m53297601
/
mindspore
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

add judgement for CL

Browse Source
pull/3862/head
pengyongrong 5 years ago
parent 6a1e6b01f7
commit d831f8df5e
9 changed files with 194 additions and 319 deletions
Show Stats Download Patch File Download Diff File

8
build.sh
Unescape View File

@ -510,12 +510,8 @@ gene_ocl_program() {
build_opencl() {
cd ${BASEPATH}
if [[ ! -d "third_party/OpenCL-Headers" ]]; then
git submodule update --init third_party/OpenCL-Headers
fi
if [[ ! -d "third_party/OpenCL-CLHPP" ]]; then
git submodule update --init third_party/OpenCL-CLHPP
fi
git submodule update --init third_party/OpenCL-Headers
git submodule update --init third_party/OpenCL-CLHPP
if [[ "${OPENCL_OFFLINE_COMPILE}" == "on" ]]; then
gene_ocl_program
else

44
mindspore/lite/src/runtime/kernel/opencl/cl/fp16/depthwise_conv2d.cl
Unescape View File

@ -13,19 +13,9 @@
__constant sampler_t smp_edge = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST;
__constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
__constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
__kernel void DepthwiseConv2d_NC4HW4(
__global FLT4* src_data,
__global FLT4* filters,
__global FLT4* biases,
float relu_clip1,
__global FLT4* dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size
) {
__kernel void DepthwiseConv2d_NC4HW4(__global FLT4 *src_data, __global FLT4 *filters, __global FLT4 *biases,
float relu_clip1, __global FLT4 *dst_data, int2 kernel_size, int2 stride,
int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -42,31 +32,21 @@ __global FLT4* dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filters[fx_c];
FLT4 src_final =src_data[(((Z) * src_size.y + (y_c)) * src_size.x + (x_c))];
FLT4 src_final = src_data[(((Z)*src_size.y + (y_c)) * src_size.x + (x_c))];
r += TO_ACCUM_TYPE(src_final * f);
};
}
fx_c++;
}
}
FLT4 bias_val = biases[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
dst_data[(((Z) * dst_size.y + (Y)) * dst_size.x + (X))] = res0;
dst_data[(((Z)*dst_size.y + (Y)) * dst_size.x + (X))] = res0;
}
__kernel void DepthwiseConv2d_NHWC4(
__global FLT4* src_data,
__global FLT4* filters,
__global FLT4* biases,
float relu_clip1,
__global FLT4* dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size
) {
__kernel void DepthwiseConv2d_NHWC4(__global FLT4 *src_data, __global FLT4 *filters, __global FLT4 *biases,
float relu_clip1, __global FLT4 *dst_data, int2 kernel_size, int2 stride,
int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -83,9 +63,9 @@ __global FLT4* dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filters[fx_c];
FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final = src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
r += TO_ACCUM_TYPE(src_final * f);
};
}
fx_c++;
}
}
@ -93,4 +73,4 @@ __global FLT4* dst_data,
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
}
}

7
mindspore/lite/src/runtime/kernel/opencl/cl/fp16/matmul.cl
Unescape View File

@ -1,9 +1,8 @@
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
#define FLT4 half4
#define FLT16 half16
__kernel void MatMul(__global FLT4 *x, __global FLT16 *weight,
__global FLT4 *buffer, __global FLT4 *bias, int2 offset_ci,
int2 offset_co, int has_bias) {
__kernel void MatMul(__global FLT4 *x, __global FLT16 *weight, __global FLT4 *buffer, __global FLT4 *bias,
int2 offset_ci, int2 offset_co, int has_bias) {
int2 gid = (int2)(get_global_id(0), get_global_id(1));
int2 lid = (int2)(get_local_id(0), get_local_id(1));
FLT4 s = (FLT4)(0.0f);
@ -29,4 +28,4 @@ __kernel void MatMul(__global FLT4 *x, __global FLT16 *weight,
buffer[gid.x] = s;
// memory pollution? or protected by opencl
}
}
}

3
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/avg_pool2d.cl
Unescape View File

