COMPMID-2923 Integrate arm_gemm per channel quantization
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Change-Id: I8667e75843fdd6ac75bd8272a86a348b830da28d
Reviewed-on: https://review.mlplatform.org/c/2548
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index 079c04a..59cd170 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -36,51 +36,51 @@
namespace arm_gemm {
-static const GemmImplementation<uint8_t, uint8_t, ARequantizeLayer32> gemm_quint8_methods[] =
+static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
{
#ifdef __ARM_FEATURE_SVE
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_1VLx8",
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; },
+ [](const GemmArgs &args, const Requantize32 &) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; },
nullptr,
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint8_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"hybrid_u8u32_dot_4VLx4",
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_4VLx4, uint8_t, uint8_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &) { return args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
+ [](const GemmArgs &args, const Requantize32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
+ [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_4VLx4, uint8_t, uint8_t>(args, qp); }
},
#endif
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_4x8",
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; },
+ [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; },
nullptr,
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint8_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_4x6",
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; },
+ [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; },
nullptr,
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint8_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"hybrid_u8u32_dot_16x4",
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
- [](const GemmArgs &args, const ARequantizeLayer32 &) { return args._Nsize<=256 && args._Ksize>128; },
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_16x4, uint8_t, uint8_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
+ [](const GemmArgs &args, const Requantize32 &) { return args._Nsize<=256 && args._Ksize>128; },
+ [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_16x4, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::QUANTIZE_WRAPPER,
"quantized_wrapper",
nullptr,
nullptr,
- [](const GemmArgs &args, const ARequantizeLayer32 &qp) { return new QuantizeWrapper<uint8_t, uint8_t, uint32_t>(args, qp); }
+ [](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<uint8_t, uint8_t, uint32_t>(args, qp); }
},
{
GemmMethod::DEFAULT,
@@ -92,13 +92,13 @@
};
template<>
-const GemmImplementation<uint8_t, uint8_t, ARequantizeLayer32> *gemm_implementation_list<uint8_t, uint8_t, ARequantizeLayer32>() {
+const GemmImplementation<uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, Requantize32>() {
return gemm_quint8_methods;
}
-template UniqueGemmCommon<uint8_t, uint8_t> gemm<uint8_t, uint8_t, ARequantizeLayer32>(const GemmArgs &args, const ARequantizeLayer32 &os);
-template KernelDescription get_gemm_method<uint8_t, uint8_t, ARequantizeLayer32>(const GemmArgs &args, const ARequantizeLayer32 &os);
-template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, ARequantizeLayer32>(const GemmArgs &args, const ARequantizeLayer32 &os);
+template UniqueGemmCommon<uint8_t, uint8_t> gemm<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template KernelDescription get_gemm_method<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
+template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
} // namespace arm_gemm