wehub-resource-sync
5030 lines
189 KiB
C++
5030 lines
189 KiB
C++
//
|
|
// CommonOptFunction.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2018/09/06.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "CommonOptFunction.h"
|
|
#include "ConvOpt.h"
|
|
#include "WinogradOptFunction.hpp"
|
|
#include "Int8FunctionsOpt.h"
|
|
#include "ImageProcessFunction.hpp"
|
|
#include <string.h>
|
|
#include <algorithm>
|
|
#include <cmath>
|
|
#include <math.h>
|
|
#include "math/Vec.hpp"
|
|
#include <vector>
|
|
#include <cstdlib>
|
|
#include "../CPURuntime.hpp"
|
|
#include "core/MemoryFormater.h"
|
|
// TODO: Find better way to optimize it
|
|
#include "../CPUBinary.hpp"
|
|
#include "../CPUUnary.hpp"
|
|
#include "../CPUPool.hpp"
|
|
#define PACK 4
|
|
#define FLOAT float
|
|
using Vec = MNN::Math::Vec<float, 4>;
|
|
#include "../GridSampler.hpp"
|
|
#ifdef MNN_LOW_MEMORY
|
|
#ifdef __aarch64__
|
|
#include "backend/cpu/arm/arm64/low_memory/MNNDynamicQuantFunctions.hpp"
|
|
#endif
|
|
#endif
|
|
|
|
#ifdef MNN_USE_RVV
|
|
extern void MNNAbsMaxFP32_RVV(const float* source, float* absmax, size_t src_depth_quad, size_t realSize, int pack);
|
|
extern void MNNAccumulateSequenceNumber_RVV(float* dst, const float* src, int size);
|
|
extern void MNNAsyQuantFunc_RVV(int8_t* dst, const float* src, float* qscale, float* qbias, const size_t* info);
|
|
extern void MNNAsyQuantInfo_FP32_RVV(float* scale, float* bias, float* qscale, float* qbias, float* dstMin,
|
|
float* dstMax, const float* src, const size_t* info);
|
|
extern void MNNDynamicQuantFP32_RVV(const float* src, int8_t* dst, const float* scale, size_t src_depth_quad,
|
|
size_t realSize, int pack, const float* bias);
|
|
extern void MNNReorderWeightInt4_RVV(uint8_t* dest, const uint8_t* source, int32_t* shape, size_t size,
|
|
float* kernelsum);
|
|
extern void MNNSumByAxisLForMatmul_A_RVV(float* dest, int8_t* source, const float* dequantScale, ssize_t realDstCount,
|
|
SumByAxisParams sumParams);
|
|
extern void MNNSumWeightInt8_RVV(float* kernelsum, int8_t* source, size_t outside, size_t reduceAxis, size_t hP,
|
|
size_t lP);
|
|
extern void generalIm2col_RVV(float* destOrigin, float const** sourceGroup, const int32_t* info, const int32_t* el,
|
|
int LP, int pack);
|
|
extern void MNNDynamicUpdateConvBiasScale_RVV(float* newbias, float* oldbias, float* weightKernelSum, float* inputBias,
|
|
size_t ocQuad);
|
|
extern void MNNPackedMatMulFP32_RVV(float* C, const float* A, const float* B, const size_t* parameter,
|
|
const float* postParameters, const float* bias, const float* k, const float* b);
|
|
extern void MNNPackedMatMulRemainFP32_RVV(float* C, const float* A, const float* B, size_t eSize,
|
|
const size_t* parameter, const float* postParameters, const float* bias,
|
|
const float* k, const float* b);
|
|
extern void MNNPackForMatMul_B_RVV(float* destC, const float* sourceC, size_t h, size_t kernelsize, size_t ic,
|
|
bool transpose);
|
|
extern void MNNQuantScaleFP32_RVV(float* absmax, float* quant_scale, float* dequant_scale, size_t thread, size_t batch);
|
|
extern void MNNGetMatMulPackMode_RVV(int* eP, int* lP, int* hP);
|
|
#endif
|
|
|
|
#ifndef MNN_USE_SSE
|
|
void MNNInt8ToInt16(int16_t* dest, const int8_t* source, size_t count) {
|
|
// Should not be called
|
|
MNN_ASSERT(false);
|
|
}
|
|
#endif
|
|
|
|
#ifndef __aarch64__
|
|
|
|
#ifdef MNN_LOW_MEMORY
|
|
void MNNQuantScaleFP32(float* absmax, float* quant_scale, float* dequant_scale, size_t thread, size_t batch) {
|
|
for (int i = 0; i < batch; ++i) {
|
|
auto absmaxPtr = absmax + i;
|
|
float absVal = 0.f;
|
|
for (int t = 0; t < thread; ++t) {
|
|
absVal = std::max(absVal, absmaxPtr[t * batch]);
|
|
}
|
|
if (absVal < 1e-7) {
|
|
quant_scale[i] = 1.f;
|
|
dequant_scale[i] = 1.f;
|
|
} else {
|
|
quant_scale[i] = 127.0f / absVal;
|
|
dequant_scale[i] = absVal / 127.0f;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNDynamicUpdateConvBiasScale(float* newbias, float* oldbias, float* weightKernelSum, float* inputBias,
|
|
size_t ocQuad) {
|
|
int ocUp4 = 4 * ocQuad;
|
|
int pack = 4;
|
|
for (int i = 0; i < ocUp4; ++i) {
|
|
newbias[i] = oldbias[i] + weightKernelSum[i] * inputBias[0];
|
|
}
|
|
}
|
|
|
|
#endif // LOW_MEMORY
|
|
#endif // not __aarch64__
|
|
|
|
static void MNNCountMaxMinValue(const float* source, float* minVal, float* maxVal, size_t size) {
|
|
#ifndef MNN_USE_NEON
|
|
int pack = 4;
|
|
float max_ = source[0], min_ = source[0];
|
|
for (int i = 1; i < size; ++i) {
|
|
if (max_ < source[i]) {
|
|
max_ = source[i];
|
|
}
|
|
if (min_ > source[i]) {
|
|
min_ = source[i];
|
|
}
|
|
}
|
|
*minVal = min_;
|
|
*maxVal = max_;
|
|
#else
|
|
auto sizeDiv4 = size / 4;
|
|
auto remain = size - 4 * sizeDiv4;
|
|
auto srcPtr = source;
|
|
auto max0 = vdupq_n_f32(srcPtr[0]);
|
|
auto min0 = vdupq_n_f32(srcPtr[0]);
|
|
while (sizeDiv4 > 15) {
|
|
sizeDiv4 -= 16;
|
|
auto data0 = vld1q_f32(srcPtr);
|
|
auto data1 = vld1q_f32(srcPtr + 4);
|
|
auto data2 = vld1q_f32(srcPtr + 8);
|
|
auto data3 = vld1q_f32(srcPtr + 12);
|
|
auto data4 = vld1q_f32(srcPtr + 16);
|
|
auto data5 = vld1q_f32(srcPtr + 20);
|
|
auto data6 = vld1q_f32(srcPtr + 24);
|
|
auto data7 = vld1q_f32(srcPtr + 28);
|
|
auto data8 = vld1q_f32(srcPtr + 32);
|
|
auto data9 = vld1q_f32(srcPtr + 36);
|
|
auto data10 = vld1q_f32(srcPtr + 40);
|
|
auto data11 = vld1q_f32(srcPtr + 44);
|
|
auto data12 = vld1q_f32(srcPtr + 48);
|
|
auto data13 = vld1q_f32(srcPtr + 52);
|
|
auto data14 = vld1q_f32(srcPtr + 56);
|
|
auto data15 = vld1q_f32(srcPtr + 60);
|
|
|
|
auto lmin0 = vminq_f32(data0, data1);
|
|
auto lmin2 = vminq_f32(data2, data3);
|
|
auto lmin4 = vminq_f32(data4, data5);
|
|
auto lmin6 = vminq_f32(data6, data7);
|
|
auto lmin8 = vminq_f32(data8, data9);
|
|
auto lmin10 = vminq_f32(data10, data11);
|
|
auto lmin12 = vminq_f32(data12, data13);
|
|
auto lmin14 = vminq_f32(data14, data15);
|
|
|
|
auto lmax0 = vmaxq_f32(data0, data1);
|
|
auto lmax2 = vmaxq_f32(data2, data3);
|
|
auto lmax4 = vmaxq_f32(data4, data5);
|
|
auto lmax6 = vmaxq_f32(data6, data7);
|
|
auto lmax8 = vmaxq_f32(data8, data9);
|
|
auto lmax10 = vmaxq_f32(data10, data11);
|
|
auto lmax12 = vmaxq_f32(data12, data13);
|
|
auto lmax14 = vmaxq_f32(data14, data15);
|
|
|
|
lmin0 = vminq_f32(lmin0, lmin2);
|
|
lmin4 = vminq_f32(lmin4, lmin6);
|
|
lmin8 = vminq_f32(lmin8, lmin10);
|
|
lmin12 = vminq_f32(lmin12, lmin14);
|
|
|
|
lmax0 = vmaxq_f32(lmax0, lmax2);
|
|
lmax4 = vmaxq_f32(lmax4, lmax6);
|
|
lmax8 = vmaxq_f32(lmax8, lmax10);
|
|
lmax12 = vmaxq_f32(lmax12, lmax14);
|
|
|
|
lmin0 = vminq_f32(lmin0, lmin8);
|
|
lmin4 = vminq_f32(lmin4, lmin12);
|
|
lmax0 = vmaxq_f32(lmax0, lmax8);
|
|
lmax4 = vmaxq_f32(lmax4, lmax12);
|
|
lmin0 = vminq_f32(lmin0, lmin4);
|
|
lmax0 = vmaxq_f32(lmax0, lmax4);
|
|
|
|
max0 = vmaxq_f32(max0, lmax0);
|
|
min0 = vminq_f32(min0, lmin0);
|
|
srcPtr += 64;
|
|
}
|
|
if (sizeDiv4 > 7) {
|
|
sizeDiv4 -= 8;
|
|
auto data0 = vld1q_f32(srcPtr);
|
|
auto data1 = vld1q_f32(srcPtr + 4);
|
|
auto data2 = vld1q_f32(srcPtr + 8);
|
|
auto data3 = vld1q_f32(srcPtr + 12);
|
|
auto data4 = vld1q_f32(srcPtr + 16);
|
|
auto data5 = vld1q_f32(srcPtr + 20);
|
|
auto data6 = vld1q_f32(srcPtr + 24);
|
|
auto data7 = vld1q_f32(srcPtr + 28);
|
|
|
|
auto lmin0 = vminq_f32(data0, data1);
|
|
auto lmin2 = vminq_f32(data2, data3);
|
|
auto lmin4 = vminq_f32(data4, data5);
|
|
auto lmin6 = vminq_f32(data6, data7);
|
|
|
|
auto lmax0 = vmaxq_f32(data0, data1);
|
|
auto lmax2 = vmaxq_f32(data2, data3);
|
|
auto lmax4 = vmaxq_f32(data4, data5);
|
|
auto lmax6 = vmaxq_f32(data6, data7);
|
|
|
|
lmin0 = vminq_f32(lmin0, lmin2);
|
|
lmin4 = vminq_f32(lmin4, lmin6);
|
|
|
|
lmax0 = vmaxq_f32(lmax0, lmax2);
|
|
lmax4 = vmaxq_f32(lmax4, lmax6);
|
|
|
|
lmin0 = vminq_f32(lmin0, lmin4);
|
|
lmax0 = vmaxq_f32(lmax0, lmax4);
|
|
|
|
max0 = vmaxq_f32(max0, lmax0);
|
|
min0 = vminq_f32(min0, lmin0);
|
|
srcPtr += 32;
|
|
}
|
|
if (sizeDiv4 > 3) {
|
|
sizeDiv4 -= 4;
|
|
auto data0 = vld1q_f32(srcPtr);
|
|
auto data1 = vld1q_f32(srcPtr + 4);
|
|
auto data2 = vld1q_f32(srcPtr + 8);
|
|
auto data3 = vld1q_f32(srcPtr + 12);
|
|
|
|
auto lmin0 = vminq_f32(data0, data1);
|
|
auto lmin2 = vminq_f32(data2, data3);
|
|
|
|
auto lmax0 = vmaxq_f32(data0, data1);
|
|
auto lmax2 = vmaxq_f32(data2, data3);
|
|
|
|
lmin0 = vminq_f32(lmin0, lmin2);
|
|
lmax0 = vmaxq_f32(lmax0, lmax2);
|
|
|
|
max0 = vmaxq_f32(max0, lmax0);
|
|
min0 = vminq_f32(min0, lmin0);
|
|
srcPtr += 16;
|
|
}
|
|
if (sizeDiv4 > 1) {
|
|
sizeDiv4 -= 2;
|
|
auto data0 = vld1q_f32(srcPtr);
|
|
auto data1 = vld1q_f32(srcPtr + 4);
|
|
|
|
auto lmin0 = vminq_f32(data0, data1);
|
|
auto lmax0 = vmaxq_f32(data0, data1);
|
|
|
|
max0 = vmaxq_f32(max0, lmax0);
|
|
min0 = vminq_f32(min0, lmin0);
|
|
srcPtr += 8;
|
|
}
|
|
if (sizeDiv4 > 0) {
|
|
sizeDiv4--;
|
|
auto data0 = vld1q_f32(srcPtr);
|
|
max0 = vmaxq_f32(max0, data0);
|
|
min0 = vminq_f32(min0, data0);
|
|
srcPtr += 4;
|
|
}
|
|
float temp0[4];
|
|
float temp1[4];
|
|
vst1q_f32(temp0, max0);
|
|
vst1q_f32(temp1, min0);
|
|
auto maxval = temp0[0];
|
|
auto minval = temp1[0];
|
|
for (int i = 1; i < 4; ++i) {
|
|
maxval = ALIMAX(maxval, temp0[i]);
|
|
minval = ALIMIN(minval, temp1[i]);
|
|
}
|
|
while (remain > 0) {
|
|
maxval = ALIMAX(maxval, srcPtr[0]);
|
|
minval = ALIMIN(minval, srcPtr[0]);
|
|
remain--;
|
|
srcPtr += 1;
|
|
}
|
|
minVal[0] = minval;
|
|
maxVal[0] = maxval;
|
|
#endif
|
|
}
|
|
|
|
#ifdef MNN_LOW_MEMORY
|
|
static void MNNAbsMaxFP32(const float* source, float* absmax, size_t src_depth_quad, size_t realSize, int pack) {
|
|
#ifdef __aarch64__
|
|
if (pack == 4) {
|
|
MNNAbsMaxFP32_Pack4(source, absmax, src_depth_quad, realSize, pack);
|
|
return;
|
|
}
|
|
if (pack == 8) {
|
|
MNNAbsMaxFP32_Pack8(source, absmax, src_depth_quad, realSize, pack);
|
|
return;
|
|
}
|
|
#endif
|
|
// source: (ic/4, N, 4)
|
|
auto srcStep = pack * realSize;
|
|
for (int i = 0; i < realSize; ++i) {
|
|
float absmaxVal = 0.f; // absmaxVal>=0
|
|
for (int c = 0; c < src_depth_quad; ++c) {
|
|
auto src = source + c * srcStep + i * pack;
|
|
for (int k = 0; k < pack; ++k) {
|
|
absmaxVal = std::max(absmaxVal, std::abs(src[k]));
|
|
}
|
|
}
|
|
absmax[i] = absmaxVal;
|
|
}
|
|
}
|
|
|
|
void MNNDynamicQuantFP32(const float* src, int8_t* dst, const float* scale, size_t src_depth_quad, size_t realSize,
|
|
int pack, const float* bias = nullptr) {
|
|
#ifdef __aarch64__
|
|
if (pack == 4) {
|
|
MNNDynamicQuantFP32_Pack4(src, dst, scale, src_depth_quad, realSize, nullptr, pack);
|
|
return;
|
|
}
|
|
if (pack == 8) {
|
|
MNNDynamicQuantFP32_Pack8(src, dst, scale, src_depth_quad, realSize, nullptr, pack);
|
|
return;
|
|
}
|
|
#endif
|
|
#ifdef MNN_USE_SSE
|
|
uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
|
|
int offset = 128;
|
|
#else
|
|
int8_t* dstPtr = dst;
|
|
int offset = 0;
|
|
#endif
|
|
for (int i = 0; i < realSize; ++i) {
|
|
auto scaleVal = scale[i];
|
|
for (int c = 0; c < src_depth_quad; ++c) {
|
|
auto srcZ = src + c * pack * realSize + i * pack;
|
|
auto dstZ = dstPtr + c * pack * realSize + i * pack;
|
|
for (int k = 0; k < pack; ++k) {
|
|
int val = (int)roundf(srcZ[k] * scaleVal);
|
|
dstZ[k] = val + offset;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
static void MNNAsyQuantFunc(int8_t* dst, const float* src, float* qscale, float* qbias, const size_t* info) {
|
|
// input shape: [kernelsize, blockNum, blockLU, EP, LP]
|
|
auto blockNum = info[0];
|
|
auto EP = info[1]; // real area for data
|
|
auto LP = info[2]; // Innermost data layout, may come from backend's pack or gemmint8 units' SRC_UNIT
|
|
auto DST_XUNIT = info[3]; // backend gemmint8 units
|
|
auto SRC_UNIT = info[4];
|
|
auto kernelsize = info[5];
|
|
auto blockLU = info[6];
|
|
auto stride0 = blockNum * blockLU * EP * LP;
|
|
auto stride1 = blockLU * EP * LP;
|
|
int int8Max = 127;
|
|
int int8Min = -128;
|
|
// qscale&qbias [blockNum, EP]
|
|
#ifdef __aarch64__
|
|
if (LP == 4 || LP == 8) {
|
|
for (int k = 0; k < kernelsize; ++k) {
|
|
for (int i = 0; i < blockNum; ++i) {
|
|
if (LP == 4) {
|
|
MNNDynamicQuantFP32_Pack4(src + k * stride0 + i * stride1, dst + k * stride0 + i * stride1,
|
|
qscale + i * EP, blockLU, EP, qbias + i * EP, LP);
|
|
}
|
|
if (LP == 8) {
|
|
MNNDynamicQuantFP32_Pack8(src + k * stride0 + i * stride1, dst + k * stride0 + i * stride1,
|
|
qscale + i * EP, blockLU, EP, qbias + i * EP, LP);
|
|
}
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
#endif
|
|
for (int i = 0; i < EP; ++i) {
|
|
for (int bk = 0; bk < blockNum; ++bk) {
|
|
float quant_scale = qscale[i + bk * EP];
|
|
float quant_bias = qbias[i + bk * EP];
|
|
for (int n = 0; n < kernelsize; ++n) {
|
|
for (int k = 0; k < blockLU; ++k) {
|
|
for (int j = 0; j < LP; ++j) {
|
|
int dataIndx = n * stride0 + bk * stride1 + k * EP * LP + i * LP + j;
|
|
float data_ = src[dataIndx];
|
|
int qval = static_cast<int32_t>(roundf(data_ * quant_scale + quant_bias));
|
|
#ifdef MNN_USE_SSE
|
|
((uint8_t*)dst)[dataIndx] = qval + 128;
|
|
#else
|
|
dst[dataIndx] = ALIMIN(int8Max, ALIMAX(int8Min, qval));
|
|
#endif
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
static void MNNAsyQuantInfo_FP32(float* scale, float* bias, float* qscale, float* qbias, float* dstMin, float* dstMax,
|
|
const float* src, const size_t* info) {
|
|
auto blockNum = info[0];
|
|
auto plane = info[1]; // real area for data
|
|
auto innerSide = info[2]; // Innermost data layout, may come from backend's pack or gemmint8 units' SRC_UNIT
|
|
auto DST_XUNIT = info[3];
|
|
auto kernelsize = info[5];
|
|
auto blockLU = info[6];
|
|
auto stride0 = blockNum * blockLU * plane * innerSide;
|
|
auto stride1 = blockLU * plane * innerSide;
|
|
|
|
if (info[7] == 1) { // scale&bias:[1]
|
|
float maxval, minval;
|
|
MNNCountMaxMinValue(src, &minval, &maxval, kernelsize * stride0);
|
|
if (info[8] == 1 && (maxval - minval) > 1e-7) {
|
|
if (minval > 0.f) {
|
|
minval = 0;
|
|
} else if (maxval < 0.f) {
|
|
maxval = 0;
|
|
}
|
|
}
|
|
auto range = maxval - minval;
|
|
if (range <= 1e-7) {
|
|
scale[0] = 1.f;
|
|
qscale[0] = 1.f;
|
|
qbias[0] = -maxval;
|
|
bias[0] = maxval;
|
|
} else {
|
|
qscale[0] = 255.f / range;
|
|
scale[0] = range / 255.f;
|
|
qbias[0] = -minval * 255.f / range - 128.f;
|
|
bias[0] = minval + 128.f * range / 255.f;
|
|
}
|
|
return;
|
|
}
|
|
|
|
// input : [kernelsize, blockNum, blockLU, plane, pack]
|
|
// dequant scale/bias : [EU, blockNum, step], step=ALIMIN(step, EP), EU=UP_DIV(plane, EP)
|
|
// quant scale/bias : [blockNum, plane]
|
|
#ifdef __aarch64__
|
|
if ((DST_XUNIT == 12 || DST_XUNIT == 16) && innerSide == 4) { // Arm82,fp32: SRC_UNIT=4, core->pack=4
|
|
// max,min shape: [blockNum, EP]
|
|
for (int i = 0; i < kernelsize; ++i) {
|
|
MNNLocalMinMaxFP32_Pack4(dstMin, dstMax, src + i * stride0, blockNum, blockLU, plane, innerSide, i);
|
|
}
|
|
// scale, bias
|
|
if (DST_XUNIT == 12) {
|
|
bool success = MNNAsyLocalQuantInfo_EP12_FP32(scale, bias, qscale, qbias, dstMin, dstMax, info);
|
|
if (!success) {
|
|
MNN_ERROR("Call error for:MNNAsyLocalQuantInfo_EP12\n");
|
|
return;
|
|
}
|
|
return;
|
|
}
|
|
if (DST_XUNIT == 16) {
|
|
bool success = MNNAsyLocalQuantInfo_EP16_FP32(scale, bias, qscale, qbias, dstMin, dstMax, info);
|
|
if (!success) {
|
|
MNN_ERROR("Call error for:MNNAsyLocalQuantInfo_EP16_FP32\n");
|
|
return;
|
|
}
|
|
return;
|
|
}
|
|
}
|
|
if (DST_XUNIT == 10) { // Arm86,fp32: SRC_UNIT=8,core->pack=4
|
|
// max,min shape: [blockNum, EP]
|
|
if (innerSide == 4) {
|
|
for (int i = 0; i < kernelsize; ++i) {
|
|
MNNLocalMinMaxFP32_Pack4(dstMin, dstMax, src + i * stride0, blockNum, blockLU, plane, innerSide, i);
|
|
}
|
|
}
|
|
if (innerSide == 8) {
|
|
for (int i = 0; i < kernelsize; ++i) {
|
|
MNNLocalMinMaxFP32_Pack8(dstMin, dstMax, src + i * stride0, blockNum, blockLU, plane, innerSide, i);
|
|
}
|
|
}
|
|
// scale, bias
|
|
bool success = MNNAsyLocalQuantInfo_EP10_FP32(scale, bias, qscale, qbias, dstMin, dstMax, info);
|
|
if (!success) {
|
|
MNN_ERROR("Call error for:MNNAsyLocalQuantInfo_EP10\n");
|
|
return;
|
|
}
|
|
return;
|
|
}
|
|
#endif
|
|
// max,min shape: [blockNum, plane]
|
|
for (int i = 0; i < plane; ++i) {
|
|
for (int bk = 0; bk < blockNum; ++bk) {
|
|
auto idx0 = i * innerSide + bk * stride1;
|
|
float max_ = src[idx0];
|
|
float min_ = max_;
|
|
for (int n = 0; n < kernelsize; ++n) {
|
|
for (int k = 0; k < blockLU; ++k) {
|
|
for (int j = 0; j < innerSide; ++j) {
|
|
auto dataIndx = idx0 + n * stride0 + k * (plane * innerSide) + j;
|
|
float data_ = src[dataIndx];
|
|
max_ = ALIMAX(max_, data_);
|
|
min_ = ALIMIN(min_, data_);
|
|
}
|
|
}
|
|
}
|
|
auto sindx = i + bk * plane;
|
|
dstMin[sindx] = min_;
|
|
dstMax[sindx] = max_;
|
|
}
|
|
}
|
|
// scale, bias
|
|
for (int i = 0; i < plane; ++i) {
|
|
auto step = ALIMIN(DST_XUNIT, plane - (i / DST_XUNIT) * DST_XUNIT);
|
|
auto sind0 = (i / DST_XUNIT) * DST_XUNIT * blockNum + (i % DST_XUNIT);
|
|
for (int k = 0; k < blockNum; ++k) {
|
|
auto sind = sind0 + k * step;
|
|
auto qind = i + k * plane;
|
|
auto max_ = dstMax[qind];
|
|
auto min_ = dstMin[qind];
|
|
if (fabs(max_ - min_) < 1e-7) {
|
|
qscale[qind] = 0.f;
|
|
qbias[qind] = 0.f;
|
|
scale[sind] = 0.f;
|
|
bias[sind] = max_;
|
|
} else {
|
|
qscale[qind] = 255.f / (max_ - min_);
|
|
qbias[qind] = roundf(-min_ * 255.f / (max_ - min_)) - 128.0f;
|
|
scale[sind] = (max_ - min_) / 255.f;
|
|
bias[sind] = min_ + (128.f / 255.f) * (max_ - min_);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
#endif // MNN_LOW_MEMORY
|
|
|
|
static void MNNReorderWeightInt4(uint8_t* dest, const uint8_t* source, int32_t* shape, size_t size, float* kernelsum) {
|
|
MNN_ASSERT(size > 4);
|
|
auto blocknum = shape[0];
|
|
auto hu = shape[1];
|
|
auto lu = shape[2];
|
|
auto hp = shape[3];
|
|
auto lp = shape[4];
|
|
auto ic = blocknum * lu * lp;
|
|
auto stride0 = blocknum * hp * lu * lp;
|
|
auto stride1 = lu * hp * lp;
|
|
auto stride2 = hp * lp;
|
|
// [oc,ic]->[hu,blocknum,lu,hp,lp]
|
|
for (int i = 0; i < hu; ++i) {
|
|
for (int k = 0; k < hp; ++k) {
|
|
for (int bl = 0; bl < blocknum; ++bl) {
|
|
for (int j = 0; j < lu; ++j) {
|
|
int srcindex = (i * hp + k) * ic + bl * (lu * lp) + j * lp;
|
|
int dstindex = i * stride0 + bl * stride1 + j * stride2 + k * lp;
|
|
memcpy(dest + dstindex, source + srcindex, lp);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
// [hu,blocknum,lu,hp,lp] address [hp,lp] for int4
|
|
auto inside = lp * hp;
|
|
auto outside = blocknum * hu;
|
|
std::vector<uint8_t> buffer(inside);
|
|
for (int i = 0; i < outside; ++i) {
|
|
std::vector<float> accum(hp, 0);
|
|
for (int k = 0; k < lu; ++k) {
|
|
for (int j = 0; j < inside / 2; ++j) {
|
|
auto w0 = dest[j + (i * lu + k) * inside] >> 4;
|
|
auto w1 = dest[j + (i * lu + k) * inside] & 0x0f;
|
|
auto w2 = dest[(i * lu + k) * inside + j + inside / 2] >> 4;
|
|
auto w3 = dest[(i * lu + k) * inside + j + inside / 2] & 0x0f;
|
|
buffer[2 * j + 0] = w0 * 16 + w2;
|
|
buffer[2 * j + 1] = w1 * 16 + w3;
|
|
// sum
|
|
accum[j / lp] += ((float)w0 + (float)w1);
|
|
accum[(j + inside / 2) / lp] += ((float)w2 + (float)w3);
|
|
}
|
|
memcpy(dest + (i * lu + k) * inside, buffer.data(), inside);
|
|
}
|
|
memcpy(kernelsum + i * hp, accum.data(), hp * sizeof(float));
|
|
}
|
|
}
|
|
#ifdef __aarch64__
|
|
static void MNNReorderWeightInt4Arm86(uint8_t* dest, const uint8_t* source, int32_t* shape, size_t size,
|
|
float* kernelsum) {
|
|
MNN_ASSERT(size > 4);
|
|
auto blocknum = shape[0];
|
|
auto hu = shape[1];
|
|
auto lu = shape[2];
|
|
auto hp = shape[3];
|
|
auto lp = shape[4];
|
|
auto ic = blocknum * lu * lp;
|
|
auto stride0 = blocknum * hp * lu * lp;
|
|
auto stride1 = lu * hp * lp;
|
|
auto stride2 = hp * lp;
|
|
auto dstPtr = (int32_t*)dest;
|
|
auto srcPtr = (int32_t*)source;
|
|
int unitpacksize = sizeof(int32_t) / sizeof(uint8_t);
|
|
|
|
for (int i = 0; i < hu; ++i) {
|
|
for (int k = 0; k < hp; ++k) {
|
|
for (int bl = 0; bl < blocknum; ++bl) {
|
|
int j = 0;
|
|
while (j + 7 < lu) {
|
|
auto srcindex0 = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto srcindex1 = ((i * hp + k) * ic + bl * (lu * lp) + (j + 4) * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex4 = (bl * stride1 + i * stride0 + (j + 4) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex5 = (bl * stride1 + i * stride0 + (j + 5) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex6 = (bl * stride1 + i * stride0 + (j + 6) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex7 = (bl * stride1 + i * stride0 + (j + 7) * stride2 + k * lp) / unitpacksize;
|
|
j += 8;
|
|
auto srcdata0 = vld1q_s32(srcPtr + srcindex0);
|
|
auto srcdata1 = vld1q_s32(srcPtr + srcindex1);
|
|
vst1q_lane_s32(dstPtr + dstindex0, srcdata0, 0);
|
|
vst1q_lane_s32(dstPtr + dstindex1, srcdata0, 1);
|
|
vst1q_lane_s32(dstPtr + dstindex2, srcdata0, 2);
|
|
vst1q_lane_s32(dstPtr + dstindex3, srcdata0, 3);
|
|