@ -31,7 +31,6 @@ __kernel void AvgPooling2d(__global float4 *input, __global float4 *output, cons
__constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
__kernel void AvgPooling2dImage2d(__read_only image2d_t input, __write_only image2d_t output, const int4 input_shape,
const int4 output_shape, const int2 stride, const int2 kernel_size,
const int2 padding) {
@ -63,4 +62,4 @@ __kernel void AvgPooling2dImage2d(__read_only image2d_t input, __write_only imag
}
float4 result = convert_float4(r / window_size);
write_imagef(output, (int2)(X, Y * output_shape.w + Z), result);
}
}

2
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/concat.cl
Unescape View File

@ -51,4 +51,4 @@ __kernel void Concat3input(__global float *input0, __global float *input1, __glo
output[index_output] = input2[input_idx];
}
}
}
}

295
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/convolution.cl
View File

File diff suppressed because it is too large Load Diff

130
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/depthwise_conv2d.cl
Unescape View File

@ -8,18 +8,9 @@
#define TO_FLT4 convert_float4
#endif
__constant sampler_t sampler_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
__kernel void DepthwiseConv2d_IMG_NC4HW4(
__read_only image2d_t src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__write_only image2d_t dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_IMG_NC4HW4(__read_only image2d_t src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __write_only image2d_t dst_data, int2 kernel_size,
int2 stride, int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -36,32 +27,23 @@ __write_only image2d_t dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
//FLT4 src_final =src_data[(((Z) * src_size.y + (y_c)) * src_size.x + (x_c))];
FLT4 src_final =read_imagef(src_data, sampler_zero, (int2)(x_c, (Z * src_size.y + y_c)));
// FLT4 src_final =src_data[(((Z) * src_size.y + (y_c)) * src_size.x + (x_c))];
FLT4 src_final = read_imagef(src_data, sampler_zero, (int2)(x_c, (Z * src_size.y + y_c)));
r += TO_FLT4(src_final * f);
};
}
fx_c++;
}
}
FLT4 bias_val = bias[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
//dst_data[(((Z) * dst_size.y + (Y)) * dst_size.x + (X))] = res0;
// dst_data[(((Z) * dst_size.y + (Y)) * dst_size.x + (X))] = res0;
write_imagef(dst_data, (int2)(X, (Z * dst_size.y + Y)), res0);
}
__kernel void DepthwiseConv2d_IMG_NHWC4(
__read_only image2d_t src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__write_only image2d_t dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_IMG_NHWC4(__read_only image2d_t src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __write_only image2d_t dst_data, int2 kernel_size,
int2 stride, int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -78,32 +60,23 @@ __write_only image2d_t dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
//FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final =read_imagef(src_data, sampler_zero, (int2)(Z+x_c*src_size.z, y_c));
// FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final = read_imagef(src_data, sampler_zero, (int2)(Z + x_c * src_size.z, y_c));
r += TO_FLT4(src_final * f);
};
}
fx_c++;
}
}
FLT4 bias_val = bias[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
//dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
write_imagef(dst_data, (int2)(X*dst_size.z+Z, Y), res0);
// dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
write_imagef(dst_data, (int2)(X * dst_size.z + Z, Y), res0);
}
__kernel void DepthwiseConv2d_IMG_NHWC4_1x1(
__read_only image2d_t src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__write_only image2d_t dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_IMG_NHWC4_1x1(__read_only image2d_t src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __write_only image2d_t dst_data, int2 kernel_size,
int2 stride, int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -120,30 +93,21 @@ __write_only image2d_t dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
//FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
// FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final = read_imagef(src_data, sampler_zero, (int2)(Z, (y_c * src_size.x + x_c) * src_size.z));
r += TO_FLT4(src_final * f);
};
}
}
}
FLT4 bias_val = bias[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
//dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
// dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
write_imagef(dst_data, (int2)(Z, (Y * dst_size.x + X) * dst_size.z), res0);
}
__kernel void DepthwiseConv2d_BUF_NC4HW4(
__global FLT4* src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__global FLT4* dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_BUF_NC4HW4(__global FLT4 *src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __global FLT4 *dst_data, int2 kernel_size, int2 stride,
int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -160,30 +124,21 @@ __global FLT4* dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
FLT4 src_final =src_data[(((Z) * src_size.y + (y_c)) * src_size.x + (x_c))];
FLT4 src_final = src_data[(((Z)*src_size.y + (y_c)) * src_size.x + (x_c))];
r += TO_FLT4(src_final * f);
};
}
fx_c++;
}
}
FLT4 bias_val = bias[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
dst_data[(((Z) * dst_size.y + (Y)) * dst_size.x + (X))] = res0;
dst_data[(((Z)*dst_size.y + (Y)) * dst_size.x + (X))] = res0;
}
__kernel void DepthwiseConv2d_BUF_NHWC4(
__global FLT4* src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__global FLT4* dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_BUF_NHWC4(__global FLT4 *src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __global FLT4 *dst_data, int2 kernel_size, int2 stride,
int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -200,9 +155,9 @@ __global FLT4* dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final = src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
r += TO_FLT4(src_final * f);
};
}
fx_c++;
}
}
@ -212,18 +167,9 @@ __global FLT4* dst_data,
dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
}
__kernel void DepthwiseConv2d_BUF_NHWC4_1x1(
__global FLT4* src_data,
__global FLT4* filter,
__global FLT4* bias,
float relu_clip1,
__global FLT4* dst_data,
int2 kernel_size,
int2 stride,
int2 padding,
int2 dilation,
int4 src_size,
int4 dst_size) {
__kernel void DepthwiseConv2d_BUF_NHWC4_1x1(__global FLT4 *src_data, __global FLT4 *filter, __global FLT4 *bias,
float relu_clip1, __global FLT4 *dst_data, int2 kernel_size, int2 stride,
int2 padding, int2 dilation, int4 src_size, int4 dst_size) {
int X = get_global_id(0);
int Y = get_global_id(1);
int Z = get_global_id(2);
@ -240,13 +186,13 @@ __global FLT4* dst_data,
bool outside_x = x_c < 0 || x_c >= src_size.x;
if (!outside_x && !outside_y) {
FLT4 f = filter[fx_c];
FLT4 src_final =src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
FLT4 src_final = src_data[((y_c * src_size.x + x_c) * src_size.z + Z)];
r += TO_FLT4(src_final * f);
};
}
}
}
FLT4 bias_val = bias[Z];
FLT4 res0 = TO_FLT4(r) + bias_val;
res0 = clamp(res0, (FLT)(0.0f), (FLT)(relu_clip1));
dst_data[((Y * dst_size.x + X) * dst_size.z + Z)] = res0;
}
}

