COMPMID-2600: Implement a new and generic depthwise convolution for CL QASYMM8 NHWC
The NCHW case is supported at function level by permuting the
inputs/outputs to NHWC.
This patch also removes CLDirectConvolutionLayerOutputStageKernel which
is deprecated and some kernels which were only used in the generic case
of depthwise convolution.
Change-Id: I91e0f02d0a2f4a4a352e08c248e648944137fe68
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2056
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index fa5193f..7b7263f 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -222,6 +222,7 @@
{ "depthwise_convolution_3x3_nhwc", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_nhwc_stride1", "depthwise_convolution.cl" },
{ "dwc_MxN_native_fp_nhwc", "depthwise_convolution.cl" },
+ { "dwc_MxN_native_quantized8_nhwc", "depthwise_convolution_quantized.cl" },
{ "dwc_3x3_native_qasymm8_nchw", "depthwise_convolution_quantized.cl" },
{ "dwc_3x3_native_qasymm8_dot8_nchw", "depthwise_convolution_quantized.cl" },
{ "dwc_3x3_reshaped_qasymm8_nhwc", "depthwise_convolution_quantized.cl" },
@@ -234,9 +235,6 @@
{ "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32", "depthwise_convolution.cl" },
{ "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32", "depthwise_convolution.cl" },
{ "depthwise_convolution_reshape_weights", "depthwise_convolution.cl" },
- { "depthwise_convolution_reshape_weights_generic", "depthwise_convolution.cl" },
- { "depthwise_im2col", "depthwise_convolution.cl" },
- { "depthwise_vector_to_tensor", "depthwise_convolution.cl" },
{ "dequantization_layer", "dequantization_layer.cl" },
{ "dequantization_layer_per_channel_nhwc", "dequantization_layer.cl" },
{ "dequantization_layer_per_channel_nchw", "dequantization_layer.cl" },
@@ -413,7 +411,6 @@
{ "NV21_to_RGB888_bt709", "color_convert.cl" },
{ "NV21_to_RGBA8888_bt709", "color_convert.cl" },
{ "NV21_to_YUV444_bt709", "color_convert.cl" },
- { "output_stage_quantized", "direct_convolution_1x1_3x3_5x5_quantized.cl" },
{ "pad_layer_constant", "pad_layer.cl" },
{ "pad_layer_symmetric_reflect", "pad_layer.cl" },
{ "permute", "permute.cl" },
diff --git a/src/core/CL/cl_kernels/activation_quant_helpers.h b/src/core/CL/cl_kernels/activation_quant_helpers.h
index 402e7ac..0e4eb2b 100644
--- a/src/core/CL/cl_kernels/activation_quant_helpers.h
+++ b/src/core/CL/cl_kernels/activation_quant_helpers.h
@@ -41,7 +41,7 @@
// Bounded RELU Activation
inline TYPE brelu_op(TYPE x)
{
- return min((TYPE)A_VAL, max(CONST_0, x));
+ return min((TYPE)A_VAL, max((TYPE)CONST_0, x));
}
// Lower Upper Bounded RELU Activation
inline TYPE lu_brelu_op(TYPE x)
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index 1b2f5cc..3a22728 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -782,173 +782,6 @@
}
#endif // defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DST_WIDTH)
-#if defined(NCHW)
-#define in_stride_x src_stride_x
-#define in_stride_y src_stride_y
-#define in_stride_z src_stride_z
-#define out_stride_x dst_stride_x
-#define out_stride_y dst_stride_y
-#define out_stride_z dst_stride_z
-#else //defined(NCHW)
-#define in_stride_x src_stride_y
-#define in_stride_y src_stride_z
-#define in_stride_z src_stride_x
-#define out_stride_x dst_stride_y
-#define out_stride_y dst_stride_z
-#define out_stride_z dst_stride_x
-#endif //defined(NCHW)
-
-#if defined(SRC_WIDTH) && defined(DATA_TYPE)
-/** This kernel reshapes each of the tensor's low three dimensions to single rows.
- *
- * @note Datatype and source width should be given as a preprocessor argument using -DDATA_TYPE=type and -DSRC_WIDTH=width. e.g. -DSRC_WIDTH=128
- *
- * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
- * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr Pointer to the destination tensor. Same as @p src_ptr
- * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: F16/F32
- * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
- */
-__kernel void depthwise_convolution_reshape_weights_generic(
- TENSOR3D_DECLARATION(src),
- IMAGE_DECLARATION(dst)
-#ifdef HAS_BIAS
- ,
- VECTOR_DECLARATION(biases)
-#endif /* HAS_BIAS */
-)
-{
-#ifdef HAS_BIAS
- Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
-#endif /* HAS_BIAS */
-
- __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z;
- __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;
-
- for(int i = 0; i < SRC_WIDTH; ++i, input_ptr += in_stride_x)
- {
- *((__global DATA_TYPE *)(output_ptr + i * dst_stride_x)) = *((__global DATA_TYPE *)input_ptr);
- }
-
-#if defined(HAS_BIAS)
- if(get_global_id(1) == 0)
- {
- *((__global DATA_TYPE *)(output_ptr + SRC_WIDTH * get_global_size(1) * dst_stride_x)) = *((__global DATA_TYPE *)(biases.ptr + get_global_id(2) * biases_stride_x));
- }
-#endif // defined(HAS_BIAS)
-}
-#endif //defined(SRC_WIDTH) && defined(DATA_TYPE)
-
-#if defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER) && defined(DILATION_X) && defined(DILATION_Y)
-/** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication.
- *
- * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The convolution information must be passed at compile time using -DSTRIDE_X, -DSTRIDE_Y, -DPAD_LEFT, -DPAD_TOP, -DPAD_RIGHT, -DPAD_BOTTOM, -DKERNEL_WIDHT, -DKERNEL_HEIGHT, -DSRC_WIDTH, -DSRC_HEIGHT, -DDEPTH_MULTIPLIER
- * @note The dilation_x and dilation_y must be passed at compile time using -DDILATION_X and -DDILATION_Y: e.g. -DDILATION_X=1, -DDILATION_Y=1
- *
- * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
- * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
- * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(dst))
-{
- Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
-
- const