aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp249
1 files changed, 165 insertions, 84 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
index 10fee472f4..0ddca04846 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2020, 2022-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -22,120 +22,200 @@
* SOFTWARE.
*/
#include "arm_gemm.hpp"
+#include "bfloat.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 "gemv_batched.hpp"
-#include "gemv_native_transposed.hpp"
#include "gemv_pretransposed.hpp"
-#include "kernels/a64_interleaved_bf16fp32_dot_12x8.hpp"
-#include "kernels/a64_interleaved_bf16fp32_mmla_12x8.hpp"
-#include "kernels/a64_sgemm_12x8.hpp"
#include "kernels/a32_sgemm_8x6.hpp"
-#include "kernels/sve_interleaved_bf16fp32_dot_3VLx8.hpp"
-#include "kernels/sve_interleaved_bf16fp32_mmla_3VLx8.hpp"
-#include "kernels/sve_native_bf16fp32_dot_4VLx4.hpp"
-#include "kernels/sve_hybrid_bf16fp32_dot_4VLx4.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_4VLx4.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_6VLx2.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_8VLx2.hpp"
-#include "bfloat.hpp"
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#include "kernels/a64_ffhybrid_bf16fp32_mmla_6x16.hpp"
+#include "kernels/a64_ffinterleaved_bf16fp32_dot_8x12.hpp"
+#include "kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp"
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#include "kernels/a64_hybrid_bf16fp32_dot_6x16.hpp"
+#include "kernels/a64_hybrid_bf16fp32_mmla_6x16.hpp"
+#include "kernels/a64_interleaved_bf16fp32_dot_8x12.hpp"
+#include "kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp"
+#include "kernels/a64_sgemm_8x12.hpp"
-namespace arm_gemm {
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#include "kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp"
+#include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+#include "kernels/sme2_gemv_bf16fp32_dot_16VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SME2
+
+#include "kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp"
+#include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp"
+#include "kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp"
+#include "kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp"
+#include "kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp"
+#include "kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SVE
+namespace arm_gemm {
static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] =
{
-#ifdef V8P6_BF
-# ifdef __ARM_FEATURE_SVE
+#ifdef __aarch64__
+#ifdef ARM_COMPUTE_ENABLE_BF16
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+// SME kernels
{
GemmMethod::GEMM_HYBRID,
- "hybrid_bf16fp32_mmla_6VLx2",
- [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); },
- [](const GemmArgs &args) { return ((args._Msize <= 4) && (args._Nsize <= hybrid_bf16fp32_mmla_6VLx2::out_width())); },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_6VLx2, bfloat16, float>(args); }
+ "sme2_gemv_bf16fp32_dot_16VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
+ nullptr,
+ [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_bf16fp32_dot_16VL, bfloat16, float>(args); }
},
{
- GemmMethod::GEMM_HYBRID,
- "hybrid_bf16fp32_mmla_8VLx2",
- [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); },
- [](const GemmArgs &args) { return (args._Msize <= 4); },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_8VLx2, bfloat16, float>(args); }
+ GemmMethod::GEMM_INTERLEAVED,
+ "sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
+ [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
+ 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_bf16fp32_mopa_1VLx4VL, bfloat16, float>(args); }
},
{
- GemmMethod::GEMM_HYBRID,
- "hybrid_bf16fp32_mmla_4VLx4",
- [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); },
- [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_4VLx4, bfloat16, float>(args); }
+ GemmMethod::GEMM_INTERLEAVED,
+ "sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
+ [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
+ return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
+ [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL, bfloat16, float>(args); }
},
{
- GemmMethod::GEMM_HYBRID,
- "hybrid_bf16fp32_dot_4VLx4",
- [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && args._pretransposed_hint); },
- [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
- [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_dot_4VLx4, bfloat16, float>(args); }
-},
-{ // gemm_bf16_native
- GemmMethod::GEMM_NATIVE,
- "native_bf16fp32_dot_4VLx4",
- [](const GemmArgs &args) { return (args._Ksize>=8 && !args._trA && !args._trB); },
- [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
- [](const GemmArgs &args) { return new GemmNative<native_bf16fp32_dot_4VLx4, bfloat16, float>(args); }
-},
-{ // gemm_bf16_interleaved
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_bf16fp32_mmla_3VLx8",
