Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 dfa731e1 编写于 作者: L liuruilong
浏览文件
操作

format files

上级 3b82bfb5
隐藏空白更改
内联 并排
Showing with 194 addition and 190 deletion
src/operators/kernel/arm/batchnorm_kernel.cpp
浏览文件 @ dfa731e1
......@@ -45,34 +45,35 @@ void BatchNormKernel<CPU, float>::Compute(const BatchNormParam &param) const {
auto scale_ptr = scale->data<float>();
auto bias_ptr = bias->data<float>();
// Tensor inv_std;
// auto inv_std_ptr = inv_std.mutable_data<float>(make_ddim({C}));
PADDLE_MOBILE_ENFORCE(C == variance->numel(), "C must equal to variance.numel()");
PADDLE_MOBILE_ENFORCE(C == variance->numel(),
"C must equal to variance.numel()");
int HXW = H * W;
if (HXW > 32) {
int NXC = N * C;
float *inv_std_ptr = new float[NXC * 4];
float * volatile new_scale_ptr = new float[NXC * 4];
float * volatile new_bias_ptr = new float[NXC * 4];
float *volatile new_scale_ptr = new float[NXC * 4];
float *volatile new_bias_ptr = new float[NXC * 4];
/// std = (var + epsilon).sqrt();
/// inv_std = 1 / std;
for (int i = 0; i < C * 4; i += 4) {
inv_std_ptr[i] =
1 / static_cast<float>(pow((variance_ptr[i/4] + epsilon), 0.5));
1 / static_cast<float>(pow((variance_ptr[i / 4] + epsilon), 0.5));
inv_std_ptr[i + 1] = inv_std_ptr[i];
inv_std_ptr[i + 2] = inv_std_ptr[i];
inv_std_ptr[i + 3] = inv_std_ptr[i];
new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i/4];
new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i / 4];
new_scale_ptr[i + 1] = new_scale_ptr[i];
new_scale_ptr[i + 2] = new_scale_ptr[i];
new_scale_ptr[i + 3] = new_scale_ptr[i];
new_bias_ptr[i] = bias_ptr[i/4] - mean_ptr[i/4] * inv_std_ptr[i] * scale_ptr[i/4];
new_bias_ptr[i] =
bias_ptr[i / 4] - mean_ptr[i / 4] * inv_std_ptr[i] * scale_ptr[i / 4];
new_bias_ptr[i + 1] = new_bias_ptr[i];
new_bias_ptr[i + 2] = new_bias_ptr[i];
......@@ -84,116 +85,116 @@ void BatchNormKernel<CPU, float>::Compute(const BatchNormParam &param) const {
new_bias_ptr[j] = new_bias_ptr[j - C * 4];
}
asm volatile(
"subs %[N], %[N], #1 \n\t"
"blt end_n_%= \n\t"
"loop_n_%=: \n\t"
"subs %[C], %[C], #1 \n\t"
"blt end_c_%= \n\t"
"loop_c_%=: \n\t"
"vld1.32 {q9}, [%[new_scale_ptr]]! \n\t"
"vld1.32 {q10}, [%[new_bias_ptr]]! \n\t"
"mov r6, %[HXW] \n\t"
"subs r6, r6, #32 \n\t"
"blt end_hw_%= \n\t"
"loop_hw_%=: \n\t"
"vld1.32 {q1, q2}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q3, q4}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q5, q6}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q7, q8}, [%[input_x_ptr]]! \n\t"
"vmul.f32 q1, q1, q9 \n\t"
"vmul.f32 q2, q2, q9 \n\t"
"vmul.f32 q3, q3, q9 \n\t"
"vmul.f32 q4, q4, q9 \n\t"
"vmul.f32 q5, q5, q9 \n\t"
"vmul.f32 q6, q6, q9 \n\t"
"vmul.f32 q7, q7, q9 \n\t"
"vmul.f32 q8, q8, q9 \n\t"
"vadd.f32 q1, q1, q10 \n\t"
"vadd.f32 q2, q2, q10 \n\t"
"vadd.f32 q3, q3, q10 \n\t"
"vadd.f32 q4, q4, q10 \n\t"
"vadd.f32 q5, q5, q10 \n\t"
"vadd.f32 q6, q6, q10 \n\t"
"vadd.f32 q7, q7, q10 \n\t"
"vadd.f32 q8, q8, q10 \n\t"
"vst1.32 {q1, q2}, [%[out_ptr]]! \n\t"
"vst1.32 {q3, q4}, [%[out_ptr]]! \n\t"
"vst1.32 {q5, q6}, [%[out_ptr]]! \n\t"
"vst1.32 {q7, q8}, [%[out_ptr]]! \n\t"
"subs r6, r6, #32 \n\t"