vst1q_lane_s32(dstPtr + dstindex4, srcdata1, 0);
|
|
vst1q_lane_s32(dstPtr + dstindex5, srcdata1, 1);
|
|
vst1q_lane_s32(dstPtr + dstindex6, srcdata1, 2);
|
|
vst1q_lane_s32(dstPtr + dstindex7, srcdata1, 3);
|
|
}
|
|
while (j + 3 < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
j += 4;
|
|
auto srcdata = vld1q_s32(srcPtr + srcindex);
|
|
vst1q_lane_s32(dstPtr + dstindex0, srcdata, 0);
|
|
vst1q_lane_s32(dstPtr + dstindex1, srcdata, 1);
|
|
vst1q_lane_s32(dstPtr + dstindex2, srcdata, 2);
|
|
vst1q_lane_s32(dstPtr + dstindex3, srcdata, 3);
|
|
}
|
|
while (j < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
dstPtr[dstindex] = srcPtr[srcindex];
|
|
j++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
MNNPermuteSumWeightInt4Arm86(dest, dest, blocknum * hu, lu, kernelsum);
|
|
}
|
|
|
|
static void MNNReorderWeightInt4Arm82(uint8_t* dest, const uint8_t* source, int32_t* shape, size_t size,
|
|
float* kernelsum) {
|
|
MNN_ASSERT(size > 4);
|
|
// dst shape: [hu, blocknum, kernelCount, lu, hp, lp], kernelCount=1 in this case
|
|
auto blocknum = shape[0];
|
|
auto hu = shape[1];
|
|
auto lu = shape[2];
|
|
auto hp = shape[3];
|
|
auto lp = shape[4];
|
|
auto ic = blocknum * lu * lp;
|
|
auto stride0 = blocknum * hp * lu * lp;
|
|
auto stride1 = lu * hp * lp;
|
|
auto stride2 = hp * lp;
|
|
auto dstPtr = (int16_t*)dest;
|
|
auto srcPtr = (int16_t*)source;
|
|
int unitpacksize = sizeof(int16_t) / sizeof(uint8_t);
|
|
for (int i = 0; i < hu; ++i) {
|
|
for (int k = 0; k < hp; ++k) {
|
|
for (int bl = 0; bl < blocknum; ++bl) {
|
|
int j = 0;
|
|
while (j + 7 < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex4 = (bl * stride1 + i * stride0 + (j + 4) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex5 = (bl * stride1 + i * stride0 + (j + 5) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex6 = (bl * stride1 + i * stride0 + (j + 6) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex7 = (bl * stride1 + i * stride0 + (j + 7) * stride2 + k * lp) / unitpacksize;
|
|
j += 8;
|
|
auto srcdata = vld1q_s16(srcPtr + srcindex);
|
|
vst1q_lane_s16(dstPtr + dstindex0, srcdata, 0);
|
|
vst1q_lane_s16(dstPtr + dstindex1, srcdata, 1);
|
|
vst1q_lane_s16(dstPtr + dstindex2, srcdata, 2);
|
|
vst1q_lane_s16(dstPtr + dstindex3, srcdata, 3);
|
|
vst1q_lane_s16(dstPtr + dstindex4, srcdata, 4);
|
|
vst1q_lane_s16(dstPtr + dstindex5, srcdata, 5);
|
|
vst1q_lane_s16(dstPtr + dstindex6, srcdata, 6);
|
|
vst1q_lane_s16(dstPtr + dstindex7, srcdata, 7);
|
|
}
|
|
while (j + 3 < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
j += 4;
|
|
auto srcdata = vld1_s16(srcPtr + srcindex);
|
|
vst1_lane_s16(dstPtr + dstindex0, srcdata, 0);
|
|
vst1_lane_s16(dstPtr + dstindex1, srcdata, 1);
|
|
vst1_lane_s16(dstPtr + dstindex2, srcdata, 2);
|
|
vst1_lane_s16(dstPtr + dstindex3, srcdata, 3);
|
|
}
|
|
while (j < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / 2;
|
|
auto dstindex = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / 2;
|
|
dstPtr[dstindex] = srcPtr[srcindex];
|
|
j++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
MNNPermuteSumWeightInt4Arm82(dest, dest, blocknum * hu, lu, kernelsum);
|
|
}
|
|
#ifdef MNN_SME2
|
|
static void MNNReorderWeightInt4Sme2(uint8_t* dest, const uint8_t* source, int32_t* shape, size_t size,
|
|
float* kernelsum) {
|
|
MNN_ASSERT(size > 4);
|
|
// dst shape: [hu, blocknum, kernelCount, lu, hp, lp], kernelCount=1 in this case
|
|
auto blocknum = shape[0];
|
|
auto hu = shape[1];
|
|
auto lu = shape[2];
|
|
auto hp = shape[3];
|
|
auto lp = shape[4];
|
|
auto ic = blocknum * lu * lp;
|
|
auto stride0 = blocknum * hp * lu * lp;
|
|
auto stride1 = lu * hp * lp;
|
|
auto stride2 = hp * lp;
|
|
auto dstPtr = (int16_t*)dest;
|
|
auto srcPtr = (int16_t*)source;
|
|
int unitpacksize = sizeof(int16_t) / sizeof(uint8_t);
|
|
for (int i = 0; i < hu; ++i) {
|
|
for (int k = 0; k < hp; ++k) {
|
|
for (int bl = 0; bl < blocknum; ++bl) {
|
|
int j = 0;
|
|
while (j + 7 < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex4 = (bl * stride1 + i * stride0 + (j + 4) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex5 = (bl * stride1 + i * stride0 + (j + 5) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex6 = (bl * stride1 + i * stride0 + (j + 6) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex7 = (bl * stride1 + i * stride0 + (j + 7) * stride2 + k * lp) / unitpacksize;
|
|
j += 8;
|
|
auto srcdata = vld1q_s16(srcPtr + srcindex);
|
|
vst1q_lane_s16(dstPtr + dstindex0, srcdata, 0);
|
|
vst1q_lane_s16(dstPtr + dstindex1, srcdata, 1);
|
|
vst1q_lane_s16(dstPtr + dstindex2, srcdata, 2);
|
|
vst1q_lane_s16(dstPtr + dstindex3, srcdata, 3);
|
|
vst1q_lane_s16(dstPtr + dstindex4, srcdata, 4);
|
|
vst1q_lane_s16(dstPtr + dstindex5, srcdata, 5);
|
|
vst1q_lane_s16(dstPtr + dstindex6, srcdata, 6);
|
|
vst1q_lane_s16(dstPtr + dstindex7, srcdata, 7);
|
|
}
|
|
while (j + 3 < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / unitpacksize;
|
|
auto dstindex0 = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex1 = (bl * stride1 + i * stride0 + (j + 1) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex2 = (bl * stride1 + i * stride0 + (j + 2) * stride2 + k * lp) / unitpacksize;
|
|
auto dstindex3 = (bl * stride1 + i * stride0 + (j + 3) * stride2 + k * lp) / unitpacksize;
|
|
j += 4;
|
|
auto srcdata = vld1_s16(srcPtr + srcindex);
|
|
vst1_lane_s16(dstPtr + dstindex0, srcdata, 0);
|
|
vst1_lane_s16(dstPtr + dstindex1, srcdata, 1);
|
|
vst1_lane_s16(dstPtr + dstindex2, srcdata, 2);
|
|
vst1_lane_s16(dstPtr + dstindex3, srcdata, 3);
|
|
}
|
|
while (j < lu) {
|
|
auto srcindex = ((i * hp + k) * ic + bl * (lu * lp) + j * lp) / 2;
|
|
auto dstindex = (bl * stride1 + i * stride0 + j * stride2 + k * lp) / 2;
|
|
dstPtr[dstindex] = srcPtr[srcindex];
|
|
j++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
int32_t table[16] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
|
|
if (hp == 32) {
|
|
MNNPermuteSumWeightInt4Sme2_Hp32(dest, dest, blocknum * hu, lu, kernelsum, table);
|
|
} else if (hp == 128) { // [hu,blocknum,lu,hp,lp]
|
|
MNNPermuteSumWeightInt4Sme2_Hp128(dest, dest, blocknum * hu, lu, kernelsum, table);
|
|
} else {
|
|
for (int i = 0; i < blocknum * hu; ++i) {
|
|
std::vector<float> sum(hp, 0);
|
|
for (int j = 0; j < lu; ++j) {
|
|
auto destPtr = dest + i * lu * lp * hp + j * lp * hp;
|
|
for (int k = 0; k < hp; ++k) {
|
|
for (int x = 0; x < lp; ++x) {
|
|
uint8_t data = destPtr[k * lp + x];
|
|
auto d0 = data / 16;
|
|
auto d1 = data % 16;
|
|
sum[k] = sum[k] + float(d0 + d1);
|
|
destPtr[k * lp + x] = d0 + d1 * 16;
|
|
}
|
|
}
|
|
}
|
|
memcpy(kernelsum + i * hp, sum.data(), hp * sizeof(float));
|
|
}
|
|
}
|
|
}
|
|
#endif // sme2
|
|
#endif // __aarch64__
|
|
|
|
static void MNNSumWeightInt8(float* kernelsum, int8_t* source, size_t outside, size_t reduceAxis, size_t hP,
|
|
size_t lP) {
|
|
// weight shape: [outside, axis, hP, lP]
|
|
// outside = blocknum * hU
|
|
// reduceAxis = kernelCount * lU
|
|
auto inside = hP * lP;
|
|
auto stride0 = inside * reduceAxis;
|
|
std::vector<float> accum(hP);
|
|
for (int i = 0; i < outside; ++i) {
|
|
memset(accum.data(), 0, hP * 4);
|
|
for (int j = 0; j < reduceAxis; ++j) {
|
|
for (int k = 0; k < hP; ++k) {
|
|
for (int x = 0; x < lP; ++x) {
|
|
accum[k] += (float)source[x + k * lP + j * inside + i * stride0];
|
|
}
|
|
}
|
|
}
|
|
memcpy(kernelsum + i * hP, accum.data(), hP * sizeof(float));
|
|
}
|
|
}
|
|
|
|
static void MNNSumByAxisLForMatmul_A(float* dest, int8_t* source, const float* scale, ssize_t realDstCount,
|
|
SumByAxisParams sumParams) {
|
|
#ifdef MNN_USE_SSE
|
|
uint8_t* srcInt8 = reinterpret_cast<uint8_t*>(source);
|
|
#else
|
|
int8_t* srcInt8 = source;
|
|
#endif
|
|
auto scalePtr = scale;
|
|
auto blockNum = sumParams.blockNum;
|
|
auto EP = sumParams.DST_XUNIT;
|
|
auto LP = sumParams.SRC_UNIT;
|
|
auto col_buffer_unit_size = sumParams.unitColBufferSize;
|
|
auto oneScale = sumParams.oneScale;
|
|
auto LU = sumParams.LU;
|
|
auto valid = sumParams.valid;
|
|
auto kernelxy = sumParams.kernelxy;
|
|
auto blockSizeQuad = LU / blockNum;
|
|
auto inputBlockQuant = sumParams.inputBlock;
|
|
auto lastL = LP;
|
|
if (valid) {
|
|
lastL = valid;
|
|
}
|
|
float singlescale = scale[0];
|
|
do {
|
|
int step = ALIMIN(EP, realDstCount);
|
|
int scaleOffset = inputBlockQuant ? (step * blockNum) : step;
|
|
|
|
for (int k = 0; k < blockNum; ++k) {
|
|
const auto src_x = srcInt8 + k * (step * LP * blockSizeQuad * kernelxy);
|
|
for (int w = 0; w < step; ++w) {
|
|
float dequantScale = singlescale;
|
|
if (oneScale == 0 && inputBlockQuant) {
|
|
dequantScale = scalePtr[w + k * step];
|
|
} else if (oneScale == 0) {
|
|
dequantScale = scalePtr[w];
|
|
}
|
|
int sumint32 = 0;
|
|
const auto src_y = src_x + w * LP;
|
|
for (int j = 0; j < kernelxy; ++j) {
|
|
for (int i = 0; i < blockSizeQuad; ++i) {
|
|
auto sumsize = i == (blockSizeQuad - 1) ? lastL : LP;
|
|
const auto src_z = src_y + j * (blockSizeQuad * step * LP) + i * step * LP;
|
|
for (int x = 0; x < sumsize; ++x) {
|
|
sumint32 += src_z[x];
|
|
}
|
|
}
|
|
}
|
|
dest[w + k * step] = dequantScale * static_cast<float>(sumint32);
|
|
}
|
|
}
|
|
scalePtr += scaleOffset;
|
|
|
|
dest += (step * blockNum);
|
|
realDstCount -= step;
|
|
srcInt8 += col_buffer_unit_size;
|
|
} while (realDstCount > 0);
|
|
}
|
|
|
|
template <typename T>
|
|
void MNNPackC4Common(T* dst, const T* src, size_t area, size_t depth, int* areaOffset) {
|
|
int depthC4 = depth / 4;
|
|
int depthRemain = depthC4 * 4;
|
|
int remain = depth - depthRemain;
|
|
int z, x, y;
|
|
const T* srcChannel[4];
|
|
const T* srcOffset = src;
|
|
for (z = 0; z < depthC4; ++z) {
|
|
auto dstZ = dst + z * areaOffset[1] * 4;
|
|
for (y = 0; y < 4; ++y) {
|
|
srcChannel[y] = srcOffset + areaOffset[0] * y;
|
|
}
|
|
for (x = 0; x < area; ++x) {
|
|
for (y = 0; y < 4; ++y) {
|
|
dstZ[0] = srcChannel[y][x];
|
|
dstZ++;
|
|
}
|
|
}
|
|
srcOffset += areaOffset[0] * 4;
|
|
}
|
|
if (remain > 0) {
|
|
auto dstZ = dst + depthC4 * areaOffset[1] * 4;
|
|
for (y = 0; y < remain; ++y) {
|
|
srcChannel[y] = srcOffset + areaOffset[0] * y;
|
|
}
|
|
for (x = 0; x < area; ++x) {
|
|
for (y = 0; y < remain; ++y) {
|
|
dstZ[0] = srcChannel[y][x];
|
|
dstZ++;
|
|
}
|
|
for (y = remain; y < 4; ++y) {
|
|
dstZ[0] = 0;
|
|
dstZ++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void MNNUnpackC4Common(T* dst, const T* src, size_t area, size_t depth, int* areaOffset) {
|
|
int depthC4 = depth / 4;
|
|
int depthRemain = depthC4 * 4;
|
|
int remain = depth - depthRemain;
|
|
int z, x, y;
|
|
const T* srcChannel[4];
|
|
const T* srcOffset = src;
|
|
for (z = 0; z < depthC4; ++z) {
|
|
for (y = 0; y < 4; ++y) {
|
|
auto dstZ = dst + (z * 4 + y) * areaOffset[1];
|
|
srcChannel[y] = srcOffset + y;
|
|
for (x = 0; x < area; ++x) {
|
|
dstZ[x] = srcChannel[y][0];
|
|
srcChannel[y] += 4;
|
|
}
|
|
}
|
|
srcOffset += areaOffset[0] * 4;
|
|
}
|
|
if (remain > 0) {
|
|
auto dstZ = dst + depthC4 * areaOffset[1] * 4;
|
|
for (y = 0; y < remain; ++y) {
|
|
srcChannel[y] = srcOffset + y;
|
|
for (x = 0; x < area; ++x) {
|
|
dstZ[x] = srcChannel[y][0];
|
|
srcChannel[y] += 4;
|
|
}
|
|
dstZ += areaOffset[1];
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void MNNPackC2Common(T* dst, const T* src, size_t area, size_t depth, int* areaOffset) {
|
|
int depthC2 = depth / 2;
|
|
int depthRemain = depthC2 * 2;
|
|
int remain = depth - depthRemain;
|
|
int z, x, y;
|
|
const T* srcChannel[2];
|
|
const T* srcOffset = src;
|
|
for (z = 0; z < depthC2; ++z) {
|
|
auto dstZ = dst + z * areaOffset[1] * 2;
|
|
for (y = 0; y < 2; ++y) {
|
|
srcChannel[y] = srcOffset + areaOffset[0] * y;
|
|
}
|
|
for (x = 0; x < area; ++x) {
|
|
for (y = 0; y < 2; ++y) {
|
|
dstZ[0] = srcChannel[y][x];
|
|
dstZ++;
|
|
}
|
|
}
|
|
srcOffset += areaOffset[0] * 2;
|
|
}
|
|
if (remain > 0) {
|
|
auto dstZ = dst + depthC2 * areaOffset[1] * 2;
|
|
for (y = 0; y < remain; ++y) {
|
|
srcChannel[y] = srcOffset + areaOffset[0] * y;
|
|
}
|
|
for (x = 0; x < area; ++x) {
|
|
for (y = 0; y < remain; ++y) {
|
|
dstZ[0] = srcChannel[y][x];
|
|
dstZ++;
|
|
}
|
|
for (y = remain; y < 2; ++y) {
|
|
dstZ[0] = 0;
|
|
dstZ++;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T>
|
|
void MNNUnpackC2Common(T* dst, const T* src, size_t area, size_t depth, int* areaOffset, int pack = 1) {
|
|
int depthC2 = depth / 2;
|
|
int depthRemain = depthC2 * 2;
|
|
int remain = depth - depthRemain;
|
|
int z, x, y;
|
|
const T* srcChannel[2];
|
|
const T* srcOffset = src;
|
|
for (z = 0; z < depthC2; ++z) {
|
|
for (y = 0; y < 2; ++y) {
|
|
auto dstZ = dst + (z * 2 + y) * areaOffset[1] * pack;
|
|
srcChannel[y] = srcOffset + y * pack;
|
|
for (x = 0; x < area; ++x) {
|
|
for (int p = 0; p < pack; ++p) {
|
|
dstZ[x * pack + p] = srcChannel[y][p];
|
|
}
|
|
srcChannel[y] += (2 * pack);
|
|
}
|
|
}
|
|
srcOffset += areaOffset[0] * 2 * pack;
|
|
}
|
|
if (remain > 0) {
|
|
auto dstZ = dst + depthC2 * areaOffset[1] * 2 * pack;
|
|
for (y = 0; y < remain; ++y) {
|
|
srcChannel[y] = srcOffset + y * pack;
|
|
for (x = 0; x < area; ++x) {
|
|
for (int p = 0; p < pack; ++p) {
|
|
dstZ[x * pack + p] = srcChannel[y][p];
|
|
}
|
|
srcChannel[y] += 2 * pack;
|
|
}
|
|
dstZ += areaOffset[1] * pack;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNN4BitcopyWithStride(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
auto src = (uint32_t*)srcO;
|
|
auto dst = (uint32_t*)dstO;
|
|
for (int i = 0; i < size; ++i) {
|
|
dst[0] = *src;
|
|
dst += ds;
|
|
src += stride;
|
|
}
|
|
}
|
|
|
|
void MNN4BitcopyFast(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
// ds=1, stride=0||1
|
|
auto src = (float*)srcO;
|
|
auto dst = (float*)dstO;
|
|
int cnt = size;
|
|
if (stride == 1) { // stride=1
|
|
#ifdef MNN_USE_NEON
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
auto v4 = vld1q_f32(src);
|
|
auto u4 = vld1q_f32(src + 4);
|
|
vst1q_f32(dst, v4);
|
|
vst1q_f32(dst + 4, u4);
|
|
dst += 8;
|
|
src += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
auto v4 = vld1q_f32(src);
|
|
vst1q_f32(dst, v4);
|
|
dst += 4;
|
|
src += 4;
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
__m128 v4 = _mm_loadu_ps(src);
|
|
__m128 u4 = _mm_loadu_ps(src + 4);
|
|
_mm_storeu_ps(dst, v4);
|
|
_mm_storeu_ps(dst + 4, u4);
|
|
dst += 8;
|
|
src += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
__m128 v4 = _mm_loadu_ps(src);
|
|
_mm_storeu_ps(dst, v4);
|
|
dst += 4;
|
|
src += 4;
|
|
}
|
|
#endif
|
|
} else { // stride=0
|
|
int i = 0;
|
|
float val = *src;
|
|
#ifdef MNN_USE_NEON
|
|
auto val4 = vdupq_n_f32(val);
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
vst1q_f32(dst, val4);
|
|
vst1q_f32(dst + 4, val4);
|
|
dst += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
vst1q_f32(dst, val4);
|
|
dst += 4;
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
__m128 val4 = _mm_set_ps(val, val, val, val);
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
_mm_storeu_ps(dst, val4);
|
|
_mm_storeu_ps((dst + 4), val4);
|
|
dst += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
_mm_storeu_ps(dst, val4);
|
|
dst += 4;
|
|
}
|
|
#endif
|
|
}
|
|
for (; cnt > 0; --cnt) {
|
|
dst[0] = *src;
|
|
dst += ds;
|
|
src += stride;
|
|
}
|
|
}
|
|
|
|
void MNN2BitcopyWithStride(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
auto src = (uint16_t*)srcO;
|
|
auto dst = (uint16_t*)dstO;
|
|
for (int i = 0; i < size; ++i) {
|
|
*dst = *src;
|
|
src += stride;
|
|
dst += ds;
|
|
}
|
|
}
|
|
|
|
void MNN2BitcopyFast(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
auto src = (uint16_t*)srcO;
|
|
auto dst = (uint16_t*)dstO;
|
|
int cnt = size;
|
|
uint16_t val = *src;
|
|
if (stride == 1) {
|
|
#ifdef MNN_USE_NEON
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
auto val8 = vld1q_u16(src);
|
|
vst1q_u16(dst, val8);
|
|
src += 8;
|
|
dst += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
auto val4 = vld1_u16(src);
|
|
vst1_u16(dst, val4);
|
|
src += 4;
|
|
dst += 4;
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
auto tmp = _mm_loadu_ps((float*)src);
|
|
_mm_storeu_ps((float*)dst, tmp);
|
|
src += 8;
|
|
dst += 8;
|
|
}
|
|
#endif
|
|
} else { // stride=0
|
|
#ifdef MNN_USE_NEON
|
|
auto val4 = vdup_n_u16(val);
|
|
auto val8 = vdupq_n_u16(val);
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
vst1q_u16(dst, val8);
|
|
dst += 8;
|
|
}
|
|
for (; cnt >= 4; cnt -= 4) {
|
|
vst1_u16(dst, val4);
|
|
dst += 4;
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
uint16_t arr[8] = {val, val, val, val, val, val, val, val};
|
|
auto val8 = _mm_loadu_ps((float*)arr);
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
_mm_storeu_ps((float*)dst, val8);
|
|
dst += 8;
|
|
}
|
|
#endif
|
|
}
|
|
for (; cnt > 0; --cnt) {
|
|
*dst = *src;
|
|
src += stride;
|
|
dst += ds;
|
|
}
|
|
}
|
|
|
|
void MNN1BitcopyWithStride(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
for (int i = 0; i < size; ++i) {
|
|
dstO[0] = *srcO;
|
|
dstO += ds;
|
|
srcO += stride;
|
|
}
|
|
}
|
|
|
|
void MNN1BitCopyFast(uint8_t* dstO, const uint8_t* srcO, int size, int stride, int ds) {
|
|
int cnt = size;
|
|
uint8_t val = *srcO;
|
|
if (stride == 1) {
|
|
#ifdef MNN_USE_SSE
|
|
for (; cnt >= 16; cnt -= 16) {
|
|
auto tmp = _mm_loadu_ps((float*)srcO);
|
|
_mm_storeu_ps((float*)dstO, tmp);
|
|
srcO += 16;
|
|
dstO += 16;
|
|
}
|
|
#elif defined(MNN_USE_NEON)
|
|
for (; cnt >= 16; cnt -= 16) {
|
|
auto val16 = vld1q_u8(srcO);
|
|
vst1q_u8(dstO, val16);
|
|
srcO += 16;
|
|
dstO += 16;
|
|
}
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
auto val8 = vld1_u8(srcO);
|
|
vst1_u8(dstO, val8);
|
|
srcO += 8;
|
|
dstO += 8;
|
|
}
|
|
#endif
|
|
} else { // stride=0
|
|
#ifdef MNN_USE_SSE
|
|
std::vector<uint8_t> arr(16, val);
|
|
auto val16 = _mm_loadu_ps((float*)arr.data());
|
|
|
|
for (; cnt >= 16; cnt -= 16) {
|
|
_mm_storeu_ps((float*)dstO, val16);
|
|
dstO += 16;
|
|
}
|
|
#elif defined(MNN_USE_NEON)
|
|
auto val16 = vdupq_n_u8(val);
|
|
auto val8 = vdup_n_u8(val);
|
|
for (; cnt >= 16; cnt -= 16) {
|
|
vst1q_u8(dstO, val16);
|
|
dstO += 16;
|
|
}
|
|
for (; cnt >= 8; cnt -= 8) {
|
|
vst1_u8(dstO, val8);
|
|
dstO += 8;
|
|
}
|
|
#endif
|
|
}
|
|
for (; cnt > 0; --cnt) {
|
|
dstO[0] = *srcO;
|
|
dstO += ds;
|
|
srcO += stride;
|
|
}
|
|
}
|
|
|
|
void MNNAccumulateSequenceNumber(float* dst, const float* src, int size) {
|
|
// mode: 0:Add, 1:Sub, 2:Min, 3:Max
|
|
int size8 = (size / 8) * 8;
|
|
int i = 0;
|
|
float sum = 0.f;
|
|
float tmp[4];
|
|
#ifdef MNN_USE_NEON
|
|
int size16 = (size / 16);
|
|
if (size >= 8) {
|
|
auto sum4_1 = vdupq_n_f32(0.f);
|
|
auto sum4_2 = vdupq_n_f32(0.f);
|
|
if (size >= 16) {
|
|
auto sum4_3 = vdupq_n_f32(0.f);
|
|
auto sum4_4 = vdupq_n_f32(0.f);
|
|
for (int v = 0; v < size16; ++v) {
|
|
auto v4 = vld1q_f32(src);
|
|
auto u4 = vld1q_f32(src + 4);
|
|
auto p4 = vld1q_f32(src + 8);
|
|
auto q4 = vld1q_f32(src + 12);
|
|
sum4_1 = vaddq_f32(sum4_1, v4);
|
|
sum4_2 = vaddq_f32(sum4_2, u4);
|
|
sum4_3 = vaddq_f32(sum4_3, p4);
|
|
sum4_4 = vaddq_f32(sum4_4, q4);
|
|
src += 16;
|
|
i += 16;
|
|
}
|
|
sum4_1 = vaddq_f32(sum4_1, sum4_3);
|
|
sum4_2 = vaddq_f32(sum4_2, sum4_4);
|
|
}
|
|
if (size - i >= 8) {
|
|
auto v4 = vld1q_f32(src);
|
|
auto u4 = vld1q_f32(src + 4);
|
|
sum4_1 = vaddq_f32(sum4_1, v4);
|
|
sum4_2 = vaddq_f32(sum4_2, u4);
|
|
src += 8;
|
|
i += 8;
|
|
}
|
|
sum4_1 = vaddq_f32(sum4_1, sum4_2);
|
|
sum = (sum4_1[0] + sum4_1[1]) + (sum4_1[2] + sum4_1[3]);
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
if (size >= 8) {
|
|
auto sum4_1 = _mm_set_ps1(0.f);
|
|
auto sum4_2 = _mm_set_ps1(0.f);
|
|
|
|
for (; i < size8; i += 8) {
|
|
auto v4 = _mm_loadu_ps(src);
|
|
auto u4 = _mm_loadu_ps(src + 4);
|
|
sum4_1 = _mm_add_ps(sum4_1, v4);
|
|
sum4_2 = _mm_add_ps(sum4_2, u4);
|
|
src += 8;
|
|
}
|
|
|
|
sum4_1 = _mm_add_ps(sum4_1, sum4_2);
|
|
_mm_storeu_ps(tmp, sum4_1);
|
|
sum += (tmp[0] + tmp[1] + tmp[2] + tmp[3]);
|
|
}
|
|
#endif
|
|
for (; i < size; ++i) {
|
|
sum += (*src);
|
|
src += 1;
|
|
}
|
|
*dst = sum;
|
|
}
|
|
|
|
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
|
|
|
|
static void MNNFlashAttentionUpdateBlockOutput(float* dst, float* src, float* scale, float* normalizeScale,
|
|
int depthQuad, int plane, int pack, int idx, int kvBlocks, int size,
|
|
int bytes, int seqStart) {
|
|
// source shape: [headDim/pack, seqLen, pack]
|
|
// scale & normalizeScale shape: [seqLen]
|
|
// dest shape: [headDim/pack, seqLen, pack]
|
|
auto stride0 = plane * pack;
|
|
|
|
if (idx > 0) {
|
|
for (int j = 0; j < depthQuad; ++j) {
|
|
int i = seqStart;
|
|
for (; i < plane; ++i) {
|
|
auto dataNew = Vec::load(src + j * stride0 + i * pack);
|
|
auto dataOld = Vec::load(dst + j * stride0 + i * pack);
|
|
auto s = Vec(scale[i]);
|
|
dataNew = Vec::fma(dataNew, dataOld, s);
|
|
Vec::save(dst + j * stride0 + i * pack, dataNew);
