aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/cl_kernels/winograd_output_transform.cl
diff options
context:
space:
mode:
authorGiorgio Arena <giorgio.arena@arm.com>2020-10-12 11:53:51 +0100
committerGiorgio Arena <giorgio.arena@arm.com>2020-10-12 14:06:24 +0000
commitd056e574f60ca731b2d078e56c6baca5a6c642ac (patch)
tree30a47f1df1341b0462f344f4234bcf5dbf5d4360 /src/core/CL/cl_kernels/winograd_output_transform.cl
parentff9612cad3288ab96f94e86485e591401753f56b (diff)
downloadComputeLibrary-d056e574f60ca731b2d078e56c6baca5a6c642ac.tar.gz
COMPMID-3826 ArmNN Nightly failing for CL
Change-Id: I09f557b5cecafc669e12764e8592457212168d62 Signed-off-by: Giorgio Arena <giorgio.arena@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4131 Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/winograd_output_transform.cl')
-rw-r--r--src/core/CL/cl_kernels/winograd_output_transform.cl69
1 files changed, 36 insertions, 33 deletions
diff --git a/src/core/CL/cl_kernels/winograd_output_transform.cl b/src/core/CL/cl_kernels/winograd_output_transform.cl
index efd8502657..e735bbafb6 100644
--- a/src/core/CL/cl_kernels/winograd_output_transform.cl
+++ b/src/core/CL/cl_kernels/winograd_output_transform.cl
@@ -158,11 +158,11 @@ __kernel void winograd_output_transform_2x2_3x3_nchw(
// Store the output tile
#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
const VEC_DATA_TYPE(DATA_TYPE, 2)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL);
*((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out0_dt.s1;
#else // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
- vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL), 0,
+ vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
@@ -172,7 +172,7 @@ __kernel void winograd_output_transform_2x2_3x3_nchw(
out10 += (DATA_TYPE)b;
out11 += (DATA_TYPE)b;
#endif // defined(HAS_BIAS)
- vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 2))(out10, out11), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL), 0,
+ vstore2(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 2))(out10, out11), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
#endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
}
@@ -281,14 +281,15 @@ __kernel void winograd_output_transform_2x2_7x7_nhwc(
offset = min(offset + (int2)(0, 1) * (int2)dst_stride_z, (int2)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
VEC_DATA_TYPE(DATA_TYPE, 2)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL);
*(__global DATA_TYPE *)(dst_ptr + offset.s0) = out0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + offset.s1) = out0_dt.s1;
#else // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
// Get output address
int offset = dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
VEC_DATA_TYPE(DATA_TYPE, 2)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 2))(out00, out01), VEC_DATA_TYPE(DATA_TYPE, 2)), A_VAL,
+ B_VAL);
*(__global DATA_TYPE *)(dst_ptr + 0 * dst_stride_y + offset) = out0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + 1 * dst_stride_y + offset) = out0_dt.s1;
#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
@@ -398,9 +399,9 @@ __kernel void winograd_output_transform_2x2_7x7_nhwc(
// Store the output tile
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col1, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col1, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
*(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 0 * (int)dst_stride_y + offset.s0) = out_col0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 1 * (int)dst_stride_y + offset.s0) = out_col1_dt.s0;
@@ -599,14 +600,14 @@ __kernel void winograd_output_transform_4x4_3x3_nchw(
// Store the output tile
#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
- B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
+ B_VAL);
*((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out0_dt.s1;
*((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y)) = out0_dt.s2;
*((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y)) = out0_dt.s3;
#else // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
@@ -628,11 +629,11 @@ __kernel void winograd_output_transform_4x4_3x3_nchw(
out32 += (float)b;
out33 += (float)b;
#endif // defined(HAS_BIAS)
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out10, out11, out12, out13), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out10, out11, out12, out13), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out20, out21, out22, out23), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out20, out21, out22, out23), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y));
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out30, out31, out32, out33), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out30, out31, out32, out33), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y));
#endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
}
@@ -840,7 +841,8 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
// Store the 1x4 output tile
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
+ B_VAL);
*((__global DATA_TYPE *)(dst_ptr + offset.s0)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_ptr + offset.s1)) = out0_dt.s1;
*((__global DATA_TYPE *)(dst_ptr + offset.s2)) = out0_dt.s2;
@@ -851,7 +853,7 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
int mult_y = min(dst_size - offset, 1);
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)),
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)),
A_VAL, B_VAL);
