From 4c30de056afe8680b42723b26a2241811715b989 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Fri, 7 Jul 2023 12:01:32 +0100 Subject: Enable premultiplication for depthwise convolution with fp16 and quantized types Resolves: COMPMID-6337 Change-Id: I81542e51c9c0329f202ac8452f173b138e51a0f6 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9883 Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/arm_conv/depthwise/depthwise_u8q.cpp | 35 +++++++++++----------- 1 file changed, 17 insertions(+), 18 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp index 9dbd89fb52..236930ee26 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021-2022 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -71,6 +71,16 @@ using arm_gemm::Requantize32; namespace arm_conv { namespace depthwise { +namespace +{ +#if defined(__aarch64__) +uint64_t not_preferred(const DepthwiseArgs &, const Requantize32 &) +{ + return std::numeric_limits::max(); +} +#endif // defined(__aarch64__) +} + static const DepthwiseImplementation depthwise_u8q_methods[] = { #if defined(__aarch64__) #if defined(ARM_COMPUTE_ENABLE_SVE) @@ -132,7 +142,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "sve_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift, cpu_has_sve2), nullptr, @@ -145,7 +154,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "sve_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift, cpu_has_sve2), nullptr, @@ -158,7 +166,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift, cpu_has_sve2), nullptr, @@ -171,7 +178,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "sve_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift, cpu_has_sve2), nullptr, @@ -187,7 +193,7 @@ static const DepthwiseImplementation de qp_has_no_left_shift, has_channel_multiplier, cpu_has_sve2), - nullptr, + not_preferred, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto strat = new sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info); return new DepthwiseDepthfirstMultiplier(strat, args, qp); @@ -200,7 +206,7 @@ static const DepthwiseImplementation de qp_has_no_left_shift, has_channel_multiplier, cpu_has_sve2), - nullptr, + not_preferred, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto strat = new sve_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info); return new DepthwiseDepthfirstMultiplier(strat, args, qp); @@ -212,7 +218,6 @@ static const DepthwiseImplementation de "a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst", constraint(is_supported, cpu_has_dot_product, - has_no_channel_multiplier, qp_has_no_left_shift), nullptr, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { @@ -225,7 +230,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_zero_a_offset, qp_has_no_left_shift), nullptr, @@ -238,7 +242,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8qa_nhwc_3x3_s2_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_zero_a_offset, qp_has_no_left_shift), nullptr, @@ -251,7 +254,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8qa_nhwc_5x5_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_zero_a_offset, qp_has_no_left_shift), nullptr, @@ -265,7 +267,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift), nullptr, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { @@ -277,7 +278,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift), nullptr, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { @@ -289,7 +289,6 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst", constraint(is_supported, - has_no_channel_multiplier, qp_has_no_left_shift), nullptr, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { @@ -300,7 +299,7 @@ static const DepthwiseImplementation de { DepthwiseMethod::DEPTHFIRST, "a64_u8q_nhwc_generic_output3x3_mla_depthfirst", - constraint(has_no_channel_multiplier), + nullptr, nullptr, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto kernel = new a64_u8q_nhwc_generic_output9_mla_depthfirst(args.cpu_info); @@ -315,7 +314,7 @@ static const DepthwiseImplementation de cpu_has_dot_product, has_channel_multiplier, qp_has_no_left_shift), - nullptr, + not_preferred, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto strat = new a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst(args.cpu_info); return new DepthwiseDepthfirstMultiplier(strat, args, qp); @@ -328,7 +327,7 @@ static const DepthwiseImplementation de cpu_has_dot_product, has_channel_multiplier, qp_has_no_left_shift), - nullptr, + not_preferred, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto strat = new a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst(args.cpu_info); return new DepthwiseDepthfirstMultiplier(strat, args, qp); @@ -338,7 +337,7 @@ static const DepthwiseImplementation de DepthwiseMethod::DEPTHFIRST, "a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst", constraint(has_channel_multiplier), - nullptr, + not_preferred, [] (const DepthwiseArgs &args, const Requantize32 &qp) -> DepthwiseCommon * { auto kern = new a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst(args.cpu_info); auto strat = new GenericDepthfirstMultiplierStrategy(kern, args); -- cgit v1.2.1