|
|
}
|
|
}
|
|
} else {
|
|
memcpy(dst, src, size * bytes);
|
|
}
|
|
if (idx == kvBlocks - 1) { // if last subBlock, exp(xi)/sum(exp(xi))
|
|
for (int j = 0; j < depthQuad; ++j) {
|
|
for (int i = 0; i < plane; ++i) {
|
|
auto dataNew = Vec::load(dst + j * stride0 + i * pack);
|
|
auto ns = Vec(1.0f / normalizeScale[i]);
|
|
dataNew = dataNew * ns;
|
|
Vec::save(dst + j * stride0 + i * pack, dataNew);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
static void MNNAttenPackAndScaleSingleHead(float* dst, const float* srcHeadBase, size_t srcRowStride,
|
|
const float* scale, const int32_t* units, size_t seqLen, size_t headDim) {
|
|
const int32_t eP = units[0];
|
|
const int32_t lP = units[1];
|
|
|
|
if (lP != 1) {
|
|
MNN_ERROR("This function only supports lP=1 or 2\n");
|
|
return;
|
|
}
|
|
|
|
const float scaleVal = scale[0];
|
|
#ifdef MNN_USE_NEON
|
|
const float32x4_t vScale = vdupq_n_f32(scaleVal);
|
|
#endif
|
|
const size_t packedHeadDim = UP_DIV(headDim, lP);
|
|
const size_t dstStrideDOuter = (size_t)eP * lP;
|
|
const size_t dstStrideSOuter = packedHeadDim * dstStrideDOuter;
|
|
|
|
for (int s = 0; s < seqLen; ++s) {
|
|
const int sOuter = s / eP;
|
|
const int sInner = s % eP;
|
|
const float* srcRowPtr = srcHeadBase + s * srcRowStride;
|
|
float* dstBasePtr = dst + sOuter * dstStrideSOuter + sInner * lP;
|
|
|
|
size_t d = 0;
|
|
#ifdef MNN_USE_NEON
|
|
for (; d + 7 < headDim; d += 8) {
|
|
float32x4_t sVec0 = vld1q_f32(srcRowPtr + d);
|
|
float32x4_t sVec1 = vld1q_f32(srcRowPtr + d + 4);
|
|
sVec0 = vmulq_f32(sVec0, vScale);
|
|
sVec1 = vmulq_f32(sVec1, vScale);
|
|
|
|
dstBasePtr[(d + 0) * dstStrideDOuter] = sVec0[0];
|
|
dstBasePtr[(d + 1) * dstStrideDOuter] = sVec0[1];
|
|
dstBasePtr[(d + 2) * dstStrideDOuter] = sVec0[2];
|
|
dstBasePtr[(d + 3) * dstStrideDOuter] = sVec0[3];
|
|
dstBasePtr[(d + 4) * dstStrideDOuter] = sVec1[0];
|
|
dstBasePtr[(d + 5) * dstStrideDOuter] = sVec1[1];
|
|
dstBasePtr[(d + 6) * dstStrideDOuter] = sVec1[2];
|
|
dstBasePtr[(d + 7) * dstStrideDOuter] = sVec1[3];
|
|
}
|
|
for (; d < headDim; ++d) {
|
|
dstBasePtr[d * dstStrideDOuter] = srcRowPtr[d] * scaleVal;
|
|
}
|
|
#else
|
|
for (; d < headDim; ++d) {
|
|
dstBasePtr[d * dstStrideDOuter] = srcRowPtr[d] * scaleVal;
|
|
}
|
|
#endif
|
|
}
|
|
}
|
|
|
|
#ifndef __aarch64__
|
|
void MNNQuantAttentionKey(int8_t* dst, const float* source, float* sumKeyPtr, float* maxKeyPtr, int32_t* params) {
|
|
int32_t kvNumHead = params[0];
|
|
int32_t seqLen = params[1];
|
|
int32_t headDim = params[2];
|
|
int32_t blockNum = params[3];
|
|
int32_t eP = params[4];
|
|
int32_t lP = params[5];
|
|
int32_t hP = params[6];
|
|
int32_t pastLength = params[7];
|
|
int32_t kvHeadIdx = params[8];
|
|
|
|
auto blockL = UP_DIV(headDim, blockNum);
|
|
auto weightStride1 = ROUND_UP(blockL, lP) * hP;
|
|
auto weightStride2 = lP * hP;
|
|
auto packedWeightStride1 = weightStride1 + 2 * 4 * hP;
|
|
|
|
if (seqLen > 1) {
|
|
// get max
|
|
for (int s = 0; s < seqLen; ++s) {
|
|
const float* keySrc = source + s * kvNumHead * headDim + kvHeadIdx * headDim;
|
|
for (int d = 0; d < headDim; d++) {
|
|
maxKeyPtr[d] = ALIMAX(maxKeyPtr[d], keySrc[d]);
|
|
}
|
|
}
|
|
}
|
|
|
|
for (int s = 0; s < seqLen; s++) {
|
|
const float* keySrc = source + s * kvNumHead * headDim + kvHeadIdx * headDim;
|
|
float minKey, maxKey;
|
|
minKey = keySrc[0] - maxKeyPtr[0];
|
|
maxKey = keySrc[0] - maxKeyPtr[0];
|
|
for (int d = 1; d < headDim; d++) {
|
|
auto keydata = keySrc[d] - maxKeyPtr[d];
|
|
minKey = ALIMIN(minKey, keydata);
|
|
maxKey = ALIMAX(maxKey, keydata);
|
|
}
|
|
|
|
int outIndex = (pastLength + s) / hP;
|
|
int inIndex = (pastLength + s) % hP;
|
|
|
|
float sumKey = 0;
|
|
for (int k = 0; k < blockNum; ++k) {
|
|
int8_t* weightDst = dst + outIndex * blockNum * packedWeightStride1 + k * packedWeightStride1;
|
|
float* scaleDst = (float*)(weightDst + weightStride1);
|
|
float* biasDst = scaleDst + hP;
|
|
|
|
scaleDst[inIndex] = (maxKey - minKey) / 255.0f;
|
|
biasDst[inIndex] = minKey + 128.f * (maxKey - minKey) / 255.f;
|
|
|
|
for (int d = 0; d < blockL; d++) {
|
|
int i = d / lP;
|
|
int j = d % lP;
|
|
|
|
int int8v = (int)(roundf((keySrc[d + k * blockL] - maxKeyPtr[d + k * blockL] - minKey) /
|
|
(maxKey - minKey) * 255.0f -
|
|
128.0f));
|
|
weightDst[i * weightStride2 + inIndex * lP + j] = int8v;
|
|
sumKey += (int8v * scaleDst[inIndex] + biasDst[inIndex]);
|
|
}
|
|
}
|
|
sumKeyPtr[outIndex * hP + inIndex] = sumKey;
|
|
}
|
|
}
|
|
|
|
void MNNQuantAttentionValue(int8_t* dst, const float* source, float* valueSum, int32_t* params) {
|
|
// float value src : [kvSeq,kvNumHead,headDim]
|
|
// int8_t value dest: [updiv(maxLength,flashAttentionBlockKv),
|
|
// updiv(headDim,hp),updiv(flashAttentionBlockKv,lp),hp,lp] float value sum:
|
|
// [updiv(maxLength,flashAttentionBlockKv), roundup(headDim,hp)]
|
|
int32_t kvNumHead = params[0];
|
|
int32_t seqLen = params[1];
|
|
int32_t headDim = params[2];
|
|
int32_t blockNum = params[3];
|
|
int32_t maxLength = params[4];
|
|
|
|
int32_t lP = params[5];
|
|
int32_t hP = params[6];
|
|
int32_t pastLength = params[7];
|
|
int32_t kvHeadIdx = params[8];
|
|
|
|
int32_t flashAttentionBlockKv = params[9];
|
|
|
|
auto blockKvseq = UP_DIV(seqLen + pastLength, blockNum);
|
|
auto weightStride2 = lP * hP;
|
|
auto weightStride1 = UP_DIV(flashAttentionBlockKv, lP) * weightStride2;
|
|
|
|
auto packedStride1 = (int)(weightStride1 + 2 * hP * sizeof(float));
|
|
auto packedStride0 = UP_DIV(headDim, hP) * packedStride1;
|
|
|
|
auto srcStride0 = kvNumHead * headDim;
|
|
|
|
auto sourceFp32 = (float*)source;
|
|
|
|
// quant scale & bias
|
|
if (pastLength == 0) {
|
|
for (int d = 0; d < headDim; ++d) {
|
|
float* scalePtr = (float*)(dst + (d / hP) * packedStride1 + weightStride1) + (d % hP);
|
|
float* biasPtr = scalePtr + hP;
|
|
|
|
// find min,max
|
|
float dMax = sourceFp32[d + kvHeadIdx * headDim];
|
|
float dMin = dMax;
|
|
for (int s = 0; s < seqLen; ++s) {
|
|
float data = sourceFp32[s * srcStride0 + d + kvHeadIdx * headDim];
|
|
dMax = ALIMAX(dMax, data);
|
|
dMin = ALIMIN(dMin, data);
|
|
}
|
|
|
|
// scale & bias
|
|
float range = dMax - dMin;
|
|
if (range < 1e-6) {
|
|
scalePtr[0] = 0.f;
|
|
biasPtr[0] = dMax;
|
|
} else {
|
|
float scale = range / 255.f;
|
|
float bias = range / 255.f * 128.f + dMin;
|
|
scalePtr[0] = scale;
|
|
biasPtr[0] = bias;
|
|
}
|
|
}
|
|
}
|
|
|
|
// copy the scale&bias to each blockKv
|
|
// pastLength == 0: First time prefill
|
|
// (seqLen + pastLength) % flashAttentionBlockKv == 0: Open a new blockKv
|
|
if (pastLength == 0 || (pastLength % flashAttentionBlockKv) == 0) {
|
|
int32_t d0 = UP_DIV(maxLength, flashAttentionBlockKv);
|
|
int32_t d1 = UP_DIV(headDim, hP);
|
|
for (int k = 0; k < d0; ++k) {
|
|
for (int r = 0; r < d1; ++r) {
|
|
float* scalePtr = (float*)(dst + k * packedStride0 + r * packedStride1 + weightStride1);
|
|
float* biasPtr = scalePtr + hP;
|
|
memcpy(scalePtr, dst + r * packedStride1 + weightStride1, hP * sizeof(float));
|
|
memcpy(biasPtr, dst + r * packedStride1 + weightStride1 + hP * sizeof(float), hP * sizeof(float));
|
|
}
|
|
}
|
|
}
|
|
|
|
for (int d = 0; d < headDim; ++d) {
|
|
// dst address
|
|
int idxBase = (d / hP) * packedStride1 + (d % hP) * lP;
|
|
int8_t* dstBase = dst + idxBase;
|
|
float* scaleBase = (float*)(dst + (d / hP) * packedStride1 + weightStride1) + (d % hP);
|
|
float* biasBase = scaleBase + hP;
|
|
float* sumBase = valueSum + (d / hP) * hP + (d % hP);
|
|
|
|
float qscale = scaleBase[0] < 1e-6 ? 0 : 1.0f / scaleBase[0];
|
|
float qbias = scaleBase[0] < 1e-6 ? 0 : (-biasBase[0] / scaleBase[0]);
|
|
// quant
|
|
for (int s = 0; s < seqLen; ++s) {
|
|
int kvSeqIndx = s + pastLength;
|
|
int idxInner = (kvSeqIndx / flashAttentionBlockKv) * packedStride0 +
|
|
(kvSeqIndx % flashAttentionBlockKv) / lP * weightStride2 +
|
|
(kvSeqIndx % flashAttentionBlockKv) % lP;
|
|
float xf = sourceFp32[s * srcStride0 + d + kvHeadIdx * headDim];
|
|
int8_t xq = ALIMAX(ALIMIN(127, static_cast<int32_t>(roundf(xf * qscale + qbias))), -128);
|
|
dstBase[idxInner] = xq;
|
|
|
|
// sum
|
|
int idxSum = (kvSeqIndx / flashAttentionBlockKv) * ROUND_UP(headDim, hP);
|
|
sumBase[idxSum] += ((float)xq * scaleBase[0] + biasBase[0]);
|
|
}
|
|
}
|
|
}
|
|
|
|
#endif
|
|
|
|
#endif // MNN_SUPPORT_TRANSFORMER_FUSE
|
|
|
|
#ifndef MNN_USE_NEON
|
|
|
|
void MNNGetMatMulPackMode(int* eP, int* lP, int* hP) {
|
|
*eP = 16;
|
|
*lP = 1;
|
|
*hP = 4;
|
|
}
|
|
|
|
void MNNGetSparseMatMulPackMode(int* eP, int* lP, int* hP) {
|
|
*eP = 16;
|
|
*lP = 1;
|
|
*hP = 4;
|
|
// hp is corresponding to sparse block along right matrix colum dimension. in ramdom sparse, it is 1.
|
|
return;
|
|
}
|
|
|
|
void MNNPackForMatMul_B(float* dest, const float* source, size_t h, size_t kernelsize, size_t ic, bool transpose) {
|
|
// src: [h, kernelsize, ic]
|
|
auto hP = h / 4;
|
|
auto hR = hP * 4;
|
|
auto l = kernelsize * ic;
|
|
if (hR != h) {
|
|
::memset(dest, 0, UP_DIV(h, 4) * 4 * l * sizeof(float));
|
|
}
|
|
if (!transpose) {
|
|
for (int y = 0; y < hP; ++y) {
|
|
auto destY = dest + y * 4 * l;
|
|
auto sourceY = source + y * 4;
|
|
for (int x = 0; x < l; ++x) {
|
|
::memcpy(destY + 4 * x, sourceY + x * h, 4 * sizeof(float));
|
|
}
|
|
}
|
|
auto hRemain = h - hR;
|
|
if (hRemain > 0) {
|
|
auto destY = dest + hP * 4 * l;
|
|
auto sourceY = source + hP * 4;
|
|
for (int x = 0; x < l; ++x) {
|
|
::memcpy(destY + 4 * x, sourceY + x * h, hRemain * sizeof(float));
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
int offset[] = {(int)l, (int)l};
|
|
MNNPackC4(dest, source, l, h, offset);
|
|
}
|
|
|
|
static void _MNNPackedMatMulRemain(float* C, const float* A, const float* B, size_t eSize, const size_t* parameter,
|
|
const float* postParameters, const float* bias, int aStride) {
|
|
auto h = parameter[2];
|
|
auto l = parameter[1];
|
|
auto cStride = parameter[3] / sizeof(float);
|
|
auto hRemain = parameter[4];
|
|
auto bExtraStride = parameter[5] / sizeof(float);
|
|
auto bStride = bExtraStride + l * 4;
|
|
auto hC4 = UP_DIV(h, 4);
|
|
for (int y = 0; y < hC4; ++y) {
|
|
::memset(C + y * cStride, 0, eSize * 4 * sizeof(float));
|
|
}
|
|
float alpha = 1.0f;
|
|
float beta = 0.0f;
|
|
float minValue = -std::numeric_limits<float>().max();
|
|
float maxValue = std::numeric_limits<float>().max();
|
|
if (nullptr != postParameters) {
|
|
minValue = postParameters[2];
|
|
maxValue = postParameters[3];
|
|
alpha = postParameters[0];
|
|
beta = postParameters[1];
|
|
}
|
|
|
|
for (int x = 0; x < eSize; ++x) {
|
|
auto dst = C + 4 * x;
|
|
auto src = A + x;
|
|
for (int y = 0; y < hC4; ++y) {
|
|
auto dstY = dst + y * cStride;
|
|
auto weight = B + y * bStride;
|
|
float summer[4] = {
|
|
0.0f,
|
|
0.0f,
|
|
0.0f,
|
|
0.0f,
|
|
};
|
|
if (nullptr != bias) {
|
|
for (int v = 0; v < 4; ++v) {
|
|
summer[v] = bias[4 * y + v];
|
|
}
|
|
}
|
|
for (int z = 0; z < l; ++z) {
|
|
auto aZ = src + z * aStride;
|
|
auto wZ = weight + z * 4;
|
|
summer[0] += wZ[0] * aZ[0];
|
|
summer[1] += wZ[1] * aZ[0];
|
|
summer[2] += wZ[2] * aZ[0];
|
|
summer[3] += wZ[3] * aZ[0];
|
|
}
|
|
for (int v = 0; v < 4; ++v) {
|
|
auto dstValue = std::min(summer[v], maxValue);
|
|
dstValue = std::max(dstValue, minValue);
|
|
dstY[v] = dstValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackedMatMul(float* C, const float* A, const float* B, const size_t* parameter, const float* postParameters,
|
|
const float* bias, const float* k, const float* b) {
|
|
return _MNNPackedMatMulRemain(C, A, B, 16, parameter, postParameters, bias, 16);
|
|
}
|
|
|
|
void MNNPackedMatMulRemain(float* C, const float* A, const float* B, size_t eSize, const size_t* parameter,
|
|
const float* postParameters, const float* bias, const float* k, const float* b) {
|
|
auto aStride = parameter[0] / sizeof(float);
|
|
_MNNPackedMatMulRemain(C, A, B, eSize, parameter, postParameters, bias, aStride);
|
|
}
|
|
|
|
void MNNPackC4ForMatMul_A(float* destOrigin, float const** sourceGroup, const int32_t* info, const int32_t* el) {
|
|
int number = info[0];
|
|
int eReal = info[1];
|
|
int eDest = info[2];
|
|
int offset = info[3];
|
|
for (int n = 0; n < number; ++n) {
|
|
int e = el[4 * n + 0];
|
|
int l = el[4 * n + 1];
|
|
int eOffset = el[4 * n + 2];
|
|
int lOffset = el[4 * n + 3];
|
|
auto dest = destOrigin + lOffset * eDest + eOffset;
|
|
auto source = sourceGroup[n];
|
|
|
|
for (int y = 0; y < e; ++y) {
|
|
auto yR = y % eDest;
|
|
for (int x = 0; x < l; ++x) {
|
|
auto xR = x % 4;
|
|
auto xC = x / 4;
|
|
dest[(x)*eDest + yR] = source[xC * eReal * 4 + y * 4 * offset + xR];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackedSparseMatMulEpx1(float* C, const float* A, const float* B, size_t eSize, const size_t* parameter,
|
|
const float* postParameters, const float* bias, unsigned int* NNZMap,
|
|
int* dataOffsetMap) {
|
|
auto eP = parameter[0] / sizeof(float);
|
|
MNN_ASSERT((eP & 0x03) == 0); // In sparse calculate, eP should be evenly divided by 4
|
|
auto h = parameter[2];
|
|
auto l = parameter[1];
|
|
auto cStride = parameter[3] / sizeof(float);
|
|
auto aStride = eP * l;
|
|
auto hRemain = parameter[4];
|
|
auto bExtraStride = parameter[5] / sizeof(float);
|
|
auto bStride = bExtraStride + l * 4;
|
|
auto hC4 = UP_DIV(h, 4);
|
|
float minValue = -std::numeric_limits<float>().max();
|
|
float maxValue = std::numeric_limits<float>().max();
|
|
if (nullptr != postParameters) {
|
|
minValue = postParameters[2];
|
|
maxValue = postParameters[3];
|
|
}
|
|
// MNN_PRINT("MNNPackedSparseMatMul eP:%lu, eSize:%lu, l:%lu, h:%lu, cStride:%lu, aStride:%lu\n", eP, eSize, l, h,
|
|
// cStride, aStride);
|
|
|
|
const float* a = A;
|
|
size_t ie = 0;
|
|
for (ie = 0; ie < eSize && eP <= eSize; ie += eP) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
for (auto ih = 0; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihPack * cStride + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
float acc4 = initValue;
|
|
float acc5 = initValue;
|
|
float acc6 = initValue;
|
|
float acc7 = initValue;
|
|
float acc8 = initValue;
|
|
float acc9 = initValue;
|
|
float acc10 = initValue;
|
|
float acc11 = initValue;
|
|
float acc12 = initValue;
|
|
float acc13 = initValue;
|
|
float acc14 = initValue;
|
|
float acc15 = initValue;
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float a8 = a[8];
|
|
const float a9 = a[9];
|
|
const float a10 = a[10];
|
|
const float a11 = a[11];
|
|
const float a12 = a[12];
|
|
const float a13 = a[13];
|
|
const float a14 = a[14];
|
|
const float a15 = a[15];
|
|
|
|
const float oneW = *w++;
|
|
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
acc4 += a4 * oneW;
|
|
acc5 += a5 * oneW;
|
|
acc6 += a6 * oneW;
|
|
acc7 += a7 * oneW;
|
|
acc8 += a8 * oneW;
|
|
acc9 += a9 * oneW;
|
|
acc10 += a10 * oneW;
|
|
acc11 += a11 * oneW;
|
|
acc12 += a12 * oneW;
|
|
acc13 += a13 * oneW;
|
|
acc14 += a14 * oneW;
|
|
acc15 += a15 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
acc4 = std::max(std::min(maxValue, acc4), minValue);
|
|
acc5 = std::max(std::min(maxValue, acc5), minValue);
|
|
acc6 = std::max(std::min(maxValue, acc6), minValue);
|
|
acc7 = std::max(std::min(maxValue, acc7), minValue);
|
|
acc8 = std::max(std::min(maxValue, acc8), minValue);
|
|
acc9 = std::max(std::min(maxValue, acc9), minValue);
|
|
acc10 = std::max(std::min(maxValue, acc10), minValue);
|
|
acc11 = std::max(std::min(maxValue, acc11), minValue);
|
|
acc12 = std::max(std::min(maxValue, acc12), minValue);
|
|
acc13 = std::max(std::min(maxValue, acc13), minValue);
|
|
acc14 = std::max(std::min(maxValue, acc14), minValue);
|
|
acc15 = std::max(std::min(maxValue, acc15), minValue);
|
|
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
c[4 * 4] = acc4;
|
|
c[4 * 5] = acc5;
|
|
c[4 * 6] = acc6;
|
|
c[4 * 7] = acc7;
|
|
c[4 * 8] = acc8;
|
|
c[4 * 9] = acc9;
|
|
c[4 * 10] = acc10;
|
|
c[4 * 11] = acc11;
|
|
c[4 * 12] = acc12;
|
|
c[4 * 13] = acc13;
|
|
c[4 * 14] = acc14;
|
|
c[4 * 15] = acc15;
|
|
}
|
|
a += aStride;
|
|
}
|
|
// const float* blockA = A + ie * l;
|
|
if (eSize & 0x08) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
for (auto ih = 0; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihPack * cStride + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
float acc4 = initValue;
|
|
float acc5 = initValue;
|
|
float acc6 = initValue;
|
|
float acc7 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("8-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-7]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {8}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
acc4 += a4 * oneW;
|
|
acc5 += a5 * oneW;
|
|
acc6 += a6 * oneW;
|
|
acc7 += a7 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
acc4 = std::max(std::min(maxValue, acc4), minValue);
|
|
acc5 = std::max(std::min(maxValue, acc5), minValue);
|
|
acc6 = std::max(std::min(maxValue, acc6), minValue);
|
|
acc7 = std::max(std::min(maxValue, acc7), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
c[4 * 4] = acc4;
|
|
c[4 * 5] = acc5;
|
|
c[4 * 6] = acc6;
|
|
c[4 * 7] = acc7;
|
|
}
|
|
ie += 8;
|
|
a += 8;
|
|
}
|
|
|
|
if (eSize & 0x04) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
for (auto ih = 0; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihPack * cStride + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("4-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-3]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {4}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
}
|
|
ie += 4;
|
|
a += 4;
|
|
}
|
|
if (eSize & 0x02) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
for (auto ih = 0; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihPack * cStride + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("2-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-1]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {2}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
}
|
|
ie += 2;
|
|
a += 2;
|
|
}
|
|
if (eSize & 0x01) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
for (auto ih = 0; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihPack * cStride + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float oneW = *w++;
|
|
|
|
// MNN_PRINT("1-loop: ie:%zu, a offset:%ld, c offset:%ld, w offset:%ld, w value:%f, a value[0]:", ie, a
|
|
// - A, w - B - 1, c - C, oneW); formatMatrix(a, {1}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
}
|
|
ie += 1;
|
|
// a += 1;
|
|
}
|
|
|
|
return;
|
|
}
|
|
|
|
void MNNPackedSparseMatMulEpx4(float* C, const float* A, const float* B, size_t eSize, const size_t* parameter,
|
|
const float* postParameters, const float* bias, unsigned int* NNZMap,
|
|
int* dataOffsetMap) {
|
|
auto eP = parameter[0] / sizeof(float);
|
|
MNN_ASSERT((eP & 0x03) == 0); // In sparse calculate, eP should be evenly divided by 4
|
|
auto h = parameter[2];
|
|
auto l = parameter[1];
|
|
auto cStride = parameter[3] / sizeof(float);
|
|
auto aStride = eP * l;
|
|
auto hRemain = parameter[4];
|
|
auto bExtraStride = parameter[5] / sizeof(float);
|
|
auto bStride = bExtraStride + l * 4;
|
|
auto hC4 = UP_DIV(h, 4);
|
|
float minValue = -std::numeric_limits<float>().max();
|
|
float maxValue = std::numeric_limits<float>().max();
|
|
if (nullptr != postParameters) {
|
|
minValue = postParameters[2];
|
|
maxValue = postParameters[3];
|
|
}
|
|
// MNN_PRINT("MNNPackedSparseMatMul 16x4 eP:%lu, eSize:%lu, l:%lu, h:%lu, cStride:%lu, aStride:%lu\n", eP, eSize, l,
|
|
// h, cStride, aStride);
|
|
const int sparseBlockOC = 4;
|
|
const float* a = A;
|
|
size_t ie = 0;
|
|
for (ie = 0; ie < eSize && eP <= eSize; ie += eP) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
|
|
size_t ih = 0;
|
|
for (; ih < (h & (~0x03)); ih += sparseBlockOC) {
|
|
auto ihPack = ih >> 2;
|
|
auto c = blockC + ihPack * cStride;
|
|
|
|
float initValue[4] = {0, 0, 0, 0};
|
|
if (nullptr != bias) {
|
|
memcpy(initValue, bias + ih, 4 * sizeof(float));
|
|
}
|
|
float acc0[4];
|
|
float acc1[4];
|
|
float acc2[4];
|
|
float acc3[4];
|
|
float acc4[4];
|
|
float acc5[4];
|
|
float acc6[4];
|
|
float acc7[4];
|
|
float acc8[4];
|
|
float acc9[4];
|
|
float acc10[4];
|
|
float acc11[4];
|
|
float acc12[4];
|
|
float acc13[4];
|
|
float acc14[4];
|
|
float acc15[4];
|
|
memcpy(acc0, initValue, 4 * sizeof(float));
|
|
memcpy(acc1, initValue, 4 * sizeof(float));
|
|
memcpy(acc2, initValue, 4 * sizeof(float));
|
|
memcpy(acc3, initValue, 4 * sizeof(float));
|
|
memcpy(acc4, initValue, 4 * sizeof(float));
|
|
memcpy(acc5, initValue, 4 * sizeof(float));
|
|
memcpy(acc6, initValue, 4 * sizeof(float));
|
|
memcpy(acc7, initValue, 4 * sizeof(float));
|
|
memcpy(acc8, initValue, 4 * sizeof(float));
|
|
memcpy(acc9, initValue, 4 * sizeof(float));
|
|
memcpy(acc10, initValue, 4 * sizeof(float));
|
|
memcpy(acc11, initValue, 4 * sizeof(float));
|
|