- [](const GemmArgs &args) { return (args._Ksize>4); },
+ "sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL",
+ [](const GemmArgs &args) { return args._ci->has_sme2(); },
nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_mmla_3VLx8, bfloat16, float>(args); }
+ [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL, bfloat16, float>(args); }
},
-{ // gemm_bf16_interleaved
+#endif // ARM_COMPUTE_ENABLE_SME2
+// gemm_bf16_interleaved
+GemmImplementation<bfloat16, float>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_bf16fp32_dot_3VLx8",
- [](const GemmArgs &args) { return (args._Ksize>2); },
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_dot_3VLx8, bfloat16, float>(args); }
-},
-# endif // SVE
-{ // gemm_bf16_interleaved
+ "sve_interleaved_bf16fp32_mmla_8x3VL",
+ [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>4); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "sve_hybrid_bf16fp32_mmla_6x4VL",
+ [](const GemmArgs &args) { return args._ci->has_svebf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "sve_hybrid_bf16fp32_dot_6x4VL",
+ [](const GemmArgs &args) { return args._ci->has_svebf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_bf16fp32_mmla_12x8",
- [](const GemmArgs &args) { return (args._Ksize>4); },
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_mmla_12x8, bfloat16, float>(args); }
-},
-{ // gemm_bf16_interleaved
+ "sve_interleaved_bf16fp32_dot_8x3VL",
+ [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>2); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); }
+),
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+GemmImplementation<bfloat16, float>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "interleaved_bf16fp32_dot_12x8",
- [](const GemmArgs &args) { return (args._Ksize>2); },
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_dot_12x8, bfloat16, float>(args); }
-},
-#endif // V8P6_BF
-#ifdef __aarch64__
-{
+ "sve_ffinterleaved_bf16fp32_mmla_8x3VL",
+ KernelWeightFormat::VL2VL_BL64,
+ [](const GemmArgs &args) { return args._ci->has_svebf16(); },
+ [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "sgemm_12x8",
- nullptr,
- nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<sgemm_12x8, bfloat16, float>(args); }
-},
-#elif defined(__arm__)
-{
+ "sve_ffhybrid_bf16fp32_mmla_6x4VL",
+ KernelWeightFormat::VL2VL_BL64,
+ [](const GemmArgs &args) { return args._ci->has_svebf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_bf16fp32_mmla_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_bf16fp32_mmla_6x4VL, bfloat16, float>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+#endif // ARM_COMPUTE_ENABLE_SVE
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "a64_hybrid_bf16fp32_mmla_6x16",
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
GemmMethod::GEMM_INTERLEAVED,
- "sgemm_8x6",
- nullptr,
+ "a64_interleaved_bf16fp32_mmla_8x12",
+ [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>4); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_HYBRID,
+ "a64_hybrid_bf16fp32_dot_6x16",
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_interleaved_bf16fp32_dot_8x12",
+ [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>2); },
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>(args); }
+),
+#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_ffinterleaved_bf16fp32_mmla_8x12",
+ KernelWeightFormat::VL256_BL64,
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_ffhybrid_bf16fp32_mmla_6x16",
+ KernelWeightFormat::VL256_BL64,
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); }
+),
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_ffinterleaved_bf16fp32_dot_8x12",
+ KernelWeightFormat::VL128_BL32,
+ [](const GemmArgs &args) { return args._ci->has_bf16(); },
+ [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_dot_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_dot_8x12, bfloat16, float>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+GemmImplementation<bfloat16, float>::with_estimate(
+ GemmMethod::GEMM_INTERLEAVED,
+ "a64_sgemm_8x12",
nullptr,
- [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, bfloat16, float>(args); }
-},
-#else
-# error "Unknown Architecture"
-#endif
+ [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); },
+ [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_BF16
+#endif // __aarch64__
{
GemmMethod::DEFAULT,
"",
@@ -152,6 +232,7 @@ const GemmImplementation<bfloat16, float> *gemm_implementation_list<bfloat16, fl
/* Explicitly instantiate the external functions for these types. */
template UniqueGemmCommon<bfloat16, float> gemm<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
+template bool has_opt_gemm<bfloat16, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &);
template KernelDescription get_gemm_method<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
template std::vector<KernelDescription> get_compatible_kernels<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &);
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-05 08:47:18 +0000