"bge loop_hw_%= \n\t"
"end_hw_%=: \n\t"
"cmp r6, #0 \n\t"
"bge end_remainder_%= \n\t"
"mov r5, #4 \n\t"
"mul r6, r6, r5 \n\t"
"add %[input_x_ptr], %[input_x_ptr], r6 \n\t"
"vld1.32 {q1, q2}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q3, q4}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q5, q6}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q7, q8}, [%[input_x_ptr]]! \n\t"
"vmul.f32 q1, q1, q9 \n\t"
"vmul.f32 q2, q2, q9 \n\t"
"vmul.f32 q3, q3, q9 \n\t"
"vmul.f32 q4, q4, q9 \n\t"
"vmul.f32 q5, q5, q9 \n\t"
"vmul.f32 q6, q6, q9 \n\t"
"vmul.f32 q7, q7, q9 \n\t"
"vmul.f32 q8, q8, q9 \n\t"
"vadd.f32 q1, q1, q10 \n\t"
"vadd.f32 q2, q2, q10 \n\t"
"vadd.f32 q3, q3, q10 \n\t"
"vadd.f32 q4, q4, q10 \n\t"
"vadd.f32 q5, q5, q10 \n\t"
"vadd.f32 q6, q6, q10 \n\t"
"vadd.f32 q7, q7, q10 \n\t"
"vadd.f32 q8, q8, q10 \n\t"
"add %[out_ptr], %[out_ptr], r6 \n\t"
"vst1.32 {q1, q2}, [%[out_ptr]]! \n\t"
"vst1.32 {q3, q4}, [%[out_ptr]]! \n\t"
"vst1.32 {q5, q6}, [%[out_ptr]]! \n\t"
"vst1.32 {q7, q8}, [%[out_ptr]]! \n\t"
"end_remainder_%=: \n\t"
"subs %[C], %[C], #1 \n\t"
"bge loop_c_%= \n\t"
"end_c_%=: \n\t"
"subs %[N], %[N], #1 \n\t"
"bge loop_n_%= \n\t"
"end_n_%=: \n\t"
:
:[input_x_ptr]"r"(input_x_ptr), [out_ptr]"r"(out_ptr), [new_scale_ptr]"r"(new_scale_ptr), [new_bias_ptr]"r"(new_bias_ptr),
[N]"r"(N), [C]"r"(C), [HXW]"r"(HXW)
:"memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9", "q10", "r5", "r6"
);
delete [] inv_std_ptr;
delete [] new_scale_ptr;
delete [] new_bias_ptr;
"subs %[N], %[N], #1 \n\t"
"blt end_n_%= \n\t"
"loop_n_%=: \n\t"
"subs %[C], %[C], #1 \n\t"
"blt end_c_%= \n\t"
"loop_c_%=: \n\t"
"vld1.32 {q9}, [%[new_scale_ptr]]! \n\t"
"vld1.32 {q10}, [%[new_bias_ptr]]! \n\t"
"mov r6, %[HXW] \n\t"
"subs r6, r6, #32 \n\t"
"blt end_hw_%= \n\t"
"loop_hw_%=: \n\t"
"vld1.32 {q1, q2}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q3, q4}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q5, q6}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q7, q8}, [%[input_x_ptr]]! \n\t"
"vmul.f32 q1, q1, q9 \n\t"
"vmul.f32 q2, q2, q9 \n\t"
"vmul.f32 q3, q3, q9 \n\t"
"vmul.f32 q4, q4, q9 \n\t"
"vmul.f32 q5, q5, q9 \n\t"
"vmul.f32 q6, q6, q9 \n\t"
"vmul.f32 q7, q7, q9 \n\t"
"vmul.f32 q8, q8, q9 \n\t"
"vadd.f32 q1, q1, q10 \n\t"
"vadd.f32 q2, q2, q10 \n\t"
"vadd.f32 q3, q3, q10 \n\t"
"vadd.f32 q4, q4, q10 \n\t"
"vadd.f32 q5, q5, q10 \n\t"
"vadd.f32 q6, q6, q10 \n\t"
"vadd.f32 q7, q7, q10 \n\t"
"vadd.f32 q8, q8, q10 \n\t"
"vst1.32 {q1, q2}, [%[out_ptr]]! \n\t"
"vst1.32 {q3, q4}, [%[out_ptr]]! \n\t"
"vst1.32 {q5, q6}, [%[out_ptr]]! \n\t"
"vst1.32 {q7, q8}, [%[out_ptr]]! \n\t"
"subs r6, r6, #32 \n\t"
"bge loop_hw_%= \n\t"
"end_hw_%=: \n\t"
"cmp r6, #0 \n\t"
"bge end_remainder_%= \n\t"
"mov r5, #4 \n\t"
"mul r6, r6, r5 \n\t"
"add %[input_x_ptr], %[input_x_ptr], r6 \n\t"
"vld1.32 {q1, q2}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q3, q4}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q5, q6}, [%[input_x_ptr]]! \n\t"
"vld1.32 {q7, q8}, [%[input_x_ptr]]! \n\t"
"vmul.f32 q1, q1, q9 \n\t"
"vmul.f32 q2, q2, q9 \n\t"
"vmul.f32 q3, q3, q9 \n\t"
"vmul.f32 q4, q4, q9 \n\t"