7
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/matmul.cl
Unescape View File

@ -1,8 +1,7 @@
#define FLT4 float4
#define FLT16 float16
__kernel void MatMul(__global FLT4 *x, __global FLT16 *weight,
__global FLT4 *buffer, __global FLT4 *bias, int2 offset_ci,
int2 offset_co, int has_bias) {
__kernel void MatMul(__global FLT4 *x, __global FLT16 *weight, __global FLT4 *buffer, __global FLT4 *bias,
int2 offset_ci, int2 offset_co, int has_bias) {
int2 gid = (int2)(get_global_id(0), get_global_id(1));
int2 lid = (int2)(get_local_id(0), get_local_id(1));
FLT4 s = (FLT4)(0.0f);
@ -28,4 +27,4 @@ __kernel void MatMul(__global FLT4 *x, __global FLT16 *weight,
buffer[gid.x] = s;
// memory pollution? or protected by opencl
}
}
}

17
mindspore/lite/src/runtime/kernel/opencl/cl/fp32/softmax.cl
Unescape View File

@ -1,16 +1,13 @@
#define SLICES 4
int DivideRoundUp(int n, int div)
{
int q = n / div;
return n % div == 0 ? q : q + 1;
int DivideRoundUp(int n, int div) {
int q = n / div;
return n % div == 0 ? q : q + 1;
}
__kernel void SoftMax(__global float4 *input,
__global float4 *output,
const int4 input_shape) {
int X = get_global_id(0); // width
int Y = get_global_id(1); // height
__kernel void SoftMax(__global float4 *input, __global float4 *output, const int4 input_shape) {
int X = get_global_id(0); // width
int Y = get_global_id(1); // height
int H = input_shape.y;
int W = input_shape.z;
int C = input_shape.w;
@ -32,4 +29,4 @@ __kernel void SoftMax(__global float4 *input,
float4 result = convert_float4(t);
output[(Y * W + X * H) * C + d] = result;
}
}
}

Write Preview
Loading…
Cancel
Save