const GemmArgs &args) { return args._ci->has_i8mm(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>(args); }
+),
{
GemmMethod::GEMM_HYBRID,
- "smallK_hybrid_s8s32_dot_4x6",
- [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; },
- nullptr,
- [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_4x6, int8_t, int32_t>(args); }
+ "a64_smallK_hybrid_s8s32_dot_8x4",
+ [](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_s8s32_dot_8x4, int8_t, int32_t>(args); }
},
{
GemmMethod::GEMM_HYBRID,
- "hybrid_s8s32_dot_16x4",
- [](const GemmArgs &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
- [](const GemmArgs &args) { return args._Nsize<=256 && args._Ksize>128; },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_s8s32_dot_16x4, int8_t, int32_t>(args); }
+ "a64_smallK_hybrid_s8s32_dot_6x4",
+ [](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_s8s32_dot_6x4, int8_t, int32_t>(args); }
},
{
GemmMethod::GEMM_INTERLEAVED,
- "gemm_s8_12x8",
- [](const GemmArgs &args) { return args._ci->has_dotprod(); },
+ "a64_gemm_s16_8x12",
nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<gemm_s8_12x8, int8_t, int32_t>(args); }
+ [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int8_t, int32_t>(args); },
},
-{
+GemmImplementation<int8_t, int32_t>::with_estimate(
+
+ GemmMethod::GEMM_HYBRID,
+ "a64_hybrid_s8s32_dot_6x16",
+ [](const GemmArgs &args) { return args._ci->has_dotprod(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "gemm_s8_4x4",
- nullptr,
+ "a64_gemm_s8_8x12",
+ [](const GemmArgs &args) { return args._ci->has_dotprod(); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>(args); }
+),
+GemmImplementation<int8_t, int32_t>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_gemm_s8_4x4",
nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<gemm_s8_4x4, int8_t, int32_t>(args); }
-},
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>::estimate_cycles<int32_t>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>(args); }
+),
+
{
GemmMethod::DEFAULT,
"",
@@ -145,6 +185,7 @@ const GemmImplementation<int8_t, int32_t> *gemm_implementation_list<int8_t, int3
/* Explicitly instantiate the external functions for these types. */
template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<int8_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
template KernelDescription get_gemm_method<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &);
template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-05 13:13:57 +0000