memcpy(acc12, initValue, 4 * sizeof(float));
|
|
memcpy(acc13, initValue, 4 * sizeof(float));
|
|
memcpy(acc14, initValue, 4 * sizeof(float));
|
|
memcpy(acc15, initValue, 4 * sizeof(float));
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float a8 = a[8];
|
|
const float a9 = a[9];
|
|
const float a10 = a[10];
|
|
const float a11 = a[11];
|
|
const float a12 = a[12];
|
|
const float a13 = a[13];
|
|
const float a14 = a[14];
|
|
const float a15 = a[15];
|
|
|
|
const float wv[4] = {*w++, *w++, *w++, *w++};
|
|
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] += a0 * wv[lane];
|
|
acc1[lane] += a1 * wv[lane];
|
|
acc2[lane] += a2 * wv[lane];
|
|
acc3[lane] += a3 * wv[lane];
|
|
acc4[lane] += a4 * wv[lane];
|
|
acc5[lane] += a5 * wv[lane];
|
|
acc6[lane] += a6 * wv[lane];
|
|
acc7[lane] += a7 * wv[lane];
|
|
acc8[lane] += a8 * wv[lane];
|
|
acc9[lane] += a9 * wv[lane];
|
|
acc10[lane] += a10 * wv[lane];
|
|
acc11[lane] += a11 * wv[lane];
|
|
acc12[lane] += a12 * wv[lane];
|
|
acc13[lane] += a13 * wv[lane];
|
|
acc14[lane] += a14 * wv[lane];
|
|
acc15[lane] += a15 * wv[lane];
|
|
}
|
|
}
|
|
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] = std::max(std::min(maxValue, acc0[lane]), minValue);
|
|
acc1[lane] = std::max(std::min(maxValue, acc1[lane]), minValue);
|
|
acc2[lane] = std::max(std::min(maxValue, acc2[lane]), minValue);
|
|
acc3[lane] = std::max(std::min(maxValue, acc3[lane]), minValue);
|
|
acc4[lane] = std::max(std::min(maxValue, acc4[lane]), minValue);
|
|
acc5[lane] = std::max(std::min(maxValue, acc5[lane]), minValue);
|
|
acc6[lane] = std::max(std::min(maxValue, acc6[lane]), minValue);
|
|
acc7[lane] = std::max(std::min(maxValue, acc7[lane]), minValue);
|
|
acc8[lane] = std::max(std::min(maxValue, acc8[lane]), minValue);
|
|
acc9[lane] = std::max(std::min(maxValue, acc9[lane]), minValue);
|
|
acc10[lane] = std::max(std::min(maxValue, acc10[lane]), minValue);
|
|
acc11[lane] = std::max(std::min(maxValue, acc11[lane]), minValue);
|
|
acc12[lane] = std::max(std::min(maxValue, acc12[lane]), minValue);
|
|
acc13[lane] = std::max(std::min(maxValue, acc13[lane]), minValue);
|
|
acc14[lane] = std::max(std::min(maxValue, acc14[lane]), minValue);
|
|
acc15[lane] = std::max(std::min(maxValue, acc15[lane]), minValue);
|
|
}
|
|
|
|
memcpy(c, acc0, 4 * sizeof(float)); // store continuous c
|
|
memcpy(c + 4, acc1, 4 * sizeof(float));
|
|
memcpy(c + 4 * 2, acc2, 4 * sizeof(float));
|
|
memcpy(c + 4 * 3, acc3, 4 * sizeof(float));
|
|
memcpy(c + 4 * 4, acc4, 4 * sizeof(float));
|
|
memcpy(c + 4 * 5, acc5, 4 * sizeof(float));
|
|
memcpy(c + 4 * 6, acc6, 4 * sizeof(float));
|
|
memcpy(c + 4 * 7, acc7, 4 * sizeof(float));
|
|
memcpy(c + 4 * 8, acc8, 4 * sizeof(float));
|
|
memcpy(c + 4 * 9, acc9, 4 * sizeof(float));
|
|
memcpy(c + 4 * 10, acc10, 4 * sizeof(float));
|
|
memcpy(c + 4 * 11, acc11, 4 * sizeof(float));
|
|
memcpy(c + 4 * 12, acc12, 4 * sizeof(float));
|
|
memcpy(c + 4 * 13, acc13, 4 * sizeof(float));
|
|
memcpy(c + 4 * 14, acc14, 4 * sizeof(float));
|
|
memcpy(c + 4 * 15, acc15, 4 * sizeof(float));
|
|
}
|
|
|
|
blockC += (h >> 2) * cStride;
|
|
for (; ih < h; ih++) {
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
float acc4 = initValue;
|
|
float acc5 = initValue;
|
|
float acc6 = initValue;
|
|
float acc7 = initValue;
|
|
float acc8 = initValue;
|
|
float acc9 = initValue;
|
|
float acc10 = initValue;
|
|
float acc11 = initValue;
|
|
float acc12 = initValue;
|
|
float acc13 = initValue;
|
|
float acc14 = initValue;
|
|
float acc15 = initValue;
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float a8 = a[8];
|
|
const float a9 = a[9];
|
|
const float a10 = a[10];
|
|
const float a11 = a[11];
|
|
const float a12 = a[12];
|
|
const float a13 = a[13];
|
|
const float a14 = a[14];
|
|
const float a15 = a[15];
|
|
|
|
const float oneW = *w++;
|
|
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
acc4 += a4 * oneW;
|
|
acc5 += a5 * oneW;
|
|
acc6 += a6 * oneW;
|
|
acc7 += a7 * oneW;
|
|
acc8 += a8 * oneW;
|
|
acc9 += a9 * oneW;
|
|
acc10 += a10 * oneW;
|
|
acc11 += a11 * oneW;
|
|
acc12 += a12 * oneW;
|
|
acc13 += a13 * oneW;
|
|
acc14 += a14 * oneW;
|
|
acc15 += a15 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
acc4 = std::max(std::min(maxValue, acc4), minValue);
|
|
acc5 = std::max(std::min(maxValue, acc5), minValue);
|
|
acc6 = std::max(std::min(maxValue, acc6), minValue);
|
|
acc7 = std::max(std::min(maxValue, acc7), minValue);
|
|
acc8 = std::max(std::min(maxValue, acc8), minValue);
|
|
acc9 = std::max(std::min(maxValue, acc9), minValue);
|
|
acc10 = std::max(std::min(maxValue, acc10), minValue);
|
|
acc11 = std::max(std::min(maxValue, acc11), minValue);
|
|
acc12 = std::max(std::min(maxValue, acc12), minValue);
|
|
acc13 = std::max(std::min(maxValue, acc13), minValue);
|
|
acc14 = std::max(std::min(maxValue, acc14), minValue);
|
|
acc15 = std::max(std::min(maxValue, acc15), minValue);
|
|
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
c[4 * 4] = acc4;
|
|
c[4 * 5] = acc5;
|
|
c[4 * 6] = acc6;
|
|
c[4 * 7] = acc7;
|
|
c[4 * 8] = acc8;
|
|
c[4 * 9] = acc9;
|
|
c[4 * 10] = acc10;
|
|
c[4 * 11] = acc11;
|
|
c[4 * 12] = acc12;
|
|
c[4 * 13] = acc13;
|
|
c[4 * 14] = acc14;
|
|
c[4 * 15] = acc15;
|
|
}
|
|
a += aStride;
|
|
}
|
|
// const float* blockA = A + ie * l;
|
|
if (eSize & 0x08) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
|
|
size_t ih = 0;
|
|
for (; ih < (h & (~0x03)); ih += sparseBlockOC) {
|
|
auto ihPack = ih >> 2;
|
|
auto c = blockC + ihPack * cStride;
|
|
float initValue[4] = {0, 0, 0, 0};
|
|
if (nullptr != bias) {
|
|
memcpy(initValue, bias + ih, 4 * sizeof(float));
|
|
}
|
|
float acc0[4];
|
|
float acc1[4];
|
|
float acc2[4];
|
|
float acc3[4];
|
|
float acc4[4];
|
|
float acc5[4];
|
|
float acc6[4];
|
|
float acc7[4];
|
|
memcpy(acc0, initValue, 4 * sizeof(float));
|
|
memcpy(acc1, initValue, 4 * sizeof(float));
|
|
memcpy(acc2, initValue, 4 * sizeof(float));
|
|
memcpy(acc3, initValue, 4 * sizeof(float));
|
|
memcpy(acc4, initValue, 4 * sizeof(float));
|
|
memcpy(acc5, initValue, 4 * sizeof(float));
|
|
memcpy(acc6, initValue, 4 * sizeof(float));
|
|
memcpy(acc7, initValue, 4 * sizeof(float));
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float wv[4] = {*w++, *w++, *w++, *w++};
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] += a0 * wv[lane];
|
|
acc1[lane] += a1 * wv[lane];
|
|
acc2[lane] += a2 * wv[lane];
|
|
acc3[lane] += a3 * wv[lane];
|
|
acc4[lane] += a4 * wv[lane];
|
|
acc5[lane] += a5 * wv[lane];
|
|
acc6[lane] += a6 * wv[lane];
|
|
acc7[lane] += a7 * wv[lane];
|
|
}
|
|
}
|
|
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] = std::max(std::min(maxValue, acc0[lane]), minValue);
|
|
acc1[lane] = std::max(std::min(maxValue, acc1[lane]), minValue);
|
|
acc2[lane] = std::max(std::min(maxValue, acc2[lane]), minValue);
|
|
acc3[lane] = std::max(std::min(maxValue, acc3[lane]), minValue);
|
|
acc4[lane] = std::max(std::min(maxValue, acc4[lane]), minValue);
|
|
acc5[lane] = std::max(std::min(maxValue, acc5[lane]), minValue);
|
|
acc6[lane] = std::max(std::min(maxValue, acc6[lane]), minValue);
|
|
acc7[lane] = std::max(std::min(maxValue, acc7[lane]), minValue);
|
|
}
|
|
|
|
memcpy(c, acc0, 4 * sizeof(float)); // store continuous c
|
|
memcpy(c + 4, acc1, 4 * sizeof(float));
|
|
memcpy(c + 4 * 2, acc2, 4 * sizeof(float));
|
|
memcpy(c + 4 * 3, acc3, 4 * sizeof(float));
|
|
memcpy(c + 4 * 4, acc4, 4 * sizeof(float));
|
|
memcpy(c + 4 * 5, acc5, 4 * sizeof(float));
|
|
memcpy(c + 4 * 6, acc6, 4 * sizeof(float));
|
|
memcpy(c + 4 * 7, acc7, 4 * sizeof(float));
|
|
}
|
|
blockC += (ih >> 2) * cStride;
|
|
for (; ih < h; ih++) {
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
float acc4 = initValue;
|
|
float acc5 = initValue;
|
|
float acc6 = initValue;
|
|
float acc7 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float a4 = a[4];
|
|
const float a5 = a[5];
|
|
const float a6 = a[6];
|
|
const float a7 = a[7];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("8-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-7]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {8}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
acc4 += a4 * oneW;
|
|
acc5 += a5 * oneW;
|
|
acc6 += a6 * oneW;
|
|
acc7 += a7 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
acc4 = std::max(std::min(maxValue, acc4), minValue);
|
|
acc5 = std::max(std::min(maxValue, acc5), minValue);
|
|
acc6 = std::max(std::min(maxValue, acc6), minValue);
|
|
acc7 = std::max(std::min(maxValue, acc7), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
c[4 * 4] = acc4;
|
|
c[4 * 5] = acc5;
|
|
c[4 * 6] = acc6;
|
|
c[4 * 7] = acc7;
|
|
}
|
|
ie += 8;
|
|
a += 8;
|
|
}
|
|
|
|
if (eSize & 0x04) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
|
|
size_t ih = 0;
|
|
for (; ih < (h & (~0x03)); ih += sparseBlockOC) {
|
|
auto ihPack = ih >> 2;
|
|
auto c = blockC + ihPack * cStride;
|
|
float initValue[4] = {0, 0, 0, 0};
|
|
if (nullptr != bias) {
|
|
memcpy(initValue, bias + ih, 4 * sizeof(float));
|
|
}
|
|
float acc0[4];
|
|
float acc1[4];
|
|
float acc2[4];
|
|
float acc3[4];
|
|
memcpy(acc0, initValue, 4 * sizeof(float));
|
|
memcpy(acc1, initValue, 4 * sizeof(float));
|
|
memcpy(acc2, initValue, 4 * sizeof(float));
|
|
memcpy(acc3, initValue, 4 * sizeof(float));
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float wv[4] = {*w++, *w++, *w++, *w++};
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] += a0 * wv[lane];
|
|
acc1[lane] += a1 * wv[lane];
|
|
acc2[lane] += a2 * wv[lane];
|
|
acc3[lane] += a3 * wv[lane];
|
|
}
|
|
}
|
|
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] = std::max(std::min(maxValue, acc0[lane]), minValue);
|
|
acc1[lane] = std::max(std::min(maxValue, acc1[lane]), minValue);
|
|
acc2[lane] = std::max(std::min(maxValue, acc2[lane]), minValue);
|
|
acc3[lane] = std::max(std::min(maxValue, acc3[lane]), minValue);
|
|
}
|
|
|
|
memcpy(c, acc0, 4 * sizeof(float)); // store continuous c
|
|
memcpy(c + 4, acc1, 4 * sizeof(float));
|
|
memcpy(c + 4 * 2, acc2, 4 * sizeof(float));
|
|
memcpy(c + 4 * 3, acc3, 4 * sizeof(float));
|
|
}
|
|
blockC += (ih >> 2) * cStride;
|
|
for (; ih < h; ih++) {
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
float acc2 = initValue;
|
|
float acc3 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float a2 = a[2];
|
|
const float a3 = a[3];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("4-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-3]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {4}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
acc2 += a2 * oneW;
|
|
acc3 += a3 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
acc2 = std::max(std::min(maxValue, acc2), minValue);
|
|
acc3 = std::max(std::min(maxValue, acc3), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
c[4 * 2] = acc2;
|
|
c[4 * 3] = acc3;
|
|
}
|
|
ie += 4;
|
|
a += 4;
|
|
}
|
|
if (eSize & 0x02) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
|
|
size_t ih = 0;
|
|
for (; ih < (h & (~0x03)); ih += sparseBlockOC) {
|
|
auto ihPack = ih >> 2;
|
|
auto c = blockC + ihPack * cStride;
|
|
float initValue[4] = {0, 0, 0, 0};
|
|
if (nullptr != bias) {
|
|
memcpy(initValue, bias + ih, 4 * sizeof(float));
|
|
}
|
|
float acc0[4];
|
|
float acc1[4];
|
|
memcpy(acc0, initValue, 4 * sizeof(float));
|
|
memcpy(acc1, initValue, 4 * sizeof(float));
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float wv[4] = {*w++, *w++, *w++, *w++};
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] += a0 * wv[lane];
|
|
acc1[lane] += a1 * wv[lane];
|
|
}
|
|
}
|
|
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] = std::max(std::min(maxValue, acc0[lane]), minValue);
|
|
acc1[lane] = std::max(std::min(maxValue, acc1[lane]), minValue);
|
|
}
|
|
|
|
memcpy(c, acc0, 4 * sizeof(float)); // store continuous c
|
|
memcpy(c + 4, acc1, 4 * sizeof(float));
|
|
}
|
|
blockC += (ih >> 2) * cStride;
|
|
for (; ih < h; ih++) {
|
|
auto ihPack = ih >> 2;
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
float acc1 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float a1 = a[1];
|
|
const float oneW = *w++;
|
|
// MNN_PRINT("2-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-1]:", ie,
|
|
// a - A, w - B - 1, c - C, oneW); formatMatrix(a, {2}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
acc1 += a1 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
acc1 = std::max(std::min(maxValue, acc1), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
c[4] = acc1;
|
|
}
|
|
ie += 2;
|
|
a += 2;
|
|
}
|
|
if (eSize & 0x01) {
|
|
const int* dataOffset = dataOffsetMap;
|
|
const int diff = *dataOffset++;
|
|
// const float* a = blockA + diff;
|
|
a += diff;
|
|
const float* w = B;
|
|
float* blockC = C + (ie << 2);
|
|
const unsigned int* nnz = NNZMap;
|
|
|
|
size_t ih = 0;
|
|
for (; ih < (h & (~0x03)); ih += sparseBlockOC) {
|
|
auto ihPack = ih >> 2;
|
|
auto c = blockC + ihPack * cStride;
|
|
float initValue[4] = {0, 0, 0, 0};
|
|
if (nullptr != bias) {
|
|
memcpy(initValue, bias + ih, 4 * sizeof(float));
|
|
}
|
|
float acc0[4];
|
|
memcpy(acc0, initValue, 4 * sizeof(float));
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float wv[4] = {*w++, *w++, *w++, *w++};
|
|
// MNN_PRINT("16-loop: ie:%zu, a offset:%ld, w offset:%ld, c offset:%ld, w value:%f, a value[0-15]:",
|
|
// ie, a - A, w - B - 1, c - C, oneW); formatMatrix(a, {16}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] += a0 * wv[lane];
|
|
}
|
|
}
|
|
|
|
for (int lane = 0; lane < 4; lane++) {
|
|
acc0[lane] = std::max(std::min(maxValue, acc0[lane]), minValue);
|
|
}
|
|
memcpy(c, acc0, 4 * sizeof(float)); // store continuous c
|
|
}
|
|
blockC += (ih >> 2) * cStride;
|
|
for (; ih < h; ih++) {
|
|
auto ihSubIndex = ih & 0x03;
|
|
auto c = blockC + ihSubIndex;
|
|
const float initValue = nullptr != bias ? bias[ih] : 0;
|
|
float acc0 = initValue;
|
|
|
|
const int lElement = *nnz++;
|
|
for (auto il = 0; il < lElement; il++) {
|
|
const int diff = *dataOffset++;
|
|
const float a0 = a[0];
|
|
const float oneW = *w++;
|
|
|
|
// MNN_PRINT("1-loop: ie:%zu, a offset:%ld, c offset:%ld, w offset:%ld, w value:%f, a value[0]:", ie, a
|
|
// - A, w - B - 1, c - C, oneW); formatMatrix(a, {1}); MNN_PRINT("\n");
|
|
a = a + diff;
|
|
acc0 += a0 * oneW;
|
|
}
|
|
acc0 = std::max(std::min(maxValue, acc0), minValue);
|
|
// how to store faster: st4 / transpose /
|
|
c[0] = acc0;
|
|
}
|
|
ie += 1;
|
|
// a += 1;
|
|
}
|
|
|
|
return;
|
|
}
|
|
|
|
#endif
|
|
|
|
#ifndef MNN_USE_SSE
|
|
#ifndef MNN_USE_NEON
|
|
void MNNTranspose32Bit(int32_t* dstO, const int32_t* srcO, int32_t* dim) {
|
|
int w = dim[0];
|
|
int h = dim[1];
|
|
int srcStride = dim[2];
|
|
int dstStride = dim[3];
|
|
for (int i = 0; i < h; ++i) {
|
|
auto si = srcO + i;
|
|
auto di = dstO + i * dstStride;
|
|
for (int j = 0; j < w; ++j) {
|
|
auto sj = si + j * srcStride;
|
|
auto dj = di + j;
|
|
*dj = *sj;
|
|
}
|
|
}
|
|
}
|
|
void MNNTranspose16Bit(int16_t* dstO, const int16_t* srcO, int32_t* dim) {
|
|
int w = dim[0];
|
|
int h = dim[1];
|
|
int srcStride = dim[2];
|
|
int dstStride = dim[3];
|
|
for (int i = 0; i < h; ++i) {
|
|
auto si = srcO + i;
|
|
auto di = dstO + i * dstStride;
|
|
for (int j = 0; j < w; ++j) {
|
|
auto sj = si + j * srcStride;
|
|
auto dj = di + j;
|
|
*dj = *sj;
|
|
}
|
|
}
|
|
}
|
|
#endif
|
|
void MNNFunctionInit() {
|
|
// Do nothing
|
|
}
|
|
#endif
|
|
|
|
#ifdef MNN_USE_NEON
|
|
#include <arm_neon.h>
|
|
#endif
|
|
|
|
#define UNIT 4
|
|
using Vec4 = MNN::Math::Vec<float, 4>;
|
|
|
|
#ifndef MNN_USE_NEON
|
|
|
|
#ifndef MNN_USE_SSE
|
|
|
|
void MNNCopyC4WithStride(const float* source, float* dest, size_t srcStride, size_t dstStride, size_t count) {
|
|
for (int i = 0; i < count; ++i) {
|
|
auto s = source + i * srcStride;
|
|
auto d = dest + i * dstStride;
|
|
for (int j = 0; j < 4; ++j) {
|
|
d[j] = s[j];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNAddC4WithStride(const float* source, float* dest, size_t srcStride, size_t dstStride, size_t count) {
|
|
for (int i = 0; i < count; ++i) {
|
|
auto s = source + i * srcStride;
|
|
auto d = dest + i * dstStride;
|
|
for (int j = 0; j < 4; ++j) {
|
|
d[j] += s[j];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNReluWithSlopeChannel(float* dst, const float* src, const float* slope, size_t sizeQuad, size_t depthQuad) {
|
|
for (int j = 0; j < depthQuad; j++) {
|
|
const float* slopeZ = slope + 4 * j;
|
|
const float* srcZ = src + 4 * j * sizeQuad;
|
|
float* dstZ = dst + 4 * j * sizeQuad;
|
|
for (int i = 0; i < sizeQuad; i++) {
|
|
for (int c = 0; c < 4; c++) {
|
|
if (srcZ[4 * i + c] < 0) {
|
|
dstZ[4 * i + c] = srcZ[4 * i + c] * slopeZ[c];
|
|
} else {
|
|
dstZ[4 * i + c] = srcZ[4 * i + c];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackC4(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNPackC4Common<float>(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC4(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNUnpackC4Common<float>(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
#ifndef MNN_USE_RVV
|
|
void MNNExpC8(float* dest, const float* source, float* offset, const float* parameters, size_t countC8) {
|
|
auto count = countC8 * 8;
|
|
auto param = parameters[0];
|
|
float xLimit = 87;
|
|
float summer = offset[3];
|
|
for (int i = 0; i < count; ++i) {
|
|
auto x = source[i] * offset[0] + offset[2];
|
|
x = ALIMAX(x, -xLimit);
|
|
x = ALIMIN(x, xLimit);
|
|
int div = (x * parameters[1]);
|
|
int div2 = (div + 127) << 23;
|
|
auto xReamin = x - div * param;
|
|
float expBasic = *(float*)(&div2);
|
|
auto t = xReamin * 0.25f;
|
|
auto expRemain =
|
|
((((parameters[7] * t + parameters[6]) * t + parameters[5]) * t + parameters[4]) * t + 1.0f) * t + 1.0f;
|
|
expRemain = expRemain * expRemain;
|
|
expRemain = expRemain * expRemain;
|
|
dest[i] = expBasic * expRemain + offset[1];
|
|
summer += dest[i];
|
|
}
|
|
offset[3] = summer;
|
|
}
|
|
#endif
|
|
|
|
void MNNSoftmax(float* softmaxDst, const float* softmaxSrc, float* runningMax, float* runningSum, float* updateScale,
|
|
int outside, int reduceSize, int kvSeqOffset, int validOffset, int pack, bool mask) {
|
|
// source shape: [reduceSizeOuter, outside, reduceSizeInner]
|
|
// for C4, [up_div(reduceSize,4), outside,4] => reduceSizeOuter=up_div(reduceSize,4), reduceSizeInner=4
|
|
// for C, [outside, reduceSize] => reduceSizeOuter=1, reduceSizeInner=reduceSize
|
|
|
|
const int packUnit = 4;
|
|
int reduceSizeOuter = 1;
|
|
int reduceSizeInner = reduceSize;
|
|
int stride0 = packUnit;
|
|
if (pack > 1) {
|
|
reduceSizeOuter = UP_DIV(reduceSize, pack);
|
|
reduceSizeInner = pack;
|
|
stride0 = outside * reduceSizeInner;
|
|
}
|
|
|
|
float exprOffset[4] = {1.0f, 0.0f, 0.0f, 0.0f};
|
|
for (int k = 0; k < outside; ++k) {
|
|
exprOffset[3] = 0.0f; // init sum to zero for each outer loop
|
|
if (mask && kvSeqOffset > k + validOffset) {
|
|
if (updateScale) {
|
|
updateScale[k] = 1;
|
|
}
|
|
for (int j = 0; j < reduceSizeOuter; ++j) {
|
|
int i = 0;
|
|
for (; i < reduceSizeInner; i += packUnit) {
|
|
auto destPtr = softmaxDst + j * stride0 + k * reduceSizeInner + i;
|
|
memset(destPtr, 0, packUnit * sizeof(float));
|
|
}
|
|
if (i < reduceSizeInner) {
|
|
memset(softmaxDst + j * stride0 + k * reduceSizeInner + i, 0,
|
|
(reduceSizeInner - i) * sizeof(float));
|
|
}
|
|
}
|
|
continue;
|
|
}
|
|
|
|
const int validReduceSize = mask ? ALIMIN(reduceSize, k + (validOffset + 1) - kvSeqOffset) : reduceSize;
|
|
const int remain = validReduceSize % packUnit;
|
|
const int sizeDiv = validReduceSize / packUnit;
|
|
|
|
// 1. newMax
|
|
float oldMax = std::numeric_limits<float>::lowest();
|
|
if (runningMax) {
|
|
oldMax = runningMax[k];
|
|
}
|
|
|
|
float newMax = std::numeric_limits<float>::lowest();
|
|
|
|
for (int j = 0; j < sizeDiv; ++j) {
|
|
auto srcPtr = softmaxSrc + j * stride0 + k * reduceSizeInner;
|
|
for (int i = 0; i < packUnit; ++i) {
|
|
newMax = ALIMAX(newMax, srcPtr[i]);
|
|
}
|
|
}
|
|
|
|
if (remain > 0) {
|
|
auto srcPtr = softmaxSrc + sizeDiv * stride0 + k * reduceSizeInner;
|
|
for (int i = 0; i < remain; ++i) {
|
|
newMax = ALIMAX(newMax, srcPtr[i]);
|
|
}
|
|
}
|
|
|
|
const float finalMax = ALIMAX(oldMax, newMax);
|
|
|
|
// 2. exp(x - finalMax)
|
|
exprOffset[2] = -finalMax;
|
|
|
|
for (int j = 0; j < sizeDiv; ++j) {
|
|
auto idx = j * stride0 + k * reduceSizeInner;
|
|
auto srcPtr = softmaxSrc + idx;
|
|
auto dstPtr = softmaxDst + idx;
|
|
MNNExp(dstPtr, srcPtr, exprOffset, packUnit);
|
|
}
|
|
|
|
float sum = exprOffset[3];
|
|
|
|
if (remain > 0) {
|
|
auto idx = sizeDiv * stride0 + k * reduceSizeInner;
|
|
auto srcPtr = softmaxSrc + idx;
|
|
auto dstPtr = softmaxDst + idx;
|
|
|
|
for (int i = 0; i < remain; ++i) {
|
|
float val = expf(srcPtr[i] - finalMax);
|
|
sum += val;
|
|
dstPtr[i] = val;
|
|
}
|
|
}
|
|
|
|
// 3.
