修改simd实现

zjhellofss · Jul 26, 2023 · e8bfc18 · e8bfc18
1 parent 89237db
commit e8bfc18
Show file tree

Hide file tree

Showing 13 changed files with 29 additions and 1,854 deletions.
diff --git a/bench/bench_exp.cpp b/bench/bench_exp.cpp
@@ -4,9 +4,8 @@
 
 #include <benchmark/benchmark.h>
 #include <armadillo>
-#include "../source/layer/details/sse_mathfun.hpp"
+#include "../source/layer/details/arma_sse.hpp"
 #include "data/tensor.hpp"
-#include "utils/math/arma_sse.hpp"
 static void BM_ExpSimd(benchmark::State& state) {
   using namespace kuiper_infer;
   uint32_t input_c = state.range(0);

diff --git a/include/utils/math/fmath.hpp b/include/utils/math/fmath.hpp
@@ -412,7 +412,7 @@ inline float exp(float x)
 	using namespace local;
 	const ExpVar<>& expVar = C<>::expVar;
 
-#if 1
+#if __SSE2__
 	__m128 x1 = _mm_set_ss(x);
 
 	int limit = _mm_cvtss_si32(x1) & 0x7fffffff;
@@ -429,19 +429,7 @@ inline float exp(float x)
 	fi.i = ((u + 127) << 23) | expVar.tbl[v];
 	return (1 + t) * fi.f;
 #else
-	x = std::min(x, expVar.maxX[0]);
-	x = std::max(x, expVar.minX[0]);
-	float t = x * expVar.a[0];
-	const float magic = (1 << 23) + (1 << 22); // to round
-	t += magic;
-	fi fi;
-	fi.f = t;
-	t = x - (t - magic) * expVar.b[0];
-	int u = ((fi.i + (127 << expVar.s)) >> expVar.s) << 23;
-	unsigned int v = fi.i & mask(expVar.s);
-	fi.i = u | expVar.tbl[v];
-	return (1 + t) * fi.f;
-//	return (1 + t) * pow(2, (float)u) * pow(2, (float)v / n);
+    return std::exp(x);
 #endif
 }
 

diff --git a/source/utils/math/arma_sse.cpp → source/layer/details/arma_sse.cpp b/source/utils/math/arma_sse.cpp → source/layer/details/arma_sse.cpp
@@ -22,7 +22,7 @@
 //
 // Created by fss on 23-7-26.
 //
-#include "utils/math/arma_sse.hpp"
+#include "arma_sse.hpp"
 #include <glog/logging.h>
 
 namespace kuiper_infer {
@@ -66,7 +66,7 @@ void ArmaSigmoid(const arma::fcube& input_data, arma::fcube& output_data) {
   if (index < in_size) {
     while (index < in_size) {
       float value = input_data.at(index);
-      output_data.at(index) = 1 / (1.f + expf(-value));
+      output_data.at(index) = 1 / (1.f + fmath::exp(-value));
       index += 1;
     }
   }
@@ -166,7 +166,7 @@ void ArmaSiLU(const arma::fcube& input_data, arma::fcube& output_data) {
   if (j < size) {
     while (j < size) {
       float value = input_data.at(j);
-      output_data.at(j) = value / (1.f + expf(-value));
+      output_data.at(j) = value / (1.f + fmath::exp(-value));
       j += 1;
     }
   }

diff --git a/include/utils/math/arma_sse.hpp → source/layer/details/arma_sse.hpp b/include/utils/math/arma_sse.hpp → source/layer/details/arma_sse.hpp
@@ -22,10 +22,10 @@
 // Created by fss on 23-7-26.
 //
 
-#ifndef KUIPER_INFER_INCLUDE_MATH_ARMA_SIGMOID
-#define KUIPER_INFER_INCLUDE_MATH_ARMA_SIGMOID
+#ifndef KUIPER_INFER_INCLUDE_MATH_ARMA_SSE
+#define KUIPER_INFER_INCLUDE_MATH_ARMA_SSE
 #include <armadillo>
-#include "fmath.hpp"
+#include "utils/math/fmath.hpp"
 namespace kuiper_infer {
 namespace math {
 void ArmaSigmoid(const arma::fcube& input_data, arma::fcube& output_data);
@@ -35,4 +35,4 @@ void ArmaReLU(const arma::fcube& input_data, arma::fcube& output_data);
 void ArmaSiLU(const arma::fcube& input_data, arma::fcube& output_data);
 }  // namespace math
 }  // namespace kuiper_infer
-#endif  // KUIPER_INFER_INCLUDE_MATH_ARMA_SIGMOID
+#endif  // KUIPER_INFER_INCLUDE_MATH_ARMA_SSE
diff --git a/source/layer/details/platform.hpp b/source/layer/details/platform.hpp
diff --git a/source/layer/details/relu.cpp b/source/layer/details/relu.cpp
@@ -21,8 +21,8 @@
 
 // Created by fss on 22-11-18.
 #include "relu.hpp"
+#include "arma_sse.hpp"
 #include "layer/abstract/layer_factory.hpp"
-#include "utils/math/arma_sse.hpp"
 
 namespace kuiper_infer {
 InferStatus ReluLayer::Forward(

diff --git a/source/layer/details/sigmoid.cpp b/source/layer/details/sigmoid.cpp
@@ -23,8 +23,8 @@
 
 #include "sigmoid.hpp"
 #include <glog/logging.h>
+#include "arma_sse.hpp"
 #include "layer/abstract/layer_factory.hpp"
-#include "utils/math/arma_sse.hpp"
 
 namespace kuiper_infer {
 

diff --git a/source/layer/details/silu.cpp b/source/layer/details/silu.cpp
@@ -22,9 +22,9 @@
 // Created by fss on 22-12-25.
 
 #include "silu.hpp"
+#include "arma_sse.hpp"
 #include "layer/abstract/layer_factory.hpp"
 #include "tick.hpp"
-#include "utils/math/arma_sse.hpp"
 
 namespace kuiper_infer {
 

diff --git a/source/layer/details/softmax.cpp b/source/layer/details/softmax.cpp
@@ -26,6 +26,7 @@
 #include <numeric>
 #include "data/tensor_util.hpp"
 #include "layer/abstract/layer_factory.hpp"
+#include "utils/math/fmath.hpp"
 namespace kuiper_infer {
 #define POS_INDEX(outer_size, inner_size, axis_size) \
   outer_size* axis_sizes* inner_sizes + axis_size* inner_sizes + inner_size;
@@ -116,7 +117,7 @@ InferStatus SoftmaxLayer::Forward(
         for (uint32_t axis_size = 0; axis_size < axis_sizes; ++axis_size) {
           uint32_t index = POS_INDEX(outer_size, inner_size, axis_size);
           float cur_value = input_values.at(index);
-          float exp_sub_value = std::exp(cur_value - max_value);
+          float exp_sub_value = fmath::exp(cur_value - max_value);
 
           sum_value += exp_sub_value;
           output_values.at(index) = exp_sub_value;