From 868f86ef8a98a8698b656b350dc8e8e3b66b8f69 Mon Sep 17 00:00:00 2001 From: Cursor Agent Date: Tue, 29 Sep 2026 15:10:31 +0000 Subject: [PATCH] -remove class AmxBf16::SynetQuantizedConvolutionNhwcSpecV0. SynetQuantizedConvolutionInit no longer selects this path. NHWC quantized convolutions that used it fall back to Avx512vnni::SynetQuantizedConvolutionNhwcSpecV0. Co-authored-by: igor.ermolaev --- docs/2026.html | 1 + prj/vs2022/AmxBf16.vcxproj | 1 - prj/vs2022/AmxBf16.vcxproj.filters | 3 - .../SimdAmxBf16SynetQuantizedConvolution.cpp | 2 - ...f16SynetQuantizedConvolutionNhwcSpecV0.cpp | 258 ------------------ src/Simd/SimdSynetQuantizedConvolution.h | 10 - 6 files changed, 1 insertion(+), 274 deletions(-) delete mode 100644 src/Simd/SimdAmxBf16SynetQuantizedConvolutionNhwcSpecV0.cpp diff --git a/docs/2026.html b/docs/2026.html index fc713f380d..8eb3e010e9 100644 --- a/docs/2026.html +++ b/docs/2026.html @@ -158,6 +158,7 @@
Removing
  • Class Base::SynetConvolution16bNhwcSpecV2.
  • Class AmxBf16::SynetConvolution16bNhwcSpecV2.
  • Class AmxBf16::SynetQuantizedConvolutionNhwcGemmV0.
    • +
  • Class AmxBf16::SynetQuantizedConvolutionNhwcSpecV0.

    • Python wrapper

    diff --git a/prj/vs2022/AmxBf16.vcxproj b/prj/vs2022/AmxBf16.vcxproj index d89d90a559..afb6093536 100644 --- a/prj/vs2022/AmxBf16.vcxproj +++ b/prj/vs2022/AmxBf16.vcxproj @@ -94,7 +94,6 @@ - diff --git a/prj/vs2022/AmxBf16.vcxproj.filters b/prj/vs2022/AmxBf16.vcxproj.filters index 3927f57a34..be88d0251c 100644 --- a/prj/vs2022/AmxBf16.vcxproj.filters +++ b/prj/vs2022/AmxBf16.vcxproj.filters @@ -266,9 +266,6 @@ AmxBf16\Synet\Quantized - - AmxBf16\Synet\Quantized - AmxBf16\Synet\Quantized diff --git a/src/Simd/SimdAmxBf16SynetQuantizedConvolution.cpp b/src/Simd/SimdAmxBf16SynetQuantizedConvolution.cpp index 15d42c67b3..457faba982 100644 --- a/src/Simd/SimdAmxBf16SynetQuantizedConvolution.cpp +++ b/src/Simd/SimdAmxBf16SynetQuantizedConvolution.cpp @@ -42,8 +42,6 @@ namespace Simd return NULL; else if (SynetQuantizedConvolutionNhwcSpecV1::Preferable(param)) return new SynetQuantizedConvolutionNhwcSpecV1(param); - else if (SynetQuantizedConvolutionNhwcSpecV0::Preferable(param)) - return new SynetQuantizedConvolutionNhwcSpecV0(param); else if (SynetQuantizedConvolutionNhwcGemmV1::Preferable(param)) return new SynetQuantizedConvolutionNhwcGemmV1(param); else if (SynetQuantizedConvolutionNchwGemm::Preferable(param)) diff --git a/src/Simd/SimdAmxBf16SynetQuantizedConvolutionNhwcSpecV0.cpp b/src/Simd/SimdAmxBf16SynetQuantizedConvolutionNhwcSpecV0.cpp deleted file mode 100644 index 229bee2ea2..0000000000 --- a/src/Simd/SimdAmxBf16SynetQuantizedConvolutionNhwcSpecV0.cpp +++ /dev/null @@ -1,258 +0,0 @@ -/* -* Simd Library (http://ermig1979.github.io/Simd). -* -* Copyright (c) 2011-2025 Yermalayeu Ihar. -* -* Permission is hereby granted, free of charge, to any person obtaining a copy -* of this software and associated documentation files (the "Software"), to deal -* in the Software without restriction, including without limitation the rights -* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -* copies of the Software, and to permit persons to whom the Software is -* furnished to do so, subject to the following conditions: -* -* The above copyright notice and this permission notice shall be included in -* all copies or substantial portions of the Software. -* -* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -* SOFTWARE. -*/ -#include "Simd/SimdSynetQuantizedConvolution.h" -#include "Simd/SimdSynetQuantizeLinear.h" -#include "Simd/SimdSynetConvolution8iCommon.h" -#include "Simd/SimdSynet.h" -#include "Simd/SimdMath.h" -#include "Simd/SimdBase.h" -#include "Simd/SimdCpu.h" -#include "Simd/SimdLog.h" -#include "Simd/SimdSet.h" -#include "Simd/SimdTile.h" - -namespace Simd -{ -#if defined(SIMD_AMXBF16_ENABLE) && defined(SIMD_SYNET_ENABLE) - namespace AmxBf16 - { - typedef Base::SynetQuantizedConvolutionNhwcSpecV0::AlgParam AlgParam; - typedef Base::SynetQuantizedConvolutionNhwcSpecV0::ConvolutionPtr Convolution; - - //----------------------------------------------------------------------------------------- - - static void QuantizedConvolutionNhwcSpecV0_32x32(const uint8_t* src0, const ConvParam& p, const AlgParam& a, const int* offs, size_t nK, int update, const int8_t* weight0, int32_t* dst0) - { - int dD = (int)a.macroD, dX = (int)a.microC, strideS = dX, dW = 1024, strideW = 64, strideD = dD * 4; - const int8_t* weight1 = weight0 + a.K * F; - const uint8_t* src1 = src0 + 16 * dX; - int32_t* dst1 = dst0 + 16 * dD; - - if (update) - { - _tile_stream_loadd(0, dst0 + 0, strideD); - _tile_stream_loadd(1, dst0 + F, strideD); - _tile_stream_loadd(2, dst1 + 0, strideD); - _tile_stream_loadd(3, dst1 + F, strideD); - } - else - { - _tile_zero(0); - _tile_zero(1); - _tile_zero(2); - _tile_zero(3); - } - - int n1 = (int)nK - 1, o = offs[0]; - _tile_stream_loadd(4, src0 + o, strideS); - _tile_loadd(6, weight0, strideW); - for (int i = 0; i < n1; ++i) - { - _tile_stream_loadd(5, src1 + o, strideS); - _tile_loadd(7, weight1, strideW); - _tile_dpbusd(0, 4, 6); - _tile_dpbusd(1, 4, 7); - o = offs[i + 1]; - _tile_stream_loadd(4, src0 + o, strideS); - _tile_dpbusd(2, 5, 6); - weight0 += dW; - _tile_loadd(6, weight0, strideW); - _tile_dpbusd(3, 5, 7); - weight1 += dW; - } - _tile_loadd(7, weight1, strideW); - _tile_stream_loadd(5, src1 + offs[n1], strideS); - - _tile_dpbusd(0, 4, 6); - _tile_stored(0, dst0 + 0, strideD); - TileMoveToMemory(dst0 + 0, dD); - - _tile_dpbusd(1, 4, 7); - _tile_stored(1, dst0 + F, strideD); - TileMoveToMemory(dst0 + F, dD); - - _tile_dpbusd(2, 5, 6); - _tile_stored(2, dst1 + 0, strideD); - TileMoveToMemory(dst1 + 0, dD); - - _tile_dpbusd(3, 5, 7); - _tile_stored(3, dst1 + F, strideD); - TileMoveToMemory(dst1 + F, dD); - } - - static void QuantizedConvolutionNhwcSpecV0_32x16(const uint8_t* src0, const ConvParam& p, const AlgParam& a, const int* offs, size_t nK, int update, const int8_t* weight0, int32_t* dst0) - { - int dD = (int)a.macroD, dX = (int)a.microC, strideS = dX, dW = 1024, strideW = 64, strideD = dD * 4; - const uint8_t* src1 = src0 + 16 * dX; - int32_t* dst1 = dst0 + 16 * dD; - - if (update) - { - _tile_stream_loadd(0, dst0 + 0, strideD); - _tile_stream_loadd(2, dst1 + 0, strideD); - } - else - { - _tile_zero(0); - _tile_zero(2); - } - - int n1 = (int)nK - 1, o = offs[0]; - _tile_loadd(4, src0 + o, strideS); - for (int i = 0; i < n1; ++i) - { - _tile_stream_loadd(6, weight0, strideW); - _tile_loadd(5, src1 + o, strideS); - _tile_dpbusd(0, 4, 6); - o = offs[i + 1]; - _tile_loadd(4, src0 + o, strideS); - _tile_dpbusd(2, 5, 6); - weight0 += dW; - } - _tile_stream_loadd(6, weight0, strideW); - _tile_loadd(5, src1 + offs[n1], strideS); - - _tile_dpbusd(0, 4, 6); - _tile_stored(0, dst0 + 0, strideD); - TileMoveToMemory(dst0 + 0, dD); - - _tile_dpbusd(2, 5, 6); - _tile_stored(2, dst1 + 0, strideD); - TileMoveToMemory(dst1 + 0, dD); - } - - static void QuantizedConvolutionNhwcSpecV0_16x32(const uint8_t* src0, const ConvParam& p, const AlgParam& a, const int* offs, size_t nK, int update, const int8_t* weight0, int32_t* dst0) - { - int dD = (int)a.macroD, dX = (int)a.microC, strideS = dX, dW = 1024, strideW = 64, strideD = dD * 4; - const int8_t* weight1 = weight0 + a.K * F; - - if (update) - { - _tile_stream_loadd(0, dst0 + 0, strideD); - _tile_stream_loadd(1, dst0 + F, strideD); - } - else - { - _tile_zero(0); - _tile_zero(1); - } - - int n1 = (int)nK - 1; - _tile_loadd(6, weight0, strideW); - for (int i = 0; i < n1; ++i) - { - _tile_stream_loadd(4, src0 + offs[i], strideS); - _tile_loadd(7, weight1, strideW); - _tile_dpbusd(0, 4, 6); - weight0 += dW; - _tile_loadd(6, weight0, strideW); - _tile_dpbusd(1, 4, 7); - weight1 += dW; - } - _tile_stream_loadd(4, src0 + offs[n1], strideS); - _tile_loadd(7, weight1, strideW); - - _tile_dpbusd(0, 4, 6); - _tile_stored(0, dst0 + 0, strideD); - TileMoveToMemory(dst0 + 0, dD); - - _tile_dpbusd(1, 4, 7); - _tile_stored(1, dst0 + F, strideD); - TileMoveToMemory(dst0 + F, dD); - } - - static void QuantizedConvolutionNhwcSpecV0_16x16(const uint8_t* src0, const ConvParam& p, const AlgParam& a, const int* offs, size_t nK, int update, const int8_t* weight0, int32_t* dst0) - { - int dD = (int)a.macroD, dX = (int)a.microC, strideS = dX, dW = 1024, strideW = 64, strideD = dD * 4; - - if (update) - { - _tile_stream_loadd(0, dst0 + 0, strideD); - } - else - { - _tile_zero(0); - } - - int n = (int)nK; - for (int i = 0; i < n; ++i) - { - _tile_stream_loadd(4, src0 + offs[i], strideS); - _tile_loadd(6, weight0, strideW); - _tile_dpbusd(0, 4, 6); - weight0 += dW; - } - - _tile_stored(0, dst0 + 0, strideD); - TileMoveToMemory(dst0 + 0, dD); - } - - typedef void (*QuantizedConvolutionNhwcSpecV0Ptr)(const uint8_t* src0, const ConvParam& p, const AlgParam& a, const int* offset, size_t nK, int update, const int8_t* weight0, int32_t* dst0); - - static void QuantizedConvolutionNhwcSpecV0_2(const uint8_t* src, const ConvParam& p, const AlgParam& a, const int* offs, size_t dstC, size_t dstH, size_t nK, int update, const int8_t* weight, int32_t* dst) - { - size_t n1 = dstH * a.srcW - a.gapH, n = 32; - size_t nn = AlignLoAny(n1, n), m = n1 - nn, dW = a.K * DF; - size_t dD = a.macroD, dS = a.microC; - QuantizedConvolutionNhwcSpecV0Ptr body_2 = QuantizedConvolutionNhwcSpecV0_32x32; - QuantizedConvolutionNhwcSpecV0Ptr tail_2 = m > 16 ? QuantizedConvolutionNhwcSpecV0_32x32 : QuantizedConvolutionNhwcSpecV0_16x32; - QuantizedConvolutionNhwcSpecV0Ptr body_1 = QuantizedConvolutionNhwcSpecV0_32x16; - QuantizedConvolutionNhwcSpecV0Ptr tail_1 = m > 16 ? QuantizedConvolutionNhwcSpecV0_32x16 : QuantizedConvolutionNhwcSpecV0_16x16; - - SetTileConfFull(); - for (size_t dc = 0; dc < dstC; dc += DF) - { - size_t dC = Simd::Min(DF, dstC - dc); - size_t i = 0; - if (dC > F) - { - for (; i < nn; i += n) - body_2(src + i * dS, p, a, offs, nK, update, weight, dst + i * dD); - if (m) - tail_2(src + i * dS, p, a, offs, nK, update, weight, dst + i * dD); - } - else - { - for (; i < nn; i += n) - body_1(src + i * dS, p, a, offs, nK, update, weight, dst + i * dD); - if (m) - tail_1(src + i * dS, p, a, offs, nK, update, weight, dst + i * dD); - } - weight += dW; - dst += DF; - } - } - - //----------------------------------------------------------------------------------------- - - SynetQuantizedConvolutionNhwcSpecV0::SynetQuantizedConvolutionNhwcSpecV0(const ConvParam& p) - : Avx512vnni::SynetQuantizedConvolutionNhwcSpecV0(p) - { - SetAlgParam(F, F * 2, F * 2, F * 4, Base::AlgCacheL1(), int(Base::AlgCacheL2() * 0.5), Base::AlgCacheL3()); - AlgParam& a = _alg; - _convolution = QuantizedConvolutionNhwcSpecV0_2; - } - } -#endif -} diff --git a/src/Simd/SimdSynetQuantizedConvolution.h b/src/Simd/SimdSynetQuantizedConvolution.h index d4ce4cc62c..bee5913b44 100644 --- a/src/Simd/SimdSynetQuantizedConvolution.h +++ b/src/Simd/SimdSynetQuantizedConvolution.h @@ -759,16 +759,6 @@ namespace Simd virtual String Ext() const { return "AmxBf16"; } }; - //------------------------------------------------------------------------------------------------ - - class SynetQuantizedConvolutionNhwcSpecV0 : public Avx512vnni::SynetQuantizedConvolutionNhwcSpecV0 - { - public: - SynetQuantizedConvolutionNhwcSpecV0(const ConvParam& p); - - virtual String Ext() const { return _alg.microC == 64 ? "AmxBf16" : "Avx512vnni"; } - }; - //------------------------------------------------------------------------------------------------ class SynetQuantizedConvolutionNhwcSpecV1 : public Base::SynetQuantizedConvolutionNhwcSpecV1