|
|
if (runningMax != nullptr && runningSum != nullptr && updateScale != nullptr) {
|
|
// update runningSum, runningMax, scale
|
|
float scaleForSum = expf(oldMax - finalMax);
|
|
runningSum[k] = runningSum[k] * scaleForSum + sum;
|
|
runningMax[k] = finalMax;
|
|
updateScale[k] = scaleForSum;
|
|
} else {
|
|
// Normalization
|
|
if (runningMax != nullptr && runningSum != nullptr) {
|
|
sum += runningSum[k] * expf(oldMax - finalMax);
|
|
}
|
|
float scale = 1.0f / (sum + 1e-20f);
|
|
|
|
for (int j = 0; j < sizeDiv; ++j) {
|
|
auto pDest = softmaxDst + j * stride0 + k * reduceSizeInner;
|
|
for (int i = 0; i < packUnit; ++i) {
|
|
pDest[i] = pDest[i] * scale;
|
|
}
|
|
}
|
|
if (remain > 0) {
|
|
auto pDest = softmaxDst + sizeDiv * stride0 + k * reduceSizeInner;
|
|
for (int i = 0; i < remain; ++i) {
|
|
pDest[i] = pDest[i] * scale;
|
|
}
|
|
}
|
|
}
|
|
|
|
// 4. memset 0
|
|
if (pack > 1) {
|
|
if (validReduceSize % packUnit > 0) {
|
|
memset(softmaxDst + sizeDiv * stride0 + k * reduceSizeInner + (validReduceSize % packUnit), 0,
|
|
(packUnit - (validReduceSize % packUnit)) * sizeof(float));
|
|
}
|
|
auto validDiv4 = UP_DIV(validReduceSize, packUnit);
|
|
auto allDiv4 = UP_DIV(reduceSize, packUnit);
|
|
for (int j = validDiv4; j < allDiv4; ++j) {
|
|
auto destPtr = softmaxDst + j * stride0 + k * reduceSizeInner;
|
|
memset(destPtr, 0, packUnit * sizeof(float));
|
|
}
|
|
} else {
|
|
memset(softmaxDst + k * reduceSizeInner + validReduceSize, 0,
|
|
(reduceSize - validReduceSize) * sizeof(float));
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNReluInt8(int8_t* dst, const int8_t* src, size_t size, ssize_t zeroPoint) {
|
|
for (int i = 0; i < size; ++i) {
|
|
if (src[i] < zeroPoint) {
|
|
dst[i] = zeroPoint;
|
|
} else {
|
|
dst[i] = src[i];
|
|
}
|
|
}
|
|
}
|
|
#endif // no MNN_USE_SSE
|
|
|
|
void MNNExp(float* dst, const float* src, float* offset, size_t dataSize) {
|
|
int countC8 = static_cast<int32_t>(dataSize) / 8;
|
|
int remain = static_cast<int32_t>(dataSize) % 8;
|
|
static const float parameters[] = {
|
|
(float)logf(2.0f), 1.0f / (float)logf(2.0f), 0.25f, 1.0f, 0.5f, 1.0f / 6.0f, 1.0f / 24.0f, 1.0f / 120.0f};
|
|
if (countC8 > 0) {
|
|
// Align to eight so asm is easier to write
|
|
MNNExpC8(dst, src, offset, parameters, countC8);
|
|
}
|
|
if (remain > 0) {
|
|
auto param = parameters[0];
|
|
float xLimit = 87;
|
|
float summer = offset[3];
|
|
auto source = src + countC8 * 8;
|
|
auto dest = dst + countC8 * 8;
|
|
for (int i = 0; i < remain; ++i) {
|
|
auto x = source[i] * offset[0] + offset[2];
|
|
x = ALIMAX(x, -xLimit);
|
|
x = ALIMIN(x, xLimit);
|
|
int div = (x * parameters[1]);
|
|
int div2 = (div + 127) << 23;
|
|
auto xReamin = x - div * param;
|
|
float expBasic = *(float*)(&div2);
|
|
auto t = xReamin * 0.25f;
|
|
auto expRemain =
|
|
((((parameters[7] * t + parameters[6]) * t + parameters[5]) * t + parameters[4]) * t + 1.0f) * t + 1.0f;
|
|
expRemain = expRemain * expRemain;
|
|
expRemain = expRemain * expRemain;
|
|
dest[i] = expBasic * expRemain + offset[1];
|
|
summer += dest[i];
|
|
}
|
|
offset[3] = summer;
|
|
}
|
|
}
|
|
|
|
inline void smartCopy(void* dest, const void* src, size_t size) {
|
|
switch (size) {
|
|
case 1:
|
|
*(uint8_t*)dest = *(const uint8_t*)src;
|
|
break;
|
|
case 2:
|
|
*(uint16_t*)dest = *(const uint16_t*)src;
|
|
break;
|
|
case 4:
|
|
*(uint32_t*)dest = *(const uint32_t*)src;
|
|
break;
|
|
case 8:
|
|
*(uint64_t*)dest = *(const uint64_t*)src;
|
|
break;
|
|
default:
|
|
::memcpy(dest, src, size);
|
|
break;
|
|
}
|
|
}
|
|
|
|
void MNNPackForMatMul_A(float* dst, const float* src, size_t E, size_t L, size_t eP, size_t lP, size_t bytes) {
|
|
if (E == 0 || L == 0) {
|
|
return;
|
|
}
|
|
// [e,l] -> [e/eP,l/lP,eP,lP]
|
|
auto eU = UP_DIV(E, eP);
|
|
auto lU = UP_DIV(L, lP);
|
|
if (lP > 1) {
|
|
const int lC = L / lP;
|
|
const int lR = L % lP;
|
|
const size_t copySizeBytes = (size_t)lP * bytes;
|
|
|
|
const size_t srcStride0 = (size_t)L * bytes;
|
|
const size_t dstStride0 = (size_t)lU * eP * lP * bytes;
|
|
const size_t dstStride1 = eP * lP * bytes;
|
|
const size_t dstStride2 = lP * bytes;
|
|
|
|
for (int i = 0; i < eU; ++i) {
|
|
const int xC = ALIMIN(eP, E - i * eP);
|
|
const uint8_t* APtr = (uint8_t*)src + (i * eP) * srcStride0;
|
|
uint8_t* ADst = (uint8_t*)dst + i * dstStride0;
|
|
|
|
if (lC > 0) {
|
|
for (int x = 0; x < xC; ++x) {
|
|
auto srcBase = APtr + x * srcStride0;
|
|
auto destBase = ADst + x * dstStride2;
|
|
|
|
for (int yy = 0; yy < lC; ++yy) {
|
|
auto srcPtr = srcBase + (size_t)yy * copySizeBytes;
|
|
auto destPtr = destBase + (size_t)yy * dstStride1;
|
|
|
|
smartCopy(destPtr, srcPtr, copySizeBytes);
|
|
}
|
|
}
|
|
}
|
|
|
|
if (lR > 0) {
|
|
const int yy = lC;
|
|
const size_t remainderCopyBytes = (size_t)lR * bytes;
|
|
|
|
for (int x = 0; x < xC; ++x) {
|
|
auto srcPtr = APtr + x * srcStride0 + lC * lP * bytes;
|
|
auto destPtr = ADst + lC * dstStride1 + x * dstStride2; // (lC * eP * lP + x * lP) * bytes;
|
|
|
|
::memcpy(destPtr, srcPtr, remainderCopyBytes);
|
|
::memset(destPtr + remainderCopyBytes, 0, copySizeBytes - remainderCopyBytes);
|
|
}
|
|
}
|
|
}
|
|
} else { // lP=1
|
|
// e, l -> eU, l, eP, 1
|
|
for (int i = 0; i < eU; ++i) {
|
|
const int xC = ALIMIN(eP, E - i * eP);
|
|
auto APtr = (uint8_t*)src + (i * eP * L) * bytes;
|
|
auto ADst = (uint8_t*)dst + (i * lU * eP * lP) * bytes;
|
|
int dims[4] = {xC, (int)L, (int)L, (int)eP};
|
|
if (bytes == 2) {
|
|
auto S = (const int16_t*)APtr;
|
|
auto D = (int16_t*)ADst;
|
|
MNNTranspose16Bit(D, S, dims);
|
|
} else if (bytes == 4) {
|
|
auto S = (const int32_t*)APtr;
|
|
auto D = (int32_t*)ADst;
|
|
MNNTranspose32Bit(D, S, dims);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNMaxFloat(float* input, float* maxBuffer, int32_t inputCountUnit) {
|
|
for (int i = 0; i < inputCountUnit; i++) {
|
|
for (int j = 0; j < UNIT; j++) {
|
|
for (int m = 0; m < 2; m++) {
|
|
maxBuffer[j] = std::max(input[i * UNIT * 2 + j * 2 + m], maxBuffer[j]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
void MNNMinFloat(float* input, float* minBuffer, int32_t inputCountUnit) {
|
|
for (int i = 0; i < inputCountUnit; i++) {
|
|
for (int j = 0; j < UNIT; j++) {
|
|
for (int m = 0; m < 2; m++) {
|
|
minBuffer[j] = std::min(input[i * UNIT * 2 + j * 2 + m], minBuffer[j]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
void MNNScaleAndAddBias(float* dst, const float* src, const float* bias, const float* alpha, size_t planeNumber,
|
|
size_t biasNumber) {
|
|
for (int z = 0; z < biasNumber; ++z) {
|
|
float* dstZ = dst + planeNumber * 4 * z;
|
|
const float* srcZ = src + planeNumber * 4 * z;
|
|
auto biasZ = Vec4::load(bias + 4 * z);
|
|
auto alphaZ = Vec4::load(alpha + 4 * z);
|
|
for (int p = 0; p < planeNumber; ++p) {
|
|
float* dstX = dstZ + 4 * p;
|
|
const float* srcX = srcZ + 4 * p;
|
|
Vec4::save(dstX, (Vec4::load(srcX) * alphaZ) + biasZ);
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNUInt8ToInt16WithOffsetC4Common(int16_t* dst, const uint8_t* src, size_t zeroPoint, size_t sizeQuad,
|
|
size_t dstStride, size_t srcStride) {
|
|
dstStride /= sizeof(int16_t);
|
|
srcStride /= sizeof(uint8_t);
|
|
for (int z = 0; z < sizeQuad; ++z) {
|
|
auto dstZ = dst + dstStride * z;
|
|
auto srcZ = src + srcStride * z;
|
|
for (int j = 0; j < 4; ++j) {
|
|
dstZ[j] = (int16_t)((int32_t)srcZ[j] - (int32_t)zeroPoint);
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNUInt8ToInt16WithOffsetC4Fast(int16_t* colAddr, const uint8_t* srcStart, size_t zeroPoint, size_t sizeQuad,
|
|
size_t depthQuad, size_t dstZStep, size_t srcZStep) {
|
|
dstZStep /= sizeof(int16_t);
|
|
srcZStep /= sizeof(uint8_t);
|
|
for (int sz = 0; sz < depthQuad; ++sz) {
|
|
auto dstZ = colAddr + sz * dstZStep;
|
|
auto srcZ = srcStart + sz * srcZStep;
|
|
MNNUInt8ToInt16WithOffsetC4Common(dstZ, srcZ, zeroPoint, sizeQuad, 4 * sizeof(int16_t), 4 * sizeof(uint8_t));
|
|
}
|
|
}
|
|
|
|
void MNNPowC8(float* dest, const float* source, const float* powfParam, size_t betaInt, size_t countC8) {
|
|
const int count = countC8 * 8;
|
|
const float powfConstant = powfParam[6];
|
|
for (int i = 0; i < count; ++i) {
|
|
float result = 1, x, xInv = 1 / source[i];
|
|
for (int j = 0; j < betaInt; result *= xInv, ++j)
|
|
;
|
|
for (x = source[i]; x >= 1.25; x /= 1.5, result *= powfConstant)
|
|
;
|
|
float t = x - 1;
|
|
float powRemain =
|
|
powfParam[0] +
|
|
t * (powfParam[1] + t * (powfParam[2] + t * (powfParam[3] + t * (powfParam[4] + t * powfParam[5]))));
|
|
result *= powRemain;
|
|
dest[i] = result;
|
|
}
|
|
}
|
|
#endif // no MNN_USE_NEON
|
|
|
|
void MNNGridSampleComputeCord(float* dst, const float* src, size_t inH, size_t inW, size_t outH, size_t outW,
|
|
bool alignCorners) {
|
|
float a = alignCorners ? 1.0f : 0.0f;
|
|
float b = alignCorners ? 0.0f : 1.0f;
|
|
int area = outH * outW;
|
|
float kx = 0.5f * ((float)inW - a);
|
|
float bx = 0.5f * ((float)inW - a - b);
|
|
float ky = 0.5f * ((float)inH - a);
|
|
float by = 0.5f * ((float)inH - a - b);
|
|
for (int w = 0; w < area; ++w) {
|
|
auto x = src[2 * w + 0];
|
|
auto y = src[2 * w + 1];
|
|
dst[2 * w + 0] = kx * x + bx;
|
|
dst[2 * w + 1] = ky * y + by;
|
|
}
|
|
}
|
|
void MNNGridSampleComputeCord3D(float* dst, const float* src, size_t inD, size_t inH, size_t inW, size_t outD,
|
|
size_t outH, size_t outW, bool alignCorners) {
|
|
int strideD = outH * outW * 3;
|
|
int strideH = outW * 3;
|
|
float a = alignCorners ? 1.0f : 0.0f;
|
|
float b = alignCorners ? 0.0f : 1.0f;
|
|
int area = outD * outH * outW;
|
|
float kx = 0.5f * ((float)inW - a);
|
|
float bx = 0.5f * ((float)inW - a - b);
|
|
float ky = 0.5f * ((float)inH - a);
|
|
float by = 0.5f * ((float)inH - a - b);
|
|
float kz = 0.5f * ((float)inD - a);
|
|
float bz = 0.5f * ((float)inD - a - b);
|
|
|
|
for (int w = 0; w < area; ++w) {
|
|
auto x = src[3 * w + 0];
|
|
auto y = src[3 * w + 1];
|
|
auto z = src[3 * w + 2];
|
|
dst[3 * w + 0] = kx * x + bx;
|
|
dst[3 * w + 1] = ky * y + by;
|
|
dst[3 * w + 2] = kz * z + bz;
|
|
}
|
|
}
|
|
|
|
#ifndef MNN_USE_SSE
|
|
#ifndef MNN_USE_RVV
|
|
void MNNNorm(float* dst, const float* src, const float* gamma, const float* beta, float epsilon, size_t size,
|
|
bool RMSNorm) {
|
|
float mean = 0;
|
|
if (false == RMSNorm) {
|
|
float sum = 0.f;
|
|
MNNAccumulateSequenceNumber(&sum, src, size);
|
|
mean = sum / size;
|
|
}
|
|
#ifdef MNN_USE_NEON
|
|
const float32x4_t vmean = vdupq_n_f32(mean);
|
|
const float32x4_t veps = vdupq_n_f32(epsilon);
|
|
float32x4_t vsqsum = vdupq_n_f32(0.0f);
|
|
float32x4_t vsqsum1 = vdupq_n_f32(0.0f);
|
|
float32x4_t vsqsum2 = vdupq_n_f32(0.0f);
|
|
float32x4_t vsqsum3 = vdupq_n_f32(0.0f);
|
|
|
|
int j = 0;
|
|
// compute square sub sum
|
|
for (; j + 15 < size; j += 16) {
|
|
float32x4_t v0 = vld1q_f32(&src[j + 0]);
|
|
float32x4_t v1 = vld1q_f32(&src[j + 4]);
|
|
float32x4_t v2 = vld1q_f32(&src[j + 8]);
|
|
float32x4_t v3 = vld1q_f32(&src[j + 12]);
|
|
|
|
v0 = vsubq_f32(v0, vmean);
|
|
v1 = vsubq_f32(v1, vmean);
|
|
v2 = vsubq_f32(v2, vmean);
|
|
v3 = vsubq_f32(v3, vmean);
|
|
|
|
vsqsum = vmlaq_f32(vsqsum, v0, v0);
|
|
vsqsum1 = vmlaq_f32(vsqsum1, v1, v1);
|
|
vsqsum2 = vmlaq_f32(vsqsum2, v2, v2);
|
|
vsqsum3 = vmlaq_f32(vsqsum3, v3, v3);
|
|
}
|
|
vsqsum = vaddq_f32(vsqsum, vsqsum1);
|
|
vsqsum2 = vaddq_f32(vsqsum2, vsqsum3);
|
|
vsqsum = vaddq_f32(vsqsum, vsqsum2);
|
|
|
|
// last 0~15
|
|
for (; j + 3 < size; j += 4) {
|
|
float32x4_t v = vld1q_f32(&src[j]);
|
|
v = vsubq_f32(v, vmean);
|
|
vsqsum = vmlaq_f32(vsqsum, v, v);
|
|
}
|
|
|
|
#ifdef __aarch64__
|
|
float square_sum = vaddvq_f32(vsqsum);
|
|
#else
|
|
float square_sum = vsqsum[0] + vsqsum[1] + vsqsum[2] + vsqsum[3];
|
|
#endif
|
|
|
|
for (; j < size; ++j) {
|
|
float diff = src[j] - mean;
|
|
square_sum += diff * diff;
|
|
}
|
|
|
|
#ifdef __aarch64__
|
|
auto vs = vadd_f32(vdiv_f32(vdup_n_f32(square_sum), vdup_n_f32(size)), vdup_n_f32(epsilon));
|
|
auto vecs = vdiv_f32(vdup_n_f32(1.0f), vsqrt_f32(vs));
|
|
float vars[2];
|
|
vst1_f32(vars, vecs);
|
|
float variable = vars[0];
|
|
#else
|
|
float variance = square_sum / static_cast<float>(size);
|
|
float variable = 1.0f / std::sqrt(variance + epsilon);
|
|
#endif
|
|
|
|
const float32x4_t vvar = vdupq_n_f32(variable);
|
|
|
|
// Normalize + scale
|
|
j = 0;
|
|
if (gamma && beta) {
|
|
const float32x4_t vzero = vdupq_n_f32(0.0f);
|
|
for (; j + 15 < size; j += 16) {
|
|
float32x4_t s0 = vld1q_f32(&src[j + 0]);
|
|
float32x4_t s1 = vld1q_f32(&src[j + 4]);
|
|
float32x4_t s2 = vld1q_f32(&src[j + 8]);
|
|
float32x4_t s3 = vld1q_f32(&src[j + 12]);
|
|
|
|
float32x4_t g0 = vld1q_f32(&gamma[j + 0]);
|
|
float32x4_t g1 = vld1q_f32(&gamma[j + 4]);
|
|
float32x4_t g2 = vld1q_f32(&gamma[j + 8]);
|
|
float32x4_t g3 = vld1q_f32(&gamma[j + 12]);
|
|
|
|
float32x4_t b0 = vld1q_f32(&beta[j + 0]);
|
|
float32x4_t b1 = vld1q_f32(&beta[j + 4]);
|
|
float32x4_t b2 = vld1q_f32(&beta[j + 8]);
|
|
float32x4_t b3 = vld1q_f32(&beta[j + 12]);
|
|
|
|
s0 = vsubq_f32(s0, vmean);
|
|
s1 = vsubq_f32(s1, vmean);
|
|
s2 = vsubq_f32(s2, vmean);
|
|
s3 = vsubq_f32(s3, vmean);
|
|
|
|
s0 = vmulq_f32(s0, vvar);
|
|
s1 = vmulq_f32(s1, vvar);
|
|
s2 = vmulq_f32(s2, vvar);
|
|
s3 = vmulq_f32(s3, vvar);
|
|
|
|
s0 = vmlaq_f32(b0, s0, g0);
|
|
s1 = vmlaq_f32(b1, s1, g1);
|
|
s2 = vmlaq_f32(b2, s2, g2);
|
|
s3 = vmlaq_f32(b3, s3, g3);
|
|
|
|
vst1q_f32(&dst[j + 0], s0);
|
|
vst1q_f32(&dst[j + 4], s1);
|
|
vst1q_f32(&dst[j + 8], s2);
|
|
vst1q_f32(&dst[j + 12], s3);
|
|
}
|
|
|
|
for (; j + 3 < size; j += 4) {
|
|
float32x4_t s = vld1q_f32(&src[j]);
|
|
float32x4_t g = vld1q_f32(&gamma[j]);
|
|
float32x4_t b = vld1q_f32(&beta[j]);
|
|
|
|
s = vsubq_f32(s, vmean);
|
|
s = vmulq_f32(s, vvar);
|
|
s = vmlaq_f32(b, s, g);
|
|
|
|
vst1q_f32(&dst[j], s);
|
|
}
|
|
|
|
for (; j < size; ++j) {
|
|
dst[j] = (src[j] - mean) * variable * gamma[j] + beta[j];
|
|
}
|
|
} else {
|
|
for (; j + 15 < size; j += 16) {
|
|
float32x4_t s0 = vld1q_f32(&src[j + 0]);
|
|
float32x4_t s1 = vld1q_f32(&src[j + 4]);
|
|
float32x4_t s2 = vld1q_f32(&src[j + 8]);
|
|
float32x4_t s3 = vld1q_f32(&src[j + 12]);
|
|
|
|
s0 = vsubq_f32(s0, vmean);
|
|
s1 = vsubq_f32(s1, vmean);
|
|
s2 = vsubq_f32(s2, vmean);
|
|
s3 = vsubq_f32(s3, vmean);
|
|
|
|
s0 = vmulq_f32(s0, vvar);
|
|
s1 = vmulq_f32(s1, vvar);
|
|
s2 = vmulq_f32(s2, vvar);
|
|
s3 = vmulq_f32(s3, vvar);
|
|
|
|
vst1q_f32(&dst[j + 0], s0);
|
|
vst1q_f32(&dst[j + 4], s1);
|
|
vst1q_f32(&dst[j + 8], s2);
|
|
vst1q_f32(&dst[j + 12], s3);
|
|
}
|
|
|
|
for (; j + 3 < size; j += 4) {
|
|
float32x4_t s = vld1q_f32(&src[j]);
|
|
s = vsubq_f32(s, vmean);
|
|
s = vmulq_f32(s, vvar);
|
|
vst1q_f32(&dst[j], s);
|
|
}
|
|
|
|
for (; j < size; ++j) {
|
|
dst[j] = (src[j] - mean) * variable;
|
|
}
|
|
}
|
|
#else
|
|
float square_sum = 0.f;
|
|
for (int j = 0; j < size; ++j) {
|
|
square_sum += (src[j] - mean) * (src[j] - mean);
|
|
}
|
|
#ifdef __aarch64__
|
|
auto vs = vadd_f32(vdiv_f32(vdup_n_f32(square_sum), vdup_n_f32(size)), vdup_n_f32(epsilon));
|
|
auto vecs = vdiv_f32(vdup_n_f32(1.0f), vsqrt_f32(vs));
|
|
float vars[2];
|
|
vst1_f32(vars, vecs);
|
|
float variable = vars[0];
|
|
#else
|
|
float variable = square_sum / size;
|
|
variable = 1.f / std::sqrt(variable + epsilon);
|
|
#endif
|
|
|
|
if (gamma && beta) {
|
|
for (int j = 0; j < size; ++j) {
|
|
dst[j] = (src[j] - mean) * variable * gamma[j] + beta[j];
|
|
}
|
|
} else {
|
|
for (int j = 0; j < size; ++j) {
|
|
dst[j] = (src[j] - mean) * variable;
|
|
}
|
|
}
|
|
#endif
|
|
}
|
|
#endif // MNN_USE_RVV
|
|
#endif // MNN_USE_SSE
|
|
|
|
void MNNRoiPoolingMax(float* dst, const float* src, int hLen, int wLen, int iw) {
|
|
Vec4 max = Vec4(-FLT_MAX);
|
|
for (int h = 0; h < hLen; h++, src += iw * UNIT) {
|
|
for (int w = 0; w < wLen; w++) {
|
|
Vec4 in = Vec4::load(src + w * UNIT);
|
|
max = Vec4::max(max, in);
|
|
}
|
|
}
|
|
Vec4::save(dst, max);
|
|
}
|
|
|
|
void MNNRoiAlignMax(float* dst, const float* src, const std::vector<std::vector<int>>& vecPos,
|
|
const std::vector<std::vector<float>>& vecArea, int samplingRatioArea, int pooledHeight,
|
|
int pooledWidth) {
|
|
for (int h = 0; h < pooledHeight; ++h, dst += pooledWidth * UNIT) {
|
|
int preCalcIdx = h * pooledWidth * samplingRatioArea;
|
|
for (int w = 0; w < pooledWidth; ++w) {
|
|
Vec4 res = Vec4(-FLT_MAX);
|
|
for (int i = 0; i < samplingRatioArea; ++i) {
|
|
const std::vector<int>& pos = vecPos[preCalcIdx];
|
|
const std::vector<float>& area = vecArea[preCalcIdx];
|
|
|
|
Vec4 val0 = Vec4::load(src + pos[0] * UNIT);
|
|
Vec4 val1 = Vec4::load(src + pos[1] * UNIT);
|
|
Vec4 val2 = Vec4::load(src + pos[2] * UNIT);
|
|
Vec4 val3 = Vec4::load(src + pos[3] * UNIT);
|
|
Vec4 mla = val0 * area[0];
|
|
mla = Vec4::fma(mla, val1, area[1]);
|
|
mla = Vec4::fma(mla, val2, area[2]);
|
|
mla = Vec4::fma(mla, val3, area[3]);
|
|
res = Vec4::max(res, mla);
|
|
preCalcIdx++;
|
|
}
|
|
Vec4::save(dst + w * UNIT, res);
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNRoiAlignAvg(float* dst, const float* src, const std::vector<std::vector<int>>& vecPos,
|
|
const std::vector<std::vector<float>>& vecArea, int samplingRatioArea, int pooledHeight,
|
|
int pooledWidth) {
|
|
float invSamplingCnt = 1.f / samplingRatioArea;
|
|
for (int h = 0; h < pooledHeight; ++h, dst += pooledWidth * UNIT) {
|
|
int preCalcIdx = h * pooledWidth * samplingRatioArea;
|
|
for (int w = 0; w < pooledWidth; ++w) {
|
|
Vec4 res = Vec4(0.f);
|
|
for (int i = 0; i < samplingRatioArea; ++i) {
|
|
const std::vector<int>& pos = vecPos[preCalcIdx];
|
|
const std::vector<float>& area = vecArea[preCalcIdx];
|
|
|
|
Vec4 val0 = Vec4::load(src + pos[0] * UNIT);
|
|
Vec4 val1 = Vec4::load(src + pos[1] * UNIT);
|
|
Vec4 val2 = Vec4::load(src + pos[2] * UNIT);
|
|
Vec4 val3 = Vec4::load(src + pos[3] * UNIT);
|
|
Vec4 mla = val0 * area[0];
|
|
mla = Vec4::fma(mla, val1, area[1]);
|
|
mla = Vec4::fma(mla, val2, area[2]);
|
|
mla = Vec4::fma(mla, val3, area[3]);
|
|
res += mla;
|
|
preCalcIdx++;
|
|
}
|
|
res = res * invSamplingCnt;
|
|
Vec4::save(dst + w * UNIT, res);
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackC4Uint8(uint8_t* dst, const uint8_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNPackC4Common(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC4Uint8(uint8_t* dst, const uint8_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNUnpackC4Common(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackTransposeUint8(uint8_t* dst, const uint8_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
if (depth == 4) {
|
|
::memcpy(dst, src, area * depth * sizeof(uint8_t));
|
|
return;
|
|
}
|
|
#ifdef MNN_USE_NEON
|
|
if (depth == 3) {
|
|
uint8x16x4_t rgba;