"vmul.f32 q5, q5, q9 \n\t"
"vmul.f32 q6, q6, q9 \n\t"
"vmul.f32 q7, q7, q9 \n\t"
"vmul.f32 q8, q8, q9 \n\t"
"vadd.f32 q1, q1, q10 \n\t"
"vadd.f32 q2, q2, q10 \n\t"
"vadd.f32 q3, q3, q10 \n\t"
"vadd.f32 q4, q4, q10 \n\t"
"vadd.f32 q5, q5, q10 \n\t"
"vadd.f32 q6, q6, q10 \n\t"
"vadd.f32 q7, q7, q10 \n\t"
"vadd.f32 q8, q8, q10 \n\t"
"add %[out_ptr], %[out_ptr], r6 \n\t"
"vst1.32 {q1, q2}, [%[out_ptr]]! \n\t"
"vst1.32 {q3, q4}, [%[out_ptr]]! \n\t"
"vst1.32 {q5, q6}, [%[out_ptr]]! \n\t"
"vst1.32 {q7, q8}, [%[out_ptr]]! \n\t"
"end_remainder_%=: \n\t"
"subs %[C], %[C], #1 \n\t"
"bge loop_c_%= \n\t"
"end_c_%=: \n\t"
"subs %[N], %[N], #1 \n\t"
"bge loop_n_%= \n\t"
"end_n_%=: \n\t"
:
: [input_x_ptr] "r"(input_x_ptr), [out_ptr] "r"(out_ptr),
[new_scale_ptr] "r"(new_scale_ptr), [new_bias_ptr] "r"(new_bias_ptr),
[N] "r"(N), [C] "r"(C), [HXW] "r"(HXW)
: "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
"q10", "r5", "r6");
delete[] inv_std_ptr;
delete[] new_scale_ptr;
delete[] new_bias_ptr;
} else {
float *inv_std_ptr = new float[C];
for (int i = 0; i < C; i++) {
inv_std_ptr[i] =
1 / static_cast<float>(pow((variance_ptr[i] + epsilon), 0.5));
1 / static_cast<float>(pow((variance_ptr[i] + epsilon), 0.5));
}
Tensor new_scale;
......@@ -205,7 +206,8 @@ void BatchNormKernel<CPU, float>::Compute(const BatchNormParam &param) const {
/// (x * inv_var * scale) + (bias - est_mean * inv_var * scale)
for (int i = 0; i < C; i++) {
new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i];
new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
new_bias_ptr[i] =
bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
{
for (int n = 0; n < N; n++) {
for (int h = 0; h < H; h++) {
......@@ -213,23 +215,22 @@ void BatchNormKernel<CPU, float>::Compute(const BatchNormParam &param) const {
for (int w = 0; w < W; w++) {
int index = tmp_index + w;
out_ptr[index] =
input_x_ptr[index] * new_scale_ptr[i] + new_bias_ptr[i];
input_x_ptr[index] * new_scale_ptr[i] + new_bias_ptr[i];
}
}
}
}
}
delete [] inv_std_ptr;
// DLOG << "input[2,5,1,0](input[102]) ,channel 5 :";
// DLOG << "input_x_ptr : " << input_x_ptr[102];
// DLOG << "variance : " << variance_ptr[5];
// DLOG << "inv_std_ptr : " << inv_std_ptr[5];
// DLOG << "new_scale_ptr : " << new_scale_ptr[5];
// DLOG << "new_bias_ptr : " << new_bias_ptr[5];
// DLOG << "out_ptr : " << out_ptr[102];
delete[] inv_std_ptr;
// DLOG << "input[2,5,1,0](input[102]) ,channel 5 :";
// DLOG << "input_x_ptr : " << input_x_ptr[102];
// DLOG << "variance : " << variance_ptr[5];
// DLOG << "inv_std_ptr : " << inv_std_ptr[5];
// DLOG << "new_scale_ptr : " << new_scale_ptr[5];
// DLOG << "new_bias_ptr : " << new_bias_ptr[5];
// DLOG << "out_ptr : " << out_ptr[102];
}
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/relu_kernel.cpp
浏览文件 @ dfa731e1
......@@ -38,70 +38,71 @@ void ReluKernel<CPU, float>::Compute(const ReluParam &param) const {
auto *out_ptr = out->mutable_data<float>();
int numel = input_x->numel();
// if (numel > 64) {
// asm volatile(
// "pld [%[input_x_ptr], #0] \n\t"
// "vmov.f32 q8, #0.0 \n\t"
// "subs %[num], %[num], #32 \n\t"
// "blt end_num_%= \n\t"
// "loop_num_%=: \n\t"
