aboutsummaryrefslogtreecommitdiff
path: root/src/core/CL/cl_kernels/pooling_layer_quantized.cl
diff options
context:
space:
mode:
authorGiorgio Arena <giorgio.arena@arm.com>2020-10-26 15:04:08 +0000
committerGiorgio Arena <giorgio.arena@arm.com>2020-11-12 12:42:51 +0000
commit2d1a835b68eb27a800838fc2b563b12eddf2c19f (patch)
tree228dee073d37d2ec5b5dfbdb3d0e1e512ecb2d22 /src/core/CL/cl_kernels/pooling_layer_quantized.cl
parent00c7601b1f9c3bec1d3b1db844abb513b9012541 (diff)
downloadComputeLibrary-2d1a835b68eb27a800838fc2b563b12eddf2c19f.tar.gz
COMPMID-3735 Remove OpenCL padding: CLSoftmaxLayerKernel
- Renamed SELECT_DATA_TYPE to SELECT_VEC_DATA_TYPE to reflect its usage with vectors. SELECT_DATA_TYPE(dt) will now return the primitive data type - Changed the interface of VEC_OFFS and V_OFFS in order to receive the primitive data type as a parameter rather than its vector form - Performed a general cleanup of the kernels, such as creating macro for sum and max reduces, remove reduntant macros, defines, variables, calculations, etc... - Using VEC_SIZE and VEC_SIZE_LEFTOVER in every kernel in order to allow computation for smaller shapes without adding paddings - Removed the actual padding from the kernel and adjusting its calculations accordingly. Added asserts for padding removal checks. Removed invalid Validate tests. Change-Id: If5ccbd5d34e255d38c7f6bfe8740e2b80b28e264 Signed-off-by: Giorgio Arena <giorgio.arena@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4277 Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: SiCong Li <sicong.li@arm.com> Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/cl_kernels/pooling_layer_quantized.cl')
-rw-r--r--src/core/CL/cl_kernels/pooling_layer_quantized.cl37
1 files changed, 17 insertions, 20 deletions
diff --git a/src/core/CL/cl_kernels/pooling_layer_quantized.cl b/src/core/CL/cl_kernels/pooling_layer_quantized.cl
index 04fef98cd0..d8cef2b4e6 100644
--- a/src/core/CL/cl_kernels/pooling_layer_quantized.cl
+++ b/src/core/CL/cl_kernels/pooling_layer_quantized.cl
@@ -197,29 +197,23 @@ __kernel void pooling_layer_MxN_quantized_nhwc(
{
// Note: If C is not multiple of VEC_SIZE, we shift back of VEC_SIZE_LEFTOVER elements to compute the leftover elements for get_global_id(0) == 0
// Note: If C is less than VEC_SIZE, VEC_SIZE should be SHRINKED to the closest smaller VEC_SIZE. This operation is performed on the host side
- int offset_c = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) * sizeof(DATA_TYPE);
+ int offset_c = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0) * sizeof(DATA_TYPE);
int idx_out_w = get_global_id(1);
#if DST_BATCH_SIZE != 1
// If batch size != 1, the batch size dimension is collapsed over the height dimension
int idx_out_h = get_global_id(2) % DST_HEIGHT;
int idx_out_n = get_global_id(2) / DST_HEIGHT;
-#else //DST_BATCH_SIZE != 1
- int idx_out_h = get_global_id(2);
- int idx_out_n = 0;
+#else //DST_BATCH_SIZE != 1
+ int idx_out_h = get_global_id(2);
+ int idx_out_n = 0;
#endif // DST_BATCH_SIZE != 1
- int idx_in_w = idx_out_w * STRIDE_X - PAD_X;
- int idx_in_h = idx_out_h * STRIDE_Y - PAD_Y;
+ int idx_in_w = idx_out_w * STRIDE_X - PAD_X;
+ int idx_in_h = idx_out_h * STRIDE_Y - PAD_Y;
- __global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes +
- offset_c +
- idx_out_n * input_stride_w;
+ __global unsigned char *in_base_ptr = input_ptr + input_offset_first_element_in_bytes + offset_c + idx_out_n * input_stride_w;
- __global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes +
- offset_c +
- idx_out_w * output_stride_y +
- idx_out_h * output_stride_z +
- idx_out_n * output_stride_w;
+ __global unsigned char *out_base_ptr = output_ptr + output_offset_first_element_in_bytes + offset_c + idx_out_w * output_stride_y + idx_out_h * output_stride_z + idx_out_n * output_stride_w;
int pool_x_s = max((int)0, -idx_in_w);
int pool_x_e = min((int)POOL_SIZE_X, (int)SRC_WIDTH - idx_in_w);
@@ -230,7 +224,7 @@ __kernel void pooling_layer_MxN_quantized_nhwc(
int filter_size = 0;
#elif defined(POOL_AVG) && !defined(EXCLUDE_PADDING) // defined(POOL_AVG) && defined(EXCLUDE_PADDING)
int filter_size = POOL_SIZE_X * POOL_SIZE_Y;
-#endif // defined(POOL_AVG) && !defined(EXCLUDE_PADDING)
+#endif // defined(POOL_AVG) && !defined(EXCLUDE_PADDING)
VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
res0 = INITIAL_VALUE;
@@ -239,10 +233,12 @@ __kernel void pooling_layer_MxN_quantized_nhwc(
{
for(int x = pool_x_s; x < pool_x_e; ++x)
{
- VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) data;
- VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE) data0;
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ data;
+ VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE)
+ data0;
- data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + (x + idx_in_w) * input_stride_y + (y + idx_in_h) * input_stride_z));
+ data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(in_base_ptr + (x + idx_in_w) * input_stride_y + (y + idx_in_h) * input_stride_z));
data0 = CONVERT(data, VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE));
res0 = POOL_OP(res0, data0);
@@ -257,7 +253,8 @@ __kernel void pooling_layer_MxN_quantized_nhwc(
res0 = (res0 + (VEC_DATA_TYPE(ACC_DATA_TYPE, VEC_SIZE))(filter_size >> 1)) / filter_size;
#endif // defined(POOL_AVG)
- VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
+ VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+ out_q0 = CONVERT(res0, VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE));
#if defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT)
REQUANTIZE(VEC_SIZE, out_q0, OFFSET_IN1, OFFSET_OUT, SCALE_IN1, SCALE_OUT, out_q0);
#endif /* defined(OFFSET_IN1) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_OUT) */
@@ -265,5 +262,5 @@ __kernel void pooling_layer_MxN_quantized_nhwc(
// Store result
STORE_VECTOR_SELECT(out_q, DATA_TYPE, out_base_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, ((VEC_SIZE_LEFTOVER != 0) && get_global_id(0) == 0));
}
-#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(SELECT_DATA_TYPE) && defined(ACC_DATA_TYPE)
+#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DST_CHANNELS) && defined(DST_HEIGHT) && defined(DST_BATCH_SIZE) && defined(ACC_DATA_TYPE)
#endif // defined(DATA_TYPE) && defined(INITIAL_VALUE) \ No newline at end of file