|
|
rgba.val[3] = vdupq_n_u8(0);
|
|
int sta = 0;
|
|
int staC16 = (int)area / 16;
|
|
for (int i = 0; i < staC16; sta += 16, ++i) {
|
|
auto rgb = vld3q_u8(src + sta * 3);
|
|
rgba.val[0] = rgb.val[0];
|
|
rgba.val[1] = rgb.val[1];
|
|
rgba.val[2] = rgb.val[2];
|
|
vst4q_u8(dst + 4 * sta, rgba);
|
|
}
|
|
sta = staC16 * 16;
|
|
|
|
for (; sta < area; ++sta) {
|
|
auto s = src + sta * 3;
|
|
auto d = dst + sta * 4;
|
|
d[0] = s[0];
|
|
d[1] = s[1];
|
|
d[2] = s[2];
|
|
d[3] = 0;
|
|
}
|
|
|
|
return;
|
|
}
|
|
if (depth == 1) {
|
|
uint8x16x4_t rgba;
|
|
rgba.val[1] = vdupq_n_u8(0);
|
|
rgba.val[2] = vdupq_n_u8(0);
|
|
rgba.val[3] = vdupq_n_u8(0);
|
|
int sta = 0;
|
|
for (; sta < area; sta += 16) {
|
|
rgba.val[0] = vld1q_u8(src + sta);
|
|
vst4q_u8(dst + 4 * sta, rgba);
|
|
}
|
|
|
|
for (; sta < area; ++sta) {
|
|
auto s = src + sta;
|
|
auto d = dst + sta * 4;
|
|
d[0] = s[0];
|
|
d[1] = 0;
|
|
d[2] = 0;
|
|
d[3] = 0;
|
|
}
|
|
|
|
return;
|
|
}
|
|
#endif
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
|
|
if (cAlign == c) {
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = reinterpret_cast<const int32_t*>(src + hi * c);
|
|
auto dstHeight = reinterpret_cast<int32_t*>(dst + hi * 4);
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
dstHeight[ci * areaOffset[1]] = srcHeight[ci];
|
|
}
|
|
}
|
|
return;
|
|
} else {
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = src + hi * c;
|
|
auto dstHeight = dst + hi * 4;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
dstHeight[ci * areaOffset[1] * 4 + 0] = srcHeight[ci * 4 + 0];
|
|
dstHeight[ci * areaOffset[1] * 4 + 1] = srcHeight[ci * 4 + 1];
|
|
dstHeight[ci * areaOffset[1] * 4 + 2] = srcHeight[ci * 4 + 2];
|
|
dstHeight[ci * areaOffset[1] * 4 + 3] = srcHeight[ci * 4 + 3];
|
|
}
|
|
}
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + cAlign;
|
|
auto dstAlign = dst + areaOffset[1] * cAlign;
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = srcAlign + hi * c;
|
|
auto dstHeight = dstAlign + hi * 4;
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[i] = 0;
|
|
}
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNUnpackTranspose(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
int srcAreaOffset = areaOffset[0];
|
|
int dstAreaOffset = areaOffset[1];
|
|
#ifdef MNN_USE_NEON
|
|
if (1 == depth) {
|
|
auto zeroValue = vmovq_n_f32(0.0f);
|
|
int areaC4 = (int)area / 4;
|
|
int remain = areaC4 * 4;
|
|
for (int i = 0; i < areaC4; ++i) {
|
|
auto srcCur = src + 4 * i;
|
|
auto dstCur = dst + 16 * i;
|
|
auto srcValue = vld1q_f32(srcCur);
|
|
float32x4x4_t dstValue;
|
|
dstValue.val[0] = srcValue;
|
|
dstValue.val[1] = zeroValue;
|
|
dstValue.val[2] = zeroValue;
|
|
dstValue.val[3] = zeroValue;
|
|
vst4q_f32(dstCur, dstValue);
|
|
}
|
|
for (int i = remain; i < area; ++i) {
|
|
dst[4 * i + 0] = src[i];
|
|
dst[4 * i + 1] = 0.0f;
|
|
dst[4 * i + 2] = 0.0f;
|
|
dst[4 * i + 3] = 0.0f;
|
|
}
|
|
return;
|
|
}
|
|
if (3 == depth) {
|
|
auto zeroValue = vmovq_n_f32(0.0f);
|
|
int areaC4 = (int)area / 4;
|
|
int remain = areaC4 * 4;
|
|
for (int i = 0; i < areaC4; ++i) {
|
|
auto srcCur = src + 12 * i;
|
|
auto dstCur = dst + 16 * i;
|
|
auto srcValue = vld3q_f32(srcCur);
|
|
float32x4x4_t dstValue;
|
|
dstValue.val[0] = srcValue.val[0];
|
|
dstValue.val[1] = srcValue.val[1];
|
|
dstValue.val[2] = srcValue.val[2];
|
|
dstValue.val[3] = zeroValue;
|
|
vst4q_f32(dstCur, dstValue);
|
|
}
|
|
for (int i = remain; i < area; ++i) {
|
|
dst[4 * i + 0] = src[3 * i + 0];
|
|
dst[4 * i + 1] = src[3 * i + 1];
|
|
dst[4 * i + 2] = src[3 * i + 2];
|
|
dst[4 * i + 3] = 0.0f;
|
|
}
|
|
return;
|
|
}
|
|
#endif
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
const float* srcHeight = src + hi * c;
|
|
float* dstHeight = dst + hi * 4;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
Vec4::save(dstHeight + 4 * ci * dstAreaOffset, Vec4::load(srcHeight + 4 * ci));
|
|
}
|
|
}
|
|
|
|
if (cAlign == c) {
|
|
return;
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + cAlign;
|
|
auto dstAlign = dst + dstAreaOffset * cAlign;
|
|
|
|
#ifdef MNN_USE_NEON
|
|
auto zeroVector = vdupq_n_f32(0.0f);
|
|
#endif
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
const float* srcHeight = srcAlign + hi * c;
|
|
float* dstHeight = dstAlign + hi * 4;
|
|
#ifdef MNN_USE_NEON
|
|
vst1q_f32(dstHeight, zeroVector);
|
|
#else
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[i] = 0;
|
|
}
|
|
#endif
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackTransposeUint8(uint8_t* dst, const uint8_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
if (cAlign == c) {
|
|
int32_t* dst32 = (int32_t*)dst;
|
|
const int32_t* src32 = (int32_t*)src;
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = src32 + hi;
|
|
auto dstHeight = dst32 + hi * cDiv4;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci * areaOffset[0]];
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = src + hi * 4;
|
|
auto dstHeight = dst + hi * c;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[ci * 4 + i] = srcHeight[4 * ci * areaOffset[0] + i];
|
|
}
|
|
}
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + areaOffset[0] * cAlign;
|
|
auto dstAlign = dst + cAlign;
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = srcAlign + hi * 4;
|
|
auto dstHeight = dstAlign + hi * c;
|
|
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNPackTranspose(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
#if defined(MNN_USE_NEON)
|
|
if (3 == depth) {
|
|
int areaC4 = (int)area / 4;
|
|
int remain = areaC4 * 4;
|
|
for (int i = 0; i < areaC4; ++i) {
|
|
auto srcCur = src + 16 * i;
|
|
auto dstCur = dst + 12 * i;
|
|
auto srcValue = vld4q_f32(srcCur);
|
|
float32x4x3_t dstValue;
|
|
dstValue.val[0] = srcValue.val[0];
|
|
dstValue.val[1] = srcValue.val[1];
|
|
dstValue.val[2] = srcValue.val[2];
|
|
vst3q_f32(dstCur, dstValue);
|
|
}
|
|
for (int i = remain; i < area; ++i) {
|
|
dst[3 * i + 0] = src[4 * i + 0];
|
|
dst[3 * i + 1] = src[4 * i + 1];
|
|
dst[3 * i + 2] = src[4 * i + 2];
|
|
}
|
|
return;
|
|
}
|
|
#elif defined(MNN_USE_SSE)
|
|
if (3 == depth) {
|
|
if (area < 1)
|
|
return;
|
|
for (int i = 0; i < area - 1; ++i) {
|
|
auto srcValue = Vec4::load(src + 4 * i);
|
|
Vec4::save(dst + 3 * i, srcValue);
|
|
}
|
|
for (int i = 0; i < 3; ++i) {
|
|
dst[3 * (area - 1) + i] = src[4 * (area - 1) + i];
|
|
}
|
|
return;
|
|
}
|
|
#endif
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
auto srcArea = areaOffset[0];
|
|
auto dstDepthOffset = areaOffset[1];
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
const float* srcHeight = src + hi * 4;
|
|
float* dstHeight = dst + hi * dstDepthOffset;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
Vec4::save(dstHeight + 4 * ci, Vec4::load(srcHeight + 4 * ci * srcArea));
|
|
}
|
|
}
|
|
|
|
if (cAlign == c) {
|
|
return;
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + srcArea * cAlign;
|
|
auto dstAlign = dst + cAlign;
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
const float* srcHeight = srcAlign + hi * 4;
|
|
float* dstHeight = dstAlign + hi * dstDepthOffset;
|
|
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
|
|
// Lambert's series with 7 divisions
|
|
// reference from
|
|
// https://varietyofsound.wordpress.com/2011/02/14/efficient-tanh-computation-using-lamberts-continued-fraction/
|
|
inline float tanhf_poly(float value) {
|
|
if (value > 5.0) {
|
|
return 1.0;
|
|
} else if (value <= -5.0) {
|
|
return -1.0;
|
|
} else {
|
|
float x2 = value * value;
|
|
float a = value * (135135.0f + x2 * (17325.0f + x2 * (378.0f + x2)));
|
|
float b = 135135.0f + x2 * (62370.0f + x2 * (3150.0f + x2 * 28.0f));
|
|
return a / b;
|
|
}
|
|
}
|
|
void MNNTanh(float* dst, const float* src, size_t dataSize) {
|
|
/* Origin Code
|
|
for (int i = 0; i < dataSize; i++) {
|
|
// outputData[i] = 1 - 2 / (expf(2 * inputData[i]) + 1);
|
|
dst[i] = tanhf_poly(src[i]);
|
|
}
|
|
*/
|
|
float offset[4] = {-2.0f, 0.0f, 0.0f, 0.0f};
|
|
MNNExp(dst, src, offset, dataSize);
|
|
for (int i = 0; i < dataSize; i++) {
|
|
// outputData[i] = 1 - 2 / (expf(2 * inputData[i]) + 1);
|
|
auto expX2 = dst[i];
|
|
dst[i] = (1.0f - expX2) / (1.0f + expX2);
|
|
}
|
|
}
|
|
|
|
void MNNReluWithSlope(float* dst, const float* src, size_t sizeQuad, float slope) {
|
|
float slopeValue[4];
|
|
for (int i = 0; i < 4; ++i) {
|
|
slopeValue[i] = slope;
|
|
}
|
|
MNNReluWithSlopeChannel(dst, src, slopeValue, sizeQuad, 1);
|
|
}
|
|
|
|
void MNNReluWithSlopeCommon(float* dst, const float* src, size_t size, float slope) {
|
|
int sizeQuad = static_cast<int32_t>(size) / 4;
|
|
int remain = static_cast<int32_t>(size) % 4;
|
|
if (sizeQuad > 0) {
|
|
MNNReluWithSlope(dst, src, sizeQuad, slope);
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[4] = {0}, outmp[4] = {0};
|
|
::memcpy(intmp, src + sizeQuad * 4, remain * sizeof(float));
|
|
MNNReluWithSlope(outmp, intmp, 1, slope);
|
|
::memcpy(dst + sizeQuad * 4, outmp, remain * sizeof(float));
|
|
}
|
|
}
|
|
|
|
void MNNHardSwishCommon(float* dst, const float* src, size_t size) {
|
|
int sizeQuad = static_cast<int32_t>(size / 4);
|
|
int remain = static_cast<int32_t>(size) % 4;
|
|
#ifdef MNN_USE_SSE
|
|
if (sizeQuad > 0) {
|
|
MNNHardSwish(dst, src, sizeQuad);
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[4] = {0}, outmp[4] = {0};
|
|
::memcpy(intmp, src + sizeQuad * 4, remain * sizeof(float));
|
|
MNNHardSwish(outmp, intmp, 1);
|
|
::memcpy(dst + sizeQuad * 4, outmp, remain * sizeof(float));
|
|
}
|
|
#else
|
|
#ifdef MNN_USE_NEON
|
|
float32x4_t zero = vdupq_n_f32(0.f);
|
|
float32x4_t three = vdupq_n_f32(3.f);
|
|
float32x4_t six = vdupq_n_f32(6.f);
|
|
float32x4_t divsix = vdupq_n_f32(1.0f / 6.f);
|
|
for (int i = 0; i < sizeQuad; i++) {
|
|
auto x = vld1q_f32(src + 4 * i);
|
|
auto y = vmulq_f32(vmulq_f32(x, vminq_f32(vmaxq_f32(vaddq_f32(x, three), zero), six)), divsix);
|
|
vst1q_f32(dst + 4 * i, y);
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[4] = {0}, outmp[4] = {0};
|
|
::memcpy(intmp, src + sizeQuad * 4, remain * sizeof(float));
|
|
auto x = vld1q_f32(intmp);
|
|
auto y = vmulq_f32(vmulq_f32(x, vminq_f32(vmaxq_f32(vaddq_f32(x, three), zero), six)), divsix);
|
|
vst1q_f32(outmp, y);
|
|
::memcpy(dst + sizeQuad * 4, outmp, remain * sizeof(float));
|
|
}
|
|
#else
|
|
for (int j = 0; j < size; j++) {
|
|
if (src[j] <= -3) {
|
|
dst[j] = 0;
|
|
} else if (src[j] >= 3) {
|
|
dst[j] = src[j];
|
|
} else {
|
|
dst[j] = src[j] * (src[j] + 3) / 6.f;
|
|
}
|
|
}
|
|
#endif
|
|
#endif
|
|
}
|
|
|
|
#ifndef MNN_USE_RVV
|
|
void MNNGeluStandardCommon(float* dst, const float* src, size_t size) {
|
|
for (int i = 0; i < size; i++) {
|
|
dst[i] = (erf(src[i] * 0.7071067932881648) + 1) * src[i] * 0.5;
|
|
}
|
|
}
|
|
|
|
void MNNGeluCommon(float* dst, const float* src, size_t size) {
|
|
int sizeQuad = static_cast<int32_t>(size / 8);
|
|
int remain = static_cast<int32_t>(size) % 8;
|
|
#if defined(MNN_USE_SSE) || defined(MNN_USE_NEON)
|
|
float parameters[8] = {0.044715f, 0.79788458f, 378.f, 17325.f, 135135.f, 28.f, 3150.f, 62370.f};
|
|
if (sizeQuad > 0) {
|
|
MNNGelu(dst, src, sizeQuad, parameters);
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[8] = {0};
|
|
float outmp[8] = {0};
|
|
::memcpy(intmp, src + 8 * sizeQuad, remain * sizeof(float));
|
|
MNNGelu(outmp, intmp, 1, parameters);
|
|
::memcpy(dst + 8 * sizeQuad, outmp, remain * sizeof(float));
|
|
}
|
|
#else
|
|
auto tanhf_poly = [](float value) -> float {
|
|
if (value > 5.0f) {
|
|
return 1.0f;
|
|
} else if (value <= -5.0f) {
|
|
return -1.0f;
|
|
} else {
|
|
float x2 = value * value;
|
|
float a = value * (135135.0f + x2 * (17325.0f + x2 * (378.0f + x2)));
|
|
float b = 135135.0f + x2 * (62370.0f + x2 * (3150.0f + x2 * 28.0f));
|
|
return a / b;
|
|
}
|
|
};
|
|
for (int i = 0; i < size; i++) {
|
|
float temp = 0.044715f * src[i] * src[i] * src[i];
|
|
temp = 0.79788458f * (temp + src[i]);
|
|
dst[i] = (1.0f + tanhf_poly(temp)) * src[i] * 0.5f;
|
|
}
|
|
#endif
|
|
}
|
|
#endif
|
|
|
|
void MNNScaleAndAddBiasScalar(float* dst, const float* src, float bias, float alpha, size_t number) {
|
|
int numberC4 = (int)number / 4;
|
|
int start = 0;
|
|
if (numberC4 > 0) {
|
|
float biasC4[4] = {bias, bias, bias, bias};
|
|
float alphaC4[4] = {alpha, alpha, alpha, alpha};
|
|
MNNScaleAndAddBias(dst, src, biasC4, alphaC4, numberC4, 1);
|
|
start = numberC4 * 4;
|
|
}
|
|
for (int i = start; i < number; ++i) {
|
|
dst[i] = src[i] * alpha + bias;
|
|
}
|
|
}
|
|
#ifndef MNN_USE_NEON
|
|
void MNNAxByClampBroadcastUnit(float* C, const float* A, const float* B, size_t width, size_t cStride, size_t aStride,
|
|
size_t height, const float* parameters) {
|
|
auto minF = Vec4(parameters[2]);
|
|
auto maxF = Vec4(parameters[3]);
|
|
auto beta = Vec4(parameters[1]);
|
|
for (int y = 0; y < height; ++y) {
|
|
auto a = A + aStride * y;
|
|
auto b = B + 4 * y;
|
|
auto bv = Vec4::load(b);
|
|
auto c = C + cStride * y;
|
|
for (int x = 0; x < width; ++x) {
|
|
auto av = Vec4::load(a + 4 * x);
|
|
auto cv = av + bv * beta;
|
|
cv = Vec4::min(cv, maxF);
|
|
cv = Vec4::max(cv, minF);
|
|
Vec4::save(c + 4 * x, cv);
|
|
}
|
|
}
|
|
}
|
|
void MNNVectorTop1Float(float* input, float* maxValue, int32_t* maxIndex, size_t inputCountUnit) {
|
|
float maxV = input[0];
|
|
int maxIdx = 0;
|
|
for (int i = 0; i < inputCountUnit; i++) {
|
|
int offset = i * UNIT;
|
|
for (int j = 0; j < UNIT; j++) {
|
|
if (input[offset + j] > maxV) {
|
|
maxV = input[offset + j];
|
|
maxIdx = offset + j;
|
|
}
|
|
}
|
|
}
|
|
maxValue[0] = maxV;
|
|
maxIndex[0] = maxIdx;
|
|
}
|
|
|
|
void MNNVectorTop1Int32(int32_t* input, int32_t* maxValue, int32_t* maxIndex, size_t inputCountUnit) {
|
|
int32_t maxV = input[0];
|
|
int maxIdx = 0;
|
|
for (int i = 0; i < inputCountUnit; i++) {
|
|
int offset = i * UNIT;
|
|
for (int j = 0; j < UNIT; j++) {
|
|
if (input[offset + j] > maxV) {
|
|
maxV = input[offset + j];
|
|
maxIdx = offset + j;
|
|
}
|
|
}
|
|
}
|
|
maxValue[0] = maxV;
|
|
maxIndex[0] = maxIdx;
|
|
}
|
|
|
|
#endif
|
|
|
|
#ifndef __aarch64__
|
|
static void MNNRankOneUpdateDefault(float* S, const float* k, const float* delta, size_t dk, size_t dv) {
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float k_val = k[i];
|
|
float* row = S + i * dv;
|
|
for (size_t j = 0; j < dv; ++j) {
|
|
row[j] += k_val * delta[j];
|
|
}
|
|
}
|
|
}
|
|
// Read-only dual MatVec: out_k = S^T @ k, out_q = S^T @ q
|
|
static void MNNDualMatVecDefault(const float* S, const float* k, const float* q, float* out_k, float* out_q, size_t dk,
|
|
size_t dv) {
|
|
::memset(out_k, 0, dv * sizeof(float));
|
|
::memset(out_q, 0, dv * sizeof(float));
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float k_val = k[i];
|
|
float q_val = q[i];
|
|
const float* row = S + i * dv;
|
|
for (size_t j = 0; j < dv; ++j) {
|
|
out_k[j] += row[j] * k_val;
|
|
out_q[j] += row[j] * q_val;
|
|
}
|
|
}
|
|
}
|
|
// Fused decay + rank-1 update: S[i,j] = decay * S[i,j] + k[i] * delta[j]
|
|
static void MNNDecayRankOneUpdateDefault(float* S, const float* k, const float* delta, float decay, size_t dk,
|
|
size_t dv) {
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float k_val = k[i];
|
|
float* row = S + i * dv;
|
|
for (size_t j = 0; j < dv; ++j) {
|
|
row[j] = decay * row[j] + k_val * delta[j];
|
|
}
|
|
}
|
|
}
|
|
#else
|
|
extern "C" {
|
|
void MNNRankOneUpdateDefault(float* S, const float* k, const float* delta, size_t dk, size_t dv);
|
|
void MNNDualMatVecDefault(const float* S, const float* k, const float* q, float* out_k, float* out_q, size_t dk,
|
|
size_t dv);
|
|
void MNNDecayRankOneUpdateDefault(float* S, const float* k, const float* delta, float decay, size_t dk, size_t dv);
|
|
}
|
|
#endif
|
|
|
|
// ─────────────────────────────────────────────────────────────────────────
|
|
// MNNFusedGatedDelta — fused gated_delta_rule recurrence step.
|
|
//
|
|
// Processes S column-wise in chunks of `kChunk` (=16 elements for fp32,
|
|
// four v.4s lanes). For each chunk j..j+kChunk-1:
|
|
// Pass 1: stream rows [0,d_k) and accumulate out_k, out_q in registers.
|
|
// Inline correction: still in registers, compute
|
|
// delta = beta * (v - decay * out_k)
|
|
// out = decay * out_q + kq * delta
|
|
// Store `out`; keep delta resident.
|
|
// Pass 2: stream rows [0,d_k) again and update S in-place using the
|
|
// in-register delta.
|
|
//
|
|
// Requires d_v to be a multiple of 16 in fp32 (Qwen3-Next-style heads use
|
|
// d_v ∈ {64, 128, 256}). A scalar tail covers the remainder defensively.
|
|
// ─────────────────────────────────────────────────────────────────────────
|
|
static void MNNFusedGatedDeltaDefault(float* S, const float* k, const float* q, const float* v, float* out, float decay,
|
|
float beta, float kq, size_t dk, size_t dv) {
|
|
#if defined(__aarch64__) && defined(MNN_USE_NEON)
|
|
// FP32 chunk = 16 elements (4 v.4s registers per accumulator).
|
|
// The inner loop is unrolled by 4 rows so a single vld1q_f32 of
|
|
// (k[i], k[i+1], k[i+2], k[i+3]) feeds 4 vfmaq_laneq_f32 ops via
|
|
// .s[lane], amortizing the scalar broadcast across 4 row iterations.
|
|
const size_t kChunk = 16;
|
|
const float32x4_t vDecay = vdupq_n_f32(decay);
|
|
const float32x4_t vBeta = vdupq_n_f32(beta);
|
|
const float32x4_t vKq = vdupq_n_f32(kq);
|
|
|
|
size_t j = 0;
|
|
for (; j + kChunk <= dv; j += kChunk) {
|
|
// ── Pass 1: out_k = S^T @ k, out_q = S^T @ q for this column chunk ──
|
|
float32x4_t ok0 = vdupq_n_f32(0.0f), ok1 = vdupq_n_f32(0.0f), ok2 = vdupq_n_f32(0.0f), ok3 = vdupq_n_f32(0.0f);
|
|
float32x4_t oq0 = vdupq_n_f32(0.0f), oq1 = vdupq_n_f32(0.0f), oq2 = vdupq_n_f32(0.0f), oq3 = vdupq_n_f32(0.0f);
|
|
|
|
size_t i = 0;
|
|
for (; i + 4 <= dk; i += 4) {
|
|
float32x4_t kVec = vld1q_f32(k + i);
|
|
float32x4_t qVec = vld1q_f32(q + i);
|
|
#define LANE_STEP_FP32(lane) \
|
|
{ \
|
|
const float* row = S + (i + (lane)) * dv + j; \
|
|
float32x4_t s0 = vld1q_f32(row); \
|
|
float32x4_t s1 = vld1q_f32(row + 4); \
|
|
float32x4_t s2 = vld1q_f32(row + 8); \
|
|
float32x4_t s3 = vld1q_f32(row + 12); \
|
|
ok0 = vfmaq_laneq_f32(ok0, s0, kVec, (lane)); \
|
|
ok1 = vfmaq_laneq_f32(ok1, s1, kVec, (lane)); \
|
|
ok2 = vfmaq_laneq_f32(ok2, s2, kVec, (lane)); \
|
|
ok3 = vfmaq_laneq_f32(ok3, s3, kVec, (lane)); \
|
|
oq0 = vfmaq_laneq_f32(oq0, s0, qVec, (lane)); \
|
|
oq1 = vfmaq_laneq_f32(oq1, s1, qVec, (lane)); \
|
|
oq2 = vfmaq_laneq_f32(oq2, s2, qVec, (lane)); \
|
|
oq3 = vfmaq_laneq_f32(oq3, s3, qVec, (lane)); \
|
|
}
|
|
LANE_STEP_FP32(0);
|
|
LANE_STEP_FP32(1);
|
|
LANE_STEP_FP32(2);
|
|
LANE_STEP_FP32(3);
|
|
#undef LANE_STEP_FP32
|
|
}
|
|
// Tail rows (dk % 4) — fall back to scalar broadcast form.