// "pld [%[input_x_ptr], #1024] \n\t"
//
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
//
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
// "subs %[num], %[num], #32 \n\t"
// "bge loop_num_%= \n\t"
// "end_num_%=: \n\t"
// "cmp %[num], #0 \n\t"
// "bge end_%= \n\t"
// "mov r6, #4 \n\t"
// "mul r5, %[num], r6 \n\t"
// "add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
// "add %[out_ptr], %[out_ptr], r5 \n\t"
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
// "end_%=: \n\t"
// :
// :
// [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num] "r"(numel)
// : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "r5",
// "r6");
// } else {
ReluFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
// }
// if (numel > 64) {
// asm volatile(
// "pld [%[input_x_ptr], #0] \n\t"
// "vmov.f32 q8, #0.0 \n\t"
// "subs %[num], %[num], #32 \n\t"
// "blt end_num_%= \n\t"
// "loop_num_%=: \n\t"
// "pld [%[input_x_ptr], #1024] \n\t"
//
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
//
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
// "subs %[num], %[num], #32 \n\t"
// "bge loop_num_%= \n\t"
// "end_num_%=: \n\t"
// "cmp %[num], #0 \n\t"
// "bge end_%= \n\t"
// "mov r6, #4 \n\t"
// "mul r5, %[num], r6 \n\t"
// "add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
// "add %[out_ptr], %[out_ptr], r5 \n\t"
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
// "end_%=: \n\t"
// :
// :
// [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num]
// "r"(numel) : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
// "q7", "q8", "r5",
// "r6");
// } else {
ReluFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
// }
}
} // namespace operators
} // namespace paddle_mobile
......
test/common/test_lib_size.h
浏览文件 @ dfa731e1
......@@ -19,9 +19,9 @@ limitations under the License. */
#ifndef PADDLE_MOBILE_TEST_LIB_SIZE_H
#define PADDLE_MOBILE_TEST_LIB_SIZE_H
#include <vector>
#include <pthread.h>
#include <thread>
#include <vector>
//#include <list>
//#include <tuple>
//#include <typeinfo>
......@@ -74,7 +74,7 @@ void foo() {
// int z = 10;
// }
// std::shared_ptr<int> s1 = std::make_shared<int>();
// std::shared_ptr<int> s1 = std::make_shared<int>();
// std::stringstream ss;
// ss << "12345";
......
test/operators/test_batchnorm_op.cpp
浏览文件 @ dfa731e1
......@@ -137,7 +137,8 @@ int main() {
auto *inputx1_ptr = inputx1.data<float>();
paddle_mobile::framework::Tensor mean;
SetupTensor<float>(&mean, {256}, static_cast<float>(0), static_cast<float>(1));
SetupTensor<float>(&mean, {256}, static_cast<float>(0),
static_cast<float>(1));
auto *mean_ptr = mean.data<float>();
paddle_mobile::framework::Tensor scale;
......@@ -151,7 +152,8 @@ int main() {
auto *variance_ptr = variance.data<float>();
paddle_mobile::framework::Tensor bias;
SetupTensor<float>(&bias, {256}, static_cast<float>(0), static_cast<float>(1));
SetupTensor<float>(&bias, {256}, static_cast<float>(0),
static_cast<float>(1));
auto *bias_ptr = bias.data<float>();
paddle_mobile::framework::TestBatchNormOp<paddle_mobile::CPU> testBatchNormOp(
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或