From 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Fri, 16 Jul 2021 16:16:43 +0100
Subject: Update GEMM assembly kernels

- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
---
 src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp | 118 ++++++++++++++++++---------
 1 file changed, 79 insertions(+), 39 deletions(-)

(limited to 'src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp')

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
index 1632e301ac..3cf84a614a 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp
@@ -31,17 +31,22 @@
 #include "gemv_pretransposed.hpp"
 
 #include "kernels/a32_sgemm_8x6.hpp"
-#include "kernels/a64_gemv_fp32_mla_32.hpp"
+#include "kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp"
+#include "kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp"
+#include "kernels/a64_hybrid_fp32_mla_4x24.hpp"
 #include "kernels/a64_hybrid_fp32_mla_6x16.hpp"
 #include "kernels/a64_hybrid_fp32_mla_8x4.hpp"
+#include "kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp"
 #include "kernels/a64_sgemm_8x12.hpp"
 #include "kernels/a64_sgemm_8x6.hpp"
 #include "kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp"
 #include "kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp"
 
-#include "kernels/sve_gemv_fp32_mla_8VL.hpp"
+#include "kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp"
+#include "kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp"
 #include "kernels/sve_hybrid_fp32_mla_6x4VL.hpp"
 #include "kernels/sve_hybrid_fp32_mla_8x1VL.hpp"
+#include "kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp"
 #include "kernels/sve_interleaved_fp32_mla_8x3VL.hpp"
 #include "kernels/sve_interleaved_fp32_mmla_8x3VL.hpp"
 #include "kernels/sve_smallK_hybrid_fp32_mla_8x1VL.hpp"
@@ -59,57 +64,94 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] =
     [](const GemmArgs &args) { return new GemvBatched<float, float>(args); }
 },
 #ifdef __aarch64__
+#ifdef ARM_COMPUTE_ENABLE_BF16
+// "fast mode" (BF16) kernels
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_INTERLEAVED,
+    "a64_interleaved_bf16fp32_mmla_8x12",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>(args); }
+),
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_HYBRID,
+    "a64_hybrid_fp32bf16fp32_mmla_6x16",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>(args); }
+),
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_HYBRID,
+    "a64_hybrid_fp32bf16fp32_mmla_4x24",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_SVE
-{
+#ifdef ARM_COMPUTE_ENABLE_BF16
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_INTERLEAVED,
+    "sve_interleaved_bf16fp32_mmla_8x3VL",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); },
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>(args); }
+),
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
-    "sve_gemv_fp32_mla_8VL",
-    [](const GemmArgs &args) { return args._ci->has_sve() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
-    [](const GemmArgs &args) { return new GemvPretransposed<cls_sve_gemv_fp32_mla_8VL, float, float>(args); }
-},
-#endif
-{
+    "sve_hybrid_fp32bf16fp32_mmla_6x4VL",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>(args); }
+),
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
-    "a64_gemv_fp32_mla_32",
-    [](const GemmArgs &args) { return args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
-    nullptr,
-    [](const GemmArgs &args) { return new GemvPretransposed<cls_a64_gemv_fp32_mla_32, float, float>(args); }
-},
-
-// MMLA next due to higher throughput (SVE only)
-#if defined(ARM_COMPUTE_ENABLE_SVE) && defined(ARM_COMPUTE_ENABLE_SVEF32MM)
+    "sve_hybrid_fp32bf16fp32_mmla_4x6VL",
+    [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); }
+),
+#endif // ARM_COMPUTE_ENABLE_BF16
+#ifdef ARM_COMPUTE_ENABLE_SVEF32MM
+// MMLA next due to higher throughput (which is SVE only)
+// Prefer this in all cases, except if fast mode is requested and BF16 is available.
 {
     GemmMethod::GEMM_INTERLEAVED,
     "sve_interleaved_fp32_mmla_8x3VL",
     [](const GemmArgs &args) { return args._ci->has_svef32mm() && (args._Ksize>4); },
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
+    [](const GemmArgs &args) { return !(args._fast_mode && args._ci->has_bf16()); },
     [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mmla_8x3VL, float, float>(args); }
 },
-#endif // ARM_COMPUTE_ENABLE_SVE && ARM_COMPUTE_ENABLE_SVEF32MM
-
-#ifdef ARM_COMPUTE_ENABLE_SVE
-// SVE smallk / hybrid methods
+#endif // ARM_COMPUTE_ENABLE_SVEF32MM
+// SVE kernels
 {
     GemmMethod::GEMM_HYBRID,
     "sve_smallK_hybrid_fp32_mla_8x1VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize <= 24 && !args._indirect_input; },
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN; },
+    nullptr,
     [](const GemmArgs &args) { return new GemmHybrid<cls_sve_smallK_hybrid_fp32_mla_8x1VL, float, float>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp32_mla_8x1VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN && (args._Nsize < 12); },
+    [](const GemmArgs &args) { return (args._Nsize < 12); },
     [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_8x1VL, float, float>(args); }
 },
-{
+GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_hybrid_fp32_mla_6x4VL",
     [](const GemmArgs &args) { return args._ci->has_sve(); },
-    [](const GemmArgs &args) { return args._ci->get_cpu_model() != CPUModel::KLEIN && (((args._Ksize <= 256) && (args._Nsize <= 256)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8))); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>(args); }
-},
+),
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_INTERLEAVED,
+    "sve_interleaved_fp32_mla_8x3VL",
+    [](const GemmArgs &args) { return args._ci->has_sve(); },
+    [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>(args); }
+),
 #endif // ARM_COMPUTE_ENABLE_SVE
 // Cortex-A35 specific kernel - use for any problem on A35, and never in any other cases.
 {
@@ -141,27 +183,25 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] =
     [](const GemmArgs &args) { return (args._Nsize < 12); },
     [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_8x4, float, float>(args); }
 },
+GemmImplementation<float, float>::with_estimate(
+    GemmMethod::GEMM_HYBRID,
+    "a64_hybrid_fp32_mla_4x24",
+    nullptr,
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>::estimate_cycles<float>(args); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>(args); }
+),
 GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_hybrid_fp32_mla_6x16",
     nullptr,
-    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>::estimate_cycles(args, cls_a64_hybrid_fp32_mla_6x16::get_performance_parameters(args._ci)); },
+    [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>(args); }
 ),
-#ifdef ARM_COMPUTE_ENABLE_SVE
-{
-    GemmMethod::GEMM_INTERLEAVED,
-    "sve_interleaved_fp32_mla_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 new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>(args); }
-},
-#endif // ARM_COMPUTE_ENABLE_SVE
 GemmImplementation<float, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,
     "a64_sgemm_8x12",
     nullptr,
-    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float>::estimate_cycles(args, cls_a64_sgemm_8x12::get_performance_parameters(args._ci)); },
+    [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float>::estimate_cycles<float>(args); },
     [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, float, float>(args); }
 ),
 #endif // __aarch64__
-- 
cgit v1.2.1