|
|
for (; i < dk; ++i) {
|
|
const float* row = S + i * dv + j;
|
|
float32x4_t s0 = vld1q_f32(row);
|
|
float32x4_t s1 = vld1q_f32(row + 4);
|
|
float32x4_t s2 = vld1q_f32(row + 8);
|
|
float32x4_t s3 = vld1q_f32(row + 12);
|
|
ok0 = vfmaq_n_f32(ok0, s0, k[i]);
|
|
ok1 = vfmaq_n_f32(ok1, s1, k[i]);
|
|
ok2 = vfmaq_n_f32(ok2, s2, k[i]);
|
|
ok3 = vfmaq_n_f32(ok3, s3, k[i]);
|
|
oq0 = vfmaq_n_f32(oq0, s0, q[i]);
|
|
oq1 = vfmaq_n_f32(oq1, s1, q[i]);
|
|
oq2 = vfmaq_n_f32(oq2, s2, q[i]);
|
|
oq3 = vfmaq_n_f32(oq3, s3, q[i]);
|
|
}
|
|
|
|
// ── Inline analytic correction (regs only) ──
|
|
float32x4_t v0 = vld1q_f32(v + j);
|
|
float32x4_t v1 = vld1q_f32(v + j + 4);
|
|
float32x4_t v2 = vld1q_f32(v + j + 8);
|
|
float32x4_t v3 = vld1q_f32(v + j + 12);
|
|
// delta = beta * (v - decay * out_k)
|
|
float32x4_t d0 = vmulq_f32(vBeta, vsubq_f32(v0, vmulq_f32(vDecay, ok0)));
|
|
float32x4_t d1 = vmulq_f32(vBeta, vsubq_f32(v1, vmulq_f32(vDecay, ok1)));
|
|
float32x4_t d2 = vmulq_f32(vBeta, vsubq_f32(v2, vmulq_f32(vDecay, ok2)));
|
|
float32x4_t d3 = vmulq_f32(vBeta, vsubq_f32(v3, vmulq_f32(vDecay, ok3)));
|
|
// out = decay * out_q + kq * delta
|
|
float32x4_t o0 = vfmaq_f32(vmulq_f32(vDecay, oq0), vKq, d0);
|
|
float32x4_t o1 = vfmaq_f32(vmulq_f32(vDecay, oq1), vKq, d1);
|
|
float32x4_t o2 = vfmaq_f32(vmulq_f32(vDecay, oq2), vKq, d2);
|
|
float32x4_t o3 = vfmaq_f32(vmulq_f32(vDecay, oq3), vKq, d3);
|
|
vst1q_f32(out + j, o0);
|
|
vst1q_f32(out + j + 4, o1);
|
|
vst1q_f32(out + j + 8, o2);
|
|
vst1q_f32(out + j + 12, o3);
|
|
|
|
// ── Pass 2: S = decay * S + k ⊗ delta (delta d0..d3 still in regs) ──
|
|
size_t i2 = 0;
|
|
for (; i2 + 4 <= dk; i2 += 4) {
|
|
float32x4_t kVec = vld1q_f32(k + i2);
|
|
#define ROW_UPDATE_FP32(lane) \
|
|
{ \
|
|
float* row = S + (i2 + (lane)) * dv + j; \
|
|
float32x4_t s0 = vld1q_f32(row); \
|
|
float32x4_t s1 = vld1q_f32(row + 4); \
|
|
float32x4_t s2 = vld1q_f32(row + 8); \
|
|
float32x4_t s3 = vld1q_f32(row + 12); \
|
|
float32x4_t r0 = vfmaq_laneq_f32(vmulq_f32(vDecay, s0), d0, kVec, (lane)); \
|
|
float32x4_t r1 = vfmaq_laneq_f32(vmulq_f32(vDecay, s1), d1, kVec, (lane)); \
|
|
float32x4_t r2 = vfmaq_laneq_f32(vmulq_f32(vDecay, s2), d2, kVec, (lane)); \
|
|
float32x4_t r3 = vfmaq_laneq_f32(vmulq_f32(vDecay, s3), d3, kVec, (lane)); \
|
|
vst1q_f32(row, r0); \
|
|
vst1q_f32(row + 4, r1); \
|
|
vst1q_f32(row + 8, r2); \
|
|
vst1q_f32(row + 12, r3); \
|
|
}
|
|
ROW_UPDATE_FP32(0);
|
|
ROW_UPDATE_FP32(1);
|
|
ROW_UPDATE_FP32(2);
|
|
ROW_UPDATE_FP32(3);
|
|
#undef ROW_UPDATE_FP32
|
|
}
|
|
for (; i2 < dk; ++i2) {
|
|
float* row = S + i2 * dv + j;
|
|
float32x4_t s0 = vld1q_f32(row);
|
|
float32x4_t s1 = vld1q_f32(row + 4);
|
|
float32x4_t s2 = vld1q_f32(row + 8);
|
|
float32x4_t s3 = vld1q_f32(row + 12);
|
|
float32x4_t r0 = vfmaq_n_f32(vmulq_f32(vDecay, s0), d0, k[i2]);
|
|
float32x4_t r1 = vfmaq_n_f32(vmulq_f32(vDecay, s1), d1, k[i2]);
|
|
float32x4_t r2 = vfmaq_n_f32(vmulq_f32(vDecay, s2), d2, k[i2]);
|
|
float32x4_t r3 = vfmaq_n_f32(vmulq_f32(vDecay, s3), d3, k[i2]);
|
|
vst1q_f32(row, r0);
|
|
vst1q_f32(row + 4, r1);
|
|
vst1q_f32(row + 8, r2);
|
|
vst1q_f32(row + 12, r3);
|
|
}
|
|
}
|
|
// Scalar tail (guards d_v not divisible by 16 — defensive only)
|
|
for (; j < dv; ++j) {
|
|
float ok = 0.0f, oq = 0.0f;
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float s = S[i * dv + j];
|
|
ok += s * k[i];
|
|
oq += s * q[i];
|
|
}
|
|
float delta_j = beta * (v[j] - decay * ok);
|
|
out[j] = decay * oq + kq * delta_j;
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
S[i * dv + j] = decay * S[i * dv + j] + k[i] * delta_j;
|
|
}
|
|
}
|
|
#else
|
|
// Pure scalar fallback (non-aarch64 / no NEON): same math, no SIMD.
|
|
// We need delta cached because Pass 2 uses it after Pass 1+correction.
|
|
std::vector<float> deltaBuf(dv);
|
|
for (size_t j = 0; j < dv; ++j) {
|
|
float ok = 0.0f, oq = 0.0f;
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float s = S[i * dv + j];
|
|
ok += s * k[i];
|
|
oq += s * q[i];
|
|
}
|
|
float delta_j = beta * (v[j] - decay * ok);
|
|
deltaBuf[j] = delta_j;
|
|
out[j] = decay * oq + kq * delta_j;
|
|
}
|
|
for (size_t i = 0; i < dk; ++i) {
|
|
float k_val = k[i];
|
|
float* row = S + i * dv;
|
|
for (size_t j = 0; j < dv; ++j) {
|
|
row[j] = decay * row[j] + k_val * deltaBuf[j];
|
|
}
|
|
}
|
|
#endif
|
|
}
|
|
|
|
void MNNComputeMatMulForE_1(const float* A, const float* B, float* C, const float* biasPtr, const MatMulParam* param,
|
|
size_t tIdL) {
|
|
auto l = param->l;
|
|
auto h = param->h;
|
|
auto numberThread = param->numberThread;
|
|
auto lC4 = l / 4;
|
|
auto lR = lC4 * 4;
|
|
auto tId = (int)tIdL;
|
|
if (param->BTranspose) {
|
|
for (int y = tId; y < h; y += numberThread) {
|
|
Vec4 sumValue = Vec4(0.0f);
|
|
auto by = B + y * l;
|
|
for (int x = 0; x < lC4; ++x) {
|
|
sumValue = Vec4::fma(sumValue, Vec4::load(A + x * 4), Vec4::load(by + x * 4));
|
|
}
|
|
float sumRemain = 0.0f;
|
|
for (int x = lR; x < l; ++x) {
|
|
sumRemain = sumRemain + A[x] * by[x];
|
|
}
|
|
if (nullptr != biasPtr) {
|
|
sumRemain += biasPtr[y];
|
|
}
|
|
C[y] = sumRemain + sumValue[0] + sumValue[1] + sumValue[2] + sumValue[3];
|
|
}
|
|
} else {
|
|
auto hC4 = h / 16;
|
|
auto hR = hC4 * 16;
|
|
for (int y = tId; y < hC4; y += numberThread) {
|
|
auto bs = B + 16 * y;
|
|
Vec4 sumValue0;
|
|
Vec4 sumValue1;
|
|
Vec4 sumValue2;
|
|
Vec4 sumValue3;
|
|
if (biasPtr != nullptr) {
|
|
sumValue0 = Vec4::load(biasPtr + 16 * y + 0);
|
|
sumValue1 = Vec4::load(biasPtr + 16 * y + 4);
|
|
sumValue2 = Vec4::load(biasPtr + 16 * y + 8);
|
|
sumValue3 = Vec4::load(biasPtr + 16 * y + 12);
|
|
} else {
|
|
sumValue0 = Vec4(0.0f);
|
|
sumValue1 = Vec4(0.0f);
|
|
sumValue2 = Vec4(0.0f);
|
|
sumValue3 = Vec4(0.0f);
|
|
}
|
|
auto srcY = A + y * l;
|
|
for (int x = 0; x < l; ++x) {
|
|
auto a = Vec4(A[x]);
|
|
sumValue0 = Vec4::fma(sumValue0, a, Vec4::load(bs + h * x));
|
|
sumValue1 = Vec4::fma(sumValue1, a, Vec4::load(bs + h * x + 4));
|
|
sumValue2 = Vec4::fma(sumValue2, a, Vec4::load(bs + h * x + 8));
|
|
sumValue3 = Vec4::fma(sumValue3, a, Vec4::load(bs + h * x + 12));
|
|
}
|
|
Vec4::save(C + 16 * y, sumValue0);
|
|
Vec4::save(C + 16 * y + 4, sumValue1);
|
|
Vec4::save(C + 16 * y + 8, sumValue2);
|
|
Vec4::save(C + 16 * y + 12, sumValue3);
|
|
}
|
|
int hEnd = hR;
|
|
if ((h - hR) >= 8) {
|
|
if (0 == tId) {
|
|
auto bs = B + hEnd;
|
|
Vec4 sumValue0;
|
|
Vec4 sumValue1;
|
|
if (biasPtr != nullptr) {
|
|
sumValue0 = Vec4::load(biasPtr + hEnd + 0);
|
|
sumValue1 = Vec4::load(biasPtr + hEnd + 4);
|
|
} else {
|
|
sumValue0 = Vec4(0.0f);
|
|
sumValue1 = Vec4(0.0f);
|
|
}
|
|
auto srcY = A + hEnd * l;
|
|
for (int x = 0; x < l; ++x) {
|
|
auto a = Vec4(A[x]);
|
|
sumValue0 = Vec4::fma(sumValue0, a, Vec4::load(bs + h * x));
|
|
sumValue1 = Vec4::fma(sumValue1, a, Vec4::load(bs + h * x + 4));
|
|
}
|
|
Vec4::save(C + hEnd, sumValue0);
|
|
Vec4::save(C + hEnd + 4, sumValue1);
|
|
}
|
|
hEnd = hEnd + 8;
|
|
}
|
|
if ((h - hEnd) >= 4) {
|
|
if (0 == tId) {
|
|
auto bs = B + hEnd;
|
|
Vec4 sumValue0;
|
|
if (biasPtr != nullptr) {
|
|
sumValue0 = Vec4::load(biasPtr + hEnd + 0);
|
|
} else {
|
|
sumValue0 = Vec4(0.0f);
|
|
}
|
|
auto srcY = A + hEnd * l;
|
|
for (int x = 0; x < l; ++x) {
|
|
auto a = Vec4(A[x]);
|
|
sumValue0 = Vec4::fma(sumValue0, a, Vec4::load(bs + h * x));
|
|
}
|
|
Vec4::save(C + hEnd, sumValue0);
|
|
}
|
|
hEnd = hEnd + 4;
|
|
}
|
|
hEnd = hEnd + tId;
|
|
for (int y = hEnd; y < h; y += numberThread) {
|
|
auto bs = B + y;
|
|
float sumValue = 0.0f;
|
|
if (biasPtr != nullptr) {
|
|
sumValue = biasPtr[y];
|
|
}
|
|
auto srcY = A + y * l;
|
|
for (int x = 0; x < l; ++x) {
|
|
sumValue = sumValue + A[x] * bs[h * x];
|
|
}
|
|
C[y] = sumValue;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNComputeMatMulForH_1(const float* A, const float* B, float* C, const float* biasPtr, const MatMulParam* param,
|
|
size_t tId) {
|
|
int e = param->e;
|
|
int l = param->l;
|
|
int numberThread = param->numberThread;
|
|
if (param->ATranspose) {
|
|
float biasValue = 0.0f;
|
|
if (nullptr != biasPtr) {
|
|
biasValue = *biasPtr;
|
|
}
|
|
auto eC4 = e / 4;
|
|
auto eR = eC4 * 4;
|
|
for (int y = tId; y < eC4; y += numberThread) {
|
|
Vec4 sumValue = Vec4(biasValue);
|
|
auto srcY = A + y * 4;
|
|
for (int x = 0; x < l; ++x) {
|
|
sumValue = sumValue + Vec4::load(srcY + x * e) * Vec4(B[x]);
|
|
}
|
|
Vec4::save(C + 4 * y, sumValue);
|
|
}
|
|
if (0 == tId) {
|
|
for (int y = eR; y < e; ++y) {
|
|
float sumValue = biasValue;
|
|
auto srcY = A + y;
|
|
for (int x = 0; x < l; ++x) {
|
|
sumValue = sumValue + srcY[x * e] * B[x];
|
|
}
|
|
C[y] = sumValue;
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
float biasValue = 0.0f;
|
|
if (nullptr != biasPtr) {
|
|
biasValue = *biasPtr;
|
|
}
|
|
auto lC4 = l / 16;
|
|
auto lRO = lC4 * 16;
|
|
for (int y = tId; y < e; y += numberThread) {
|
|
auto lR = lRO;
|
|
Vec4 sumValue = Vec4(biasValue);
|
|
Vec4 sum1(0.0f);
|
|
Vec4 sum2(0.0f);
|
|
Vec4 sum3(0.0f);
|
|
auto srcY = A + y * l;
|
|
for (int x = 0; x < lC4; ++x) {
|
|
sumValue = Vec::fma(sumValue, Vec4::load(srcY + 16 * x + 0), Vec4::load(B + 16 * x + 0));
|
|
sum1 = Vec::fma(sum1, Vec4::load(srcY + 16 * x + 4), Vec4::load(B + 16 * x + 4));
|
|
sum2 = Vec::fma(sum2, Vec4::load(srcY + 16 * x + 8), Vec4::load(B + 16 * x + 8));
|
|
sum3 = Vec::fma(sum3, Vec4::load(srcY + 16 * x + 12), Vec4::load(B + 16 * x + 12));
|
|
}
|
|
if (l - lR >= 8) {
|
|
sumValue = Vec::fma(sumValue, Vec4::load(srcY + lR), Vec4::load(B + lR));
|
|
sum1 = Vec::fma(sum1, Vec4::load(srcY + lR + 4), Vec4::load(B + lR + 4));
|
|
lR += 8;
|
|
}
|
|
if (l - lR >= 4) {
|
|
sumValue = Vec::fma(sumValue, Vec4::load(srcY + lR), Vec4::load(B + lR));
|
|
lR += 4;
|
|
}
|
|
sum2 = sum2 + sum3;
|
|
sumValue = sumValue + sum1;
|
|
sumValue = sumValue + sum2;
|
|
float sumSingle = sumValue[0] + sumValue[1] + sumValue[2] + sumValue[3];
|
|
for (int x = lR; x < l; ++x) {
|
|
sumSingle += srcY[x] * B[x];
|
|
}
|
|
C[y] = sumSingle;
|
|
}
|
|
}
|
|
|
|
void MNNPackC4Int16(int16_t* dst, const int16_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNPackC4Common(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC4Int16(int16_t* dst, const int16_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNUnpackC4Common(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackTransposeInt16(int16_t* dst, const int16_t* src, size_t area, size_t depth, int* areaOffset) {
|
|
if (depth == 4) {
|
|
::memcpy(dst, src, area * depth * sizeof(int16_t));
|
|
return;
|
|
}
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = (src + hi * c);
|
|
auto dstHeight = (dst + hi * 4);
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[ci * areaOffset[1] * 4 + i] = srcHeight[4 * ci + i];
|
|
}
|
|
}
|
|
}
|
|
|
|
if (cAlign == c) {
|
|
return;
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + cAlign;
|
|
auto dstAlign = dst + areaOffset[1] * cAlign;
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = srcAlign + hi * c;
|
|
auto dstHeight = dstAlign + hi * 4;
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[i] = 0;
|
|
}
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
void MNNPackTransposeInt16(int16_t* dst, const int16_t* src, size_t area, size_t depth, int* offset) {
|
|
int c = (int)depth;
|
|
int cDiv4 = c / 4;
|
|
int cAlign = cDiv4 * 4;
|
|
int srcAreaOffset = offset[0];
|
|
int dstDepthOffset = offset[1];
|
|
if (cAlign == c) {
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = (int64_t*)src + hi;
|
|
auto dstHeight = (int64_t*)(dst + hi * dstDepthOffset);
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci * srcAreaOffset];
|
|
}
|
|
}
|
|
return;
|
|
}
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = src + hi * 4;
|
|
auto dstHeight = dst + hi * dstDepthOffset;
|
|
for (int ci = 0; ci < cDiv4; ++ci) {
|
|
for (int i = 0; i < 4; ++i) {
|
|
dstHeight[ci * 4 + i] = srcHeight[4 * ci * srcAreaOffset + i];
|
|
}
|
|
}
|
|
}
|
|
|
|
int cReamin = c - cAlign;
|
|
auto srcAlign = src + srcAreaOffset * cAlign;
|
|
auto dstAlign = dst + cAlign;
|
|
|
|
for (int hi = 0; hi < area; ++hi) {
|
|
auto srcHeight = srcAlign + hi * 4;
|
|
auto dstHeight = dstAlign + hi * dstDepthOffset;
|
|
|
|
for (int ci = 0; ci < cReamin; ++ci) {
|
|
dstHeight[ci] = srcHeight[ci];
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNCopyC4Int16WithStride(const float* sourceF, float* destF, size_t srcStride, size_t dstStride, size_t count) {
|
|
auto source = (int16_t*)sourceF;
|
|
auto dest = (int16_t*)destF;
|
|
for (int i = 0; i < count; ++i) {
|
|
auto s = source + i * srcStride;
|
|
auto d = dest + i * dstStride;
|
|
*(int64_t*)(d) = *((int64_t*)s);
|
|
}
|
|
}
|
|
|
|
void MNNSin(float* dst, const float* src, size_t dataSize) {
|
|
for (int i = 0; i < dataSize; i++) {
|
|
dst[i] = sinf(src[i]);
|
|
}
|
|
}
|
|
|
|
void MNNSigmoid(float* dst, const float* src, size_t dataSize) {
|
|
float offset[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
|
|
MNNExp(dst, src, offset, dataSize);
|
|
for (int i = 0; i < dataSize; ++i) {
|
|
dst[i] = 1.0f / (1.0f + dst[i]);
|
|
}
|
|
}
|
|
|
|
#ifndef MNN_USE_RVV
|
|
void MNNSiLu(float* dst, const float* src, size_t dataSize) {
|
|
float offset[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
|
|
MNNExp(dst, src, offset, dataSize);
|
|
for (int i = 0; i < dataSize; ++i) {
|
|
dst[i] = src[i] / (1.0f + dst[i]);
|
|
}
|
|
}
|
|
#endif
|
|
|
|
/**
|
|
Modified from https://github.com/alibaba/MNN/pull/1359
|
|
Thanks for https://github.com/hroken
|
|
*/
|
|
void MNNSigmoidLowp(float* dst, const float* src, size_t dataSize) {
|
|
float offset[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
|
|
MNNExp(dst, src, offset, dataSize);
|
|
#ifdef MNN_USE_NEON
|
|
int dataC4 = static_cast<int32_t>(dataSize) / 4;
|
|
int remain = static_cast<int32_t>(dataSize) % 4;
|
|
float32x4_t value = vdupq_n_f32(1.0f);
|
|
|
|
if (dataC4 > 0) {
|
|
float32x4_t out = vld1q_f32(dst);
|
|
// neon optimization for sigmid cpu
|
|
for (int i = 1; i < dataC4; ++i) {
|
|
out = vrecpeq_f32(vaddq_f32(value, out));
|
|
vst1q_f32(dst, out);
|
|
dst += 4;
|
|
out = vld1q_f32(dst);
|
|
}
|
|
out = vrecpeq_f32(vaddq_f32(value, out));
|
|
vst1q_f32(dst, out);
|
|
dst += 4;
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[4] = {0};
|
|
::memcpy(intmp, dst, remain * sizeof(float));
|
|
float32x4_t out = vld1q_f32(intmp);
|
|
out = vrecpeq_f32(vaddq_f32(value, out));
|
|
vst1q_f32(intmp, out);
|
|
::memcpy(dst, intmp, remain * sizeof(float));
|
|
}
|
|
#else
|
|
for (int i = 0; i < dataSize; ++i) {
|
|
dst[i] = 1.0f / (1.0f + dst[i]);
|
|
}
|
|
#endif
|
|
}
|
|
|
|
#ifndef MNN_USE_RVV
|
|
void MNNSiLuLowp(float* dst, const float* src, size_t dataSize) {
|
|
float offset[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
|
|
MNNExp(dst, src, offset, dataSize);
|
|
#ifdef __aarch64__
|
|
int dataC4 = static_cast<int32_t>(dataSize) / 4;
|
|
int remain = static_cast<int32_t>(dataSize) % 4;
|
|
float32x4_t one = vdupq_n_f32(1.0f);
|
|
|
|
if (dataC4 > 0) {
|
|
float32x4_t out = vld1q_f32(dst);
|
|
float32x4_t in = vld1q_f32(src);
|
|
// neon optimization for sigmid cpu
|
|
for (int i = 1; i < dataC4; ++i) {
|
|
out = vdivq_f32(in, vaddq_f32(one, out));
|
|
vst1q_f32(dst, out);
|
|
dst += 4;
|
|
src += 4;
|
|
out = vld1q_f32(dst);
|
|
in = vld1q_f32(src);
|
|
}
|
|
out = vdivq_f32(in, vaddq_f32(one, out));
|
|
vst1q_f32(dst, out);
|
|
dst += 4;
|
|
src += 4;
|
|
}
|
|
if (remain > 0) {
|
|
float intmp[4] = {0};
|
|
float atmp[4] = {0};
|
|
::memcpy(intmp, dst, remain * sizeof(float));
|
|
::memcpy(atmp, src, remain * sizeof(float));
|
|
float32x4_t out = vld1q_f32(intmp);
|
|
float32x4_t in = vld1q_f32(atmp);
|
|
out = vdivq_f32(in, vaddq_f32(one, out));
|
|
vst1q_f32(intmp, out);
|
|
::memcpy(dst, intmp, remain * sizeof(float));
|
|
}
|
|
#else
|
|
for (int i = 0; i < dataSize; ++i) {
|
|
dst[i] = src[i] / (1.0f + dst[i]);
|
|
}
|
|
#endif
|
|
}
|
|
#endif
|
|
|
|
static void _MNNAdjustOptimalSparseKernel(int& sparseBlockOC,
|
|
MNN::CoreFunctions::MNNPackedSparseMatMul& packedSparseMatMul) {
|
|
if (sparseBlockOC == 4) {
|
|
packedSparseMatMul = MNNPackedSparseMatMulEpx4;
|
|
return;
|
|
} else if (sparseBlockOC % 4 == 0) {
|
|
sparseBlockOC = 4;
|
|
packedSparseMatMul = MNNPackedSparseMatMulEpx4;
|
|
// MNN_PRINT("common downgrade sparse to:%d\n",sparseBlockOC);
|
|
return;
|
|
} else {
|
|
sparseBlockOC = 1;
|
|
packedSparseMatMul = MNNPackedSparseMatMulEpx1;
|
|
return;
|
|
}
|
|
}
|
|
|
|
#ifdef MNN_LOW_MEMORY
|
|
static void generalIm2col(float* destOrigin, float const** sourceGroup, const int32_t* info, const int32_t* el, int LP,
|
|
int pack) {
|
|
// LP >= pack
|
|
int number = info[0];
|
|
int eReal = info[1];
|
|
int eDest = info[2];
|
|
int offset = info[3];
|
|
for (int n = 0; n < number; ++n) {
|
|
int e = el[4 * n + 0];
|
|
int l = el[4 * n + 1];
|
|
int eOffset = el[4 * n + 2];
|
|
int lOffset = el[4 * n + 3];
|
|
int lC = lOffset / LP;
|
|
int lR = lOffset % LP;
|
|
auto dest = destOrigin + eOffset * LP + lC * eDest * LP + lR;
|
|
auto source = sourceGroup[n];
|
|
|
|
for (int y = 0; y < e; ++y) {
|
|
auto yR = y % eDest;
|
|
for (int x = 0; x < l; ++x) {
|
|
auto xR = x % pack;
|
|
auto xC = x / pack;
|
|
auto xOut = x / LP;
|
|
auto xIn = x % LP;
|
|
dest[xOut * eDest * LP + yR * LP + xIn] = source[xC * eReal * pack + y * pack * offset + xR];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
#endif // MNN_LOW_MEMORY
|
|
|
|
#ifdef MNN_SME2
|
|
#define SME2_MATMUL_EP 16
|
|
#define SME2_MATMUL_LP 1
|
|
#define SME2_MATMUL_HP 64
|
|
|
|
static void SME2MNNGetMatMulPackMode(int* eP, int* lP, int* hP) {
|
|
*eP = SME2_MATMUL_EP;
|
|
*lP = SME2_MATMUL_LP;
|
|
*hP = SME2_MATMUL_HP;
|
|
}
|
|
static void MNNPackedMatMulFP32_SME2(float* C, const float* A, const float* B, const size_t* parameter,
|
|
const float* postParameters, const float* bias, const float* k, const float* b) {
|
|
MNNPackedMatMulRemainFP32_SME2(C, A, B, 16, parameter, postParameters, bias, k, b);
|
|
return;
|
|
}
|
|
static void Sme2MNNPackForMatMul_B(float* destC, const float* sourceC, size_t h, size_t kernelsize, size_t ic,
|
|
bool transpose) {
|
|
// src: [h, kernelsize, ic]
|
|
// dst: [h/hp, kernelsize, ic/lp, hp, lp]
|
|
auto dest = (int32_t*)destC;
|
|
auto source = (int32_t*)sourceC;
|
|
int LP = SME2_MATMUL_LP;
|
|
int HP = SME2_MATMUL_HP;