*((__global DATA_TYPE *)(dst_ptr + mult_y * 0 * dst_stride_y + offset)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_ptr + mult_y * 1 * dst_stride_y + offset)) = out0_dt.s1;
@@ -869,14 +871,14 @@ __kernel void winograd_output_transform_4x4_3x3_nhwc(
// Store the 4x4 output tile
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, 4)
- out1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out10, out11, out12, out13), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
+ out1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out10, out11, out12, out13), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, 4)
- out2_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out20, out21, out22, out23), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
+ out2_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out20, out21, out22, out23), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, 4)
- out3_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out30, out31, out32, out33),
- VEC_DATA_TYPE(DATA_TYPE, 4)),
+ out3_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out30, out31, out32, out33),
+ VEC_DATA_TYPE(DATA_TYPE, 4)),
A_VAL, B_VAL);
*((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 0 * dst_stride_y + offset.s0)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 1 * dst_stride_y + offset.s0)) = out0_dt.s1;
@@ -1012,14 +1014,14 @@ __kernel void winograd_output_transform_4x4_5x5_nchw(
// Store the output tile
#if defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
- B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
+ B_VAL);
*((__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y)) = out0_dt.s0;
*((__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y)) = out0_dt.s1;
*((__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y)) = out0_dt.s2;
*((__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y)) = out0_dt.s3;
#else // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr));
#endif // defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
@@ -1135,13 +1137,13 @@ __kernel void winograd_output_transform_4x4_5x5_nchw(
#endif // defined(HAS_BIAS)
// Store the output tile
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s0, out_col1.s0, out_col2.s0, out_col3.s0), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s0, out_col1.s0, out_col2.s0, out_col3.s0), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s1, out_col1.s1, out_col2.s1, out_col3.s1), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s1, out_col1.s1, out_col2.s1, out_col3.s1), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s2, out_col1.s2, out_col2.s2, out_col3.s2), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s2, out_col1.s2, out_col2.s2, out_col3.s2), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y));
- vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s3, out_col1.s3, out_col2.s3, out_col3.s3), A_VAL, B_VAL), 0,
+ vstore4(ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, (VEC_DATA_TYPE(DATA_TYPE, 4))(out_col0.s3, out_col1.s3, out_col2.s3, out_col3.s3), A_VAL, B_VAL), 0,
(__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y));
#endif // !defined(WINOGRAD_OUTPUT_TRANSFORM_HORIZONTAL) && !defined(WINOGRAD_OUTPUT_TRANSFORM_VERTICAL)
}
@@ -1241,7 +1243,8 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc(
offset = min(offset + (int4)(0, 1, 2, 3) * (int4)dst_stride_z, (int4)dst_size); // If address is beyond the last plane, clamp it to dst_size (which points to the last padding).
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL, B_VAL);
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
+ B_VAL);
*(__global DATA_TYPE *)(dst_ptr + offset.s0) = out0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + offset.s1) = out0_dt.s1;
*(__global DATA_TYPE *)(dst_ptr + offset.s2) = out0_dt.s2;
@@ -1250,7 +1253,7 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc(
// Get output address
int offset = dst_offset_first_element_in_bytes + x_out * sizeof(DATA_TYPE) + y_out * dst_stride_y + z_out * dst_stride_z;
VEC_DATA_TYPE(DATA_TYPE, 4)
- out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
+ out0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT((VEC_DATA_TYPE(float, 4))(out00, out01, out02, out03), VEC_DATA_TYPE(DATA_TYPE, 4)), A_VAL,
B_VAL);
*(__global DATA_TYPE *)(dst_ptr + 0 * dst_stride_y + offset) = out0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + 1 * dst_stride_y + offset) = out0_dt.s1;
@@ -1379,13 +1382,13 @@ __kernel void winograd_output_transform_4x4_5x5_nhwc(
// Store the output tile
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col0_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col1, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col1_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col1, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col2_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col2, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col2_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col2, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
- out_col3_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, CONVERT(out_col3, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
+ out_col3_dt = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, CONVERT(out_col3, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)), A_VAL, B_VAL);
*(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 0 * (int)dst_stride_y + offset.s0) = out_col0_dt.s0;
*(__global DATA_TYPE *)(dst_ptr + mult_y.s0 * 1 * (int)dst_stride_y + offset.s0) = out_col1_dt.s0;