|
|
auto l = kernelsize * ic;
|
|
memset(dest, 0, ROUND_UP(h, HP) * ROUND_UP(ic, LP) * kernelsize * 4);
|
|
auto stride0 = kernelsize * ROUND_UP(ic, LP) * HP;
|
|
auto stride1 = ROUND_UP(ic, LP) * HP;
|
|
auto stride2 = HP * LP;
|
|
|
|
auto srcStride0 = l; // [h,l]->[hu,lu,hp,lp]
|
|
auto srcStride1 = 1;
|
|
if (!transpose) { // [l,h]->[hu,lu,hp,lp]
|
|
srcStride0 = 1;
|
|
srcStride1 = h;
|
|
}
|
|
for (int y = 0; y < h; ++y) {
|
|
auto yHu = y / HP;
|
|
auto yHp = y % HP;
|
|
for (int k = 0; k < kernelsize; ++k) {
|
|
for (int x = 0; x < ic; ++x) {
|
|
auto xLu = x / LP;
|
|
auto xLp = x % LP;
|
|
dest[yHu * stride0 + k * stride1 + xLu * stride2 + yHp * LP + xLp] =
|
|
source[y * srcStride0 + (x + k * ic) * srcStride1];
|
|
}
|
|
}
|
|
}
|
|
}
|
|
static void Sme2MNNPackC4ForMatMul_A(float* destOrigin, float const** sourceGroup, const int32_t* info,
|
|
const int32_t* el) {
|
|
MNNPackC4ForMatMul_A(destOrigin, sourceGroup, info, el);
|
|
return;
|
|
}
|
|
#endif
|
|
|
|
namespace MNN {
|
|
|
|
static CoreFunctions* gCoreFunction = nullptr;
|
|
|
|
static void MNNRoPEComputeBasic(void* dst, const void* src, const void* cosEven, const void* cosOdd,
|
|
const void* sinEven, const void* sinOdd, int numHead, int headDim, int ropeCutHeadDim) {
|
|
const int halfHeadDim = headDim / 2;
|
|
int ropeDim = ropeCutHeadDim;
|
|
if (ropeDim <= 0 || ropeDim > headDim) {
|
|
ropeDim = headDim;
|
|
}
|
|
ropeDim = (ropeDim / 2) * 2;
|
|
const int ropeHalfHeadDim = ropeDim / 2;
|
|
|
|
auto srcFloat = static_cast<const float*>(src);
|
|
auto dstFloat = static_cast<float*>(dst);
|
|
auto cosEvenFloat = static_cast<const float*>(cosEven);
|
|
auto cosOddFloat = static_cast<const float*>(cosOdd);
|
|
auto sinEvenFloat = static_cast<const float*>(sinEven);
|
|
auto sinOddFloat = static_cast<const float*>(sinOdd);
|
|
for (int j = 0; j < numHead; ++j) {
|
|
auto src0 = srcFloat + j * headDim;
|
|
auto src1 = src0 + halfHeadDim;
|
|
auto dst0 = dstFloat + j * headDim;
|
|
auto dst1 = dst0 + halfHeadDim;
|
|
int k = 0;
|
|
for (; k <= ropeHalfHeadDim - 4; k += 4) {
|
|
auto q0 = Vec4::load(src0 + k);
|
|
auto q1 = Vec4::load(src1 + k);
|
|
auto c0 = Vec4::load(cosEvenFloat + k);
|
|
auto c1 = Vec4::load(cosOddFloat + k);
|
|
auto s0 = Vec4::load(sinEvenFloat + k);
|
|
auto s1 = Vec4::load(sinOddFloat + k);
|
|
Vec4::save(dst0 + k, Vec4::fms(q0 * c0, q1, s0));
|
|
Vec4::save(dst1 + k, Vec4::fma(q1 * c1, q0, s1));
|
|
}
|
|
for (; k < ropeHalfHeadDim; ++k) {
|
|
auto q0 = src0[k];
|
|
auto q1 = src1[k];
|
|
dst0[k] = q0 * cosEvenFloat[k] - q1 * sinEvenFloat[k];
|
|
dst1[k] = q1 * cosOddFloat[k] + q0 * sinOddFloat[k];
|
|
}
|
|
if (ropeHalfHeadDim < halfHeadDim) {
|
|
::memcpy(dst0 + ropeHalfHeadDim, src0 + ropeHalfHeadDim, (halfHeadDim - ropeHalfHeadDim) * sizeof(float));
|
|
::memcpy(dst1 + ropeHalfHeadDim, src1 + ropeHalfHeadDim, (halfHeadDim - ropeHalfHeadDim) * sizeof(float));
|
|
}
|
|
}
|
|
}
|
|
|
|
template <int Pack>
|
|
static void MNNNormPackedFloat(float* dest, const float* source, const float* gamma, const float* beta, float epsilon,
|
|
size_t batch, size_t channels, bool RMSNorm) {
|
|
const size_t channelUnit = UP_DIV(channels, Pack);
|
|
for (size_t n = 0; n < batch; ++n) {
|
|
float mean = 0.0f;
|
|
if (!RMSNorm) {
|
|
float sum = 0.0f;
|
|
for (size_t c = 0; c < channels; ++c) {
|
|
const size_t cu = c / Pack;
|
|
const size_t cr = c - cu * Pack;
|
|
sum += source[(cu * batch + n) * Pack + cr];
|
|
}
|
|
mean = sum / static_cast<float>(channels);
|
|
}
|
|
|
|
float squareSum = 0.0f;
|
|
for (size_t c = 0; c < channels; ++c) {
|
|
const size_t cu = c / Pack;
|
|
const size_t cr = c - cu * Pack;
|
|
float v = source[(cu * batch + n) * Pack + cr];
|
|
float d = RMSNorm ? v : (v - mean);
|
|
squareSum += d * d;
|
|
}
|
|
|
|
const float invStd = 1.0f / std::sqrt(squareSum / static_cast<float>(channels) + epsilon);
|
|
for (size_t c = 0; c < channels; ++c) {
|
|
const size_t cu = c / Pack;
|
|
const size_t cr = c - cu * Pack;
|
|
const size_t index = (cu * batch + n) * Pack + cr;
|
|
float v = source[index];
|
|
float norm = RMSNorm ? (v * invStd) : ((v - mean) * invStd);
|
|
if (gamma && beta) {
|
|
norm = norm * gamma[c] + beta[c];
|
|
}
|
|
dest[index] = norm;
|
|
}
|
|
for (size_t c = channels; c < channelUnit * Pack; ++c) {
|
|
const size_t cu = c / Pack;
|
|
const size_t cr = c - cu * Pack;
|
|
dest[(cu * batch + n) * Pack + cr] = 0.0f;
|
|
}
|
|
}
|
|
}
|
|
|
|
void MNNCoreFunctionInit() {
|
|
gCoreFunction = new CoreFunctions;
|
|
|
|
// MatMul
|
|
gCoreFunction->MNNGetMatMulPackMode = MNNGetMatMulPackMode;
|
|
gCoreFunction->MNNPackC4ForMatMul_A = MNNPackC4ForMatMul_A;
|
|
gCoreFunction->MNNPackForMatMul_B = MNNPackForMatMul_B;
|
|
gCoreFunction->MNNPackedMatMul = MNNPackedMatMul;
|
|
gCoreFunction->MNNPackedMatMulRemain = MNNPackedMatMulRemain;
|
|
gCoreFunction->MNNCountMaxMinValue = MNNCountMaxMinValue;
|
|
gCoreFunction->MNNNormPacked = MNNNormPackedFloat<4>;
|
|
#ifdef MNN_USE_SPARSE_COMPUTE
|
|
gCoreFunction->MNNGetSparseMatMulPackMode = MNNGetSparseMatMulPackMode;
|
|
gCoreFunction->MNNAdjustOptimalSparseKernel = _MNNAdjustOptimalSparseKernel;
|
|
#endif
|
|
|
|
gCoreFunction->MNNComputeMatMulForE_1 = MNNComputeMatMulForE_1;
|
|
gCoreFunction->MNNComputeMatMulForH_1 = MNNComputeMatMulForH_1;
|
|
gCoreFunction->MNNRankOneUpdate = MNNRankOneUpdateDefault;
|
|
gCoreFunction->MNNDualMatVec = MNNDualMatVecDefault;
|
|
gCoreFunction->MNNDecayRankOneUpdate = MNNDecayRankOneUpdateDefault;
|
|
gCoreFunction->MNNFusedGatedDelta = MNNFusedGatedDeltaDefault;
|
|
|
|
// Lowp
|
|
gCoreFunction->MNNFp32ToLowp = nullptr;
|
|
gCoreFunction->MNNLowpToFp32 = nullptr;
|
|
gCoreFunction->bytes = 4; // sizeof(float)
|
|
|
|
// Packed Function
|
|
gCoreFunction->pack = 4;
|
|
// FIXME: MNNPackTranspose and MNNUnpackTranspose is reverted
|
|
gCoreFunction->MNNPackCUnit = MNNPackC4;
|
|
gCoreFunction->MNNUnpackCUnit = MNNUnpackC4;
|
|
gCoreFunction->MNNUnpackCUnitTranspose = MNNPackTranspose;
|
|
gCoreFunction->MNNPackCUnitTranspose = MNNUnpackTranspose;
|
|
gCoreFunction->MNNPackCUnitInt8 = decltype(gCoreFunction->MNNPackCUnitInt8)(MNNPackC4Uint8);
|
|
gCoreFunction->MNNUnpackCUnitInt8 = decltype(gCoreFunction->MNNUnpackCUnitInt8)(MNNUnpackC4Uint8);
|
|
gCoreFunction->MNNPackCUnitTransposeInt8 =
|
|
decltype(gCoreFunction->MNNPackCUnitTransposeInt8)(MNNUnpackTransposeUint8);
|
|
gCoreFunction->MNNUnpackCUnitTransposeInt8 =
|
|
decltype(gCoreFunction->MNNUnpackCUnitTransposeInt8)(MNNPackTransposeUint8);
|
|
gCoreFunction->MNNPackCUnitInt16 = MNNPackC4Int16;
|
|
gCoreFunction->MNNUnpackCUnitInt16 = MNNUnpackC4Int16;
|
|
gCoreFunction->MNNPackCUnitTransposeInt16 = MNNUnpackTransposeInt16;
|
|
gCoreFunction->MNNUnpackCUnitTransposeInt16 = MNNPackTransposeInt16;
|
|
|
|
gCoreFunction->MNNAxByClampBroadcastUnit = MNNAxByClampBroadcastUnit;
|
|
gCoreFunction->MNNConvRunForLineDepthwise = MNNConvRunForLineDepthwise;
|
|
gCoreFunction->MNNMatrixAdd = MNNMatrixAdd;
|
|
gCoreFunction->MNNMatrixSub = MNNMatrixSub;
|
|
gCoreFunction->MNNStrassenMergeCFunction = MNNStrassenMergeCFunction;
|
|
gCoreFunction->penalty = 1.5f;
|
|
gCoreFunction->MNNScaleAndAddBias = MNNScaleAndAddBias;
|
|
gCoreFunction->MNNGridSampleComputeCord = MNNGridSampleComputeCord;
|
|
gCoreFunction->MNNGridSampleInterp = MNNGridSampleInterp;
|
|
#ifndef MNN_REDUCE_SIZE
|
|
gCoreFunction->MNNGridSampleInterpGrad = MNNGridSampleInterpGrad;
|
|
#endif
|
|
gCoreFunction->MNNGridSampleComputeCord3D = MNNGridSampleComputeCord3D;
|
|
gCoreFunction->MNNGridSampleInterp3D = MNNGridSampleInterp3D;
|
|
gCoreFunction->MNNRoiPoolingMax = MNNRoiPoolingMax;
|
|
gCoreFunction->MNNRoiAlignMax = MNNRoiAlignMax;
|
|
gCoreFunction->MNNRoiAlignAvg = MNNRoiAlignAvg;
|
|
gCoreFunction->MNNAddC4WithStride = MNNAddC4WithStride;
|
|
gCoreFunction->MNNCopyC4WithStride = MNNCopyC4WithStride;
|
|
|
|
gCoreFunction->chooseWinoSourceTransformPack = WinogradFunction::chooseWinoSourceTransformPack;
|
|
gCoreFunction->chooseWinoSourceUnrollTransform = WinogradFunction::chooseSourceUnrollTransform;
|
|
gCoreFunction->chooseWinoDestUnrollTransform = WinogradFunction::chooseWinoDestUnrollTransform;
|
|
gCoreFunction->MNNDeconvRunForLineDepthwise = MNNDeconvRunForLineDepthwise;
|
|
gCoreFunction->MNNDeconvRunForUnitDepthWise = MNNDeconvRunForUnitDepthWise;
|
|
gCoreFunction->MNNSoftmax = MNNSoftmax;
|
|
#ifdef MNN_USE_NEON
|
|
gCoreFunction->MNNDepthwiseConvFastKernel = MNNDepthwiseConvFastKernel;
|
|
#endif
|
|
gCoreFunction->MNNSelectBinaryFunctionForFloat = CPUBinary::selectForFloat;
|
|
gCoreFunction->MNNSelectUnaryFunctionForFloat = CPUUnary::selectForFloat;
|
|
#ifdef MNN_SUPPORT_QUANT_EXTEND
|
|
gCoreFunction->MNNSelectUnaryFunctionForInt8 = CPUUnary::selectForInt8;
|
|
#endif
|
|
|
|
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
|
|
gCoreFunction->MNNAttenPackAndScaleSingleHead = MNNAttenPackAndScaleSingleHead;
|
|
gCoreFunction->MNNFlashAttentionUpdateBlockOutput = MNNFlashAttentionUpdateBlockOutput;
|
|
gCoreFunction->MNNQuantAttentionKey = MNNQuantAttentionKey;
|
|
gCoreFunction->MNNQuantAttentionValue = MNNQuantAttentionValue;
|
|
#endif // MNN_SUPPORT_TRANSFORMER_FUSE
|
|
gCoreFunction->MNNRoPECompute = MNNRoPEComputeBasic;
|
|
|
|
gCoreFunction->MNNReluWithSlopeChannel = MNNReluWithSlopeChannel;
|
|
gCoreFunction->MNNPoolingAvg = (decltype(gCoreFunction->MNNPoolingAvg))(poolingAvg<float, Vec4, 4>);
|
|
// Set min value as 1 << 24
|
|
gCoreFunction->MNNPoolingMax = (decltype(gCoreFunction->MNNPoolingMax))(poolingMax<float, Vec4, 4, -16777216>);
|
|
|
|
gCoreFunction->MNNPoolingMaxWithRedice =
|
|
(decltype(gCoreFunction->MNNPoolingMaxWithRedice))(poolingMaxWithRedice<float, -16777216>);
|
|
// ImageProcess Functions
|
|
gCoreFunction->MNNRGBAToBGRA = MNNRGBAToBGRA;
|
|
gCoreFunction->MNNNV21ToRGBA = MNNNV21ToRGBA;
|
|
gCoreFunction->MNNNV21ToRGB = MNNNV21ToRGB;
|
|
gCoreFunction->MNNNV21ToBGRA = MNNNV21ToBGRA;
|
|
gCoreFunction->MNNNV21ToBGR = MNNNV21ToBGR;
|
|
gCoreFunction->MNNC1ToFloatC1 = MNNC1ToFloatC1;
|
|
gCoreFunction->MNNC3ToFloatC3 = MNNC3ToFloatC3;
|
|
gCoreFunction->MNNC3ToFloatRGBA = MNNC3ToFloatRGBA;
|
|
gCoreFunction->MNNSamplerC4Nearest = MNNSamplerC4Nearest;
|
|
gCoreFunction->MNNSamplerC4Bilinear = MNNSamplerC4Bilinear;
|
|
|
|
gCoreFunction->MNN4BitcopyWithStride = MNN4BitcopyWithStride;
|
|
gCoreFunction->MNN1BitcopyWithStride = MNN1BitcopyWithStride;
|
|
gCoreFunction->MNN2BitcopyWithStride = MNN2BitcopyWithStride;
|
|
gCoreFunction->MNN4BitcopyFast = MNN4BitcopyFast;
|
|
gCoreFunction->MNN2BitcopyFast = MNN2BitcopyFast;
|
|
gCoreFunction->MNN1BitcopyFast = MNN1BitCopyFast;
|
|
|
|
gCoreFunction->MNNAccumulateSequenceNumber = MNNAccumulateSequenceNumber;
|
|
|
|
const MNNCPUInfo& gCPUInfo = *MNNGetCPUInfo();
|
|
gCoreFunction->supportFp16arith = gCPUInfo.fp16arith;
|
|
gCoreFunction->supportSDot = gCPUInfo.dot;
|
|
gCoreFunction->supportI8mm = gCPUInfo.i8mm;
|
|
gCoreFunction->supportSME2 = gCPUInfo.sme2;
|
|
// add rvv support
|
|
gCoreFunction->supportRVV = gCPUInfo.rvv;
|
|
|
|
gCoreFunction->smeCoreNumber = gCPUInfo.smeCoreNumber;
|
|
#ifdef MNN_PIPELINE_PROFILE
|
|
if (const char* cpuTarget = std::getenv("MNN_CPU_TARGET")) {
|
|
int target = ::atoi(cpuTarget);
|
|
target = std::max(0, std::min(target, 3));
|
|
gCoreFunction->supportFp16arith = gCoreFunction->supportFp16arith && target >= 1;
|
|
gCoreFunction->supportSDot = gCoreFunction->supportSDot && target >= 1;
|
|
gCoreFunction->supportI8mm = gCoreFunction->supportI8mm && target >= 2;
|
|
gCoreFunction->supportSME2 = gCoreFunction->supportSME2 && target >= 3;
|
|
if (!gCoreFunction->supportSME2) {
|
|
gCoreFunction->smeCoreNumber = 0;
|
|
}
|
|
MNN_PRINT("MNN_CPU_TARGET=%d effective ARM features: fp16=%d, i8sdot=%d, i8mm=%d, sme2=%d\n", target,
|
|
gCoreFunction->supportFp16arith, gCoreFunction->supportSDot, gCoreFunction->supportI8mm,
|
|
gCoreFunction->supportSME2);
|
|
}
|
|
#endif
|
|
gCoreFunction->MNNSumByAxisLForMatmul_A = MNNSumByAxisLForMatmul_A;
|
|
gCoreFunction->MNNReorderWeightInt4 = MNNReorderWeightInt4;
|
|
gCoreFunction->MNNSumWeightInt8 = MNNSumWeightInt8;
|
|
#ifdef __aarch64__
|
|
if (gCoreFunction->supportSDot) {
|
|
gCoreFunction->MNNReorderWeightInt4 = MNNReorderWeightInt4Arm82;
|
|
gCoreFunction->MNNSumWeightInt8 = MNNSumWeightInt8Arm82;
|
|
gCoreFunction->arm82MatmulRelatedFunctions.MNNReorderWeightInt4 = MNNReorderWeightInt4Arm82;
|
|
gCoreFunction->arm82MatmulRelatedFunctions.MNNSumWeightInt8 = MNNSumWeightInt8Arm82;
|
|
}
|
|
if (gCoreFunction->supportI8mm) {
|
|
gCoreFunction->MNNReorderWeightInt4 = MNNReorderWeightInt4Arm86;
|
|
gCoreFunction->MNNSumWeightInt8 = MNNSumWeightInt8Arm86;
|
|
}
|
|
#endif
|
|
#ifdef MNN_LOW_MEMORY
|
|
gCoreFunction->MNNAbsMax = MNNAbsMaxFP32; // abs max value for [icDiv4,plane,4] -> abs max:[plane]
|
|
gCoreFunction->MNNDynamicQuant = MNNDynamicQuantFP32; // symmetric 'batch' quant for [icDiv4,plane,4]
|
|
gCoreFunction->MNNAsyQuantFunc = MNNAsyQuantFunc; // asymmetric 'batch' quant for [icDiv4,plane,4]
|
|
gCoreFunction->MNNAsyQuantInfo =
|
|
MNNAsyQuantInfo_FP32; // asymmetric quant/dequant scale&bias for [icDiv4,plane,4] -> scale&bias:[blockNum,plane]
|
|
gCoreFunction->MNNQuantScale =
|
|
MNNQuantScaleFP32; // symmetric quant/dequant scale&bias for [icDiv4,plane,4] -> scale&bias:[plane]
|
|
gCoreFunction->MNNGeneralIm2Col = generalIm2col; // Im2Col based on float data -> output:[eU,kernelsize,lU,ep,lp]
|
|
gCoreFunction->MNNDynamicUpdateConvBiasScale = MNNDynamicUpdateConvBiasScale;
|
|
#ifdef __aarch64__
|
|
if (gCoreFunction->supportSDot) {
|
|
gCoreFunction->MNNGeneralIm2Col = MNNGeneralIm2col_Fp32Arm82;
|
|
gCoreFunction->arm82MatmulRelatedFunctions.MNNGeneralIm2Col = MNNGeneralIm2col_Fp32Arm82;
|
|
}
|
|
if (gCoreFunction->supportI8mm) {
|
|
gCoreFunction->MNNGeneralIm2Col = MNNGeneralIm2col_Fp32Arm86;
|
|
}
|
|
#endif
|
|
#endif
|
|
|
|
#if defined(__riscv) && defined(MNN_USE_RVV)
|
|
if (gCoreFunction->supportRVV) {
|
|
gCoreFunction->MNNAccumulateSequenceNumber = MNNAccumulateSequenceNumber_RVV;
|
|
gCoreFunction->MNNSumByAxisLForMatmul_A = MNNSumByAxisLForMatmul_A_RVV;
|
|
gCoreFunction->MNNReorderWeightInt4 = MNNReorderWeightInt4_RVV;
|
|
gCoreFunction->MNNSumWeightInt8 = MNNSumWeightInt8_RVV;
|
|
gCoreFunction->MNNPackedMatMul = MNNPackedMatMulFP32_RVV;
|
|
gCoreFunction->MNNPackedMatMulRemain = MNNPackedMatMulRemainFP32_RVV;
|
|
gCoreFunction->MNNPackForMatMul_B = MNNPackForMatMul_B_RVV;
|
|
gCoreFunction->MNNGetMatMulPackMode = MNNGetMatMulPackMode_RVV;
|
|
#ifdef MNN_LOW_MEMORY
|
|
gCoreFunction->MNNAbsMax = MNNAbsMaxFP32_RVV;
|
|
gCoreFunction->MNNDynamicQuant = MNNDynamicQuantFP32_RVV;
|
|
gCoreFunction->MNNAsyQuantFunc = MNNAsyQuantFunc_RVV;
|
|
gCoreFunction->MNNAsyQuantInfo = MNNAsyQuantInfo_FP32_RVV;
|
|
gCoreFunction->MNNGeneralIm2Col = generalIm2col_RVV;
|
|
gCoreFunction->MNNDynamicUpdateConvBiasScale = MNNDynamicUpdateConvBiasScale_RVV;
|
|
gCoreFunction->MNNQuantScale = MNNQuantScaleFP32_RVV;
|
|
#endif
|
|
}
|
|
#endif
|
|
|
|
#ifdef __aarch64__
|
|
#ifdef MNN_SME2
|
|
if (gCoreFunction->supportSME2) {
|
|
// Int8 Gemm related
|
|
gCoreFunction->MNNSumWeightInt8 = MNNSumWeightInt8Sme2_Hp32;
|
|
gCoreFunction->MNNSumWeightInt8SmeHp128 = MNNSumWeightInt8Sme2_Hp128;
|
|
gCoreFunction->MNNReorderWeightInt4 = MNNReorderWeightInt4Sme2;
|
|
|
|
#ifdef MNN_LOW_MEMORY
|
|
gCoreFunction->MNNGeneralIm2Col = MNNGeneralIm2col_Fp32Sme2;
|
|
#endif
|
|
|
|
gCoreFunction->int8MatmulRelatedFunctions.MNNSumWeightInt8SmeHp128 = MNNSumWeightInt8Sme2_Hp128;
|
|
|
|
// Float Gemm related
|
|
gCoreFunction->MNNPackedMatMul = MNNPackedMatMulFP32_SME2;
|
|
gCoreFunction->MNNPackedMatMulRemain = MNNPackedMatMulRemainFP32_SME2;
|
|
gCoreFunction->MNNGetMatMulPackMode = SME2MNNGetMatMulPackMode;
|
|
gCoreFunction->MNNPackC4ForMatMul_A = Sme2MNNPackC4ForMatMul_A;
|
|
gCoreFunction->MNNPackForMatMul_B = Sme2MNNPackForMatMul_B;
|
|
}
|
|
#endif // MNN_SME2
|
|
#endif // __aarch64__
|
|
|
|
{ // Update the function pointers in the int8MatmulRelatedFunctions struct.
|
|
gCoreFunction->int8MatmulRelatedFunctions.MNNReorderWeightInt4 = gCoreFunction->MNNReorderWeightInt4;
|
|
gCoreFunction->int8MatmulRelatedFunctions.MNNSumWeightInt8 = gCoreFunction->MNNSumWeightInt8;
|
|
gCoreFunction->int8MatmulRelatedFunctions.MNNGeneralIm2Col = gCoreFunction->MNNGeneralIm2Col;
|
|
}
|
|
MNNCoreInt8FunctionInit();
|
|
MNNFunctionInit();
|
|
}
|
|
CoreFunctions* MNNGetCoreFunctions() {
|
|
return gCoreFunction;
|
|
}
|
|
}; // namespace MNN
|
|
|
|
void MNNUnpackC4Origin(float* dst, const float* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNUnpackC4(dst, src, area, depth, offset);
|
|
}
|
|
void MNNPackC4Origin(float* dst, const float* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNPackC4(dst, src, area, depth, offset);
|
|
}
|
|
|
|
void MNNPackC2(double* dst, const double* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNPackC2Common<double>(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC2(double* dst, const double* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNUnpackC2Common<double>(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC2Float(float* dst, const float* src, size_t area, size_t depth, int* areaOffset, int pack) {
|
|
MNNUnpackC2Common<float>(dst, src, area, depth, areaOffset, pack);
|
|
}
|
|
#ifndef __aarch64__
|
|
void MNNPackInt8C2(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNPackC2Common<float>(dst, src, area, depth, areaOffset);
|
|
}
|
|
#endif
|
|
|
|
void MNNUnpackInt8C2(float* dst, const float* src, size_t area, size_t depth, int* areaOffset) {
|
|
MNNUnpackC2Common<float>(dst, src, area, depth, areaOffset);
|
|
}
|
|
|
|
void MNNUnpackC2Origin(double* dst, const double* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNUnpackC2(dst, src, area, depth, offset);
|
|
}
|
|
void MNNPackC2Origin(double* dst, const double* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNPackC2(dst, src, area, depth, offset);
|
|
}
|
|
|
|
void MNNUnpackInt8C2Origin(float* dst, const float* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNUnpackInt8C2(dst, src, area, depth, offset);
|
|
}
|
|
void MNNPackInt8C2Origin(float* dst, const float* src, size_t area, size_t depth, int areaOffset) {
|
|
int offset[] = {
|
|
areaOffset,
|
|
areaOffset,
|
|
};
|
|
MNNPackInt8C2(dst, src, area, depth, offset);
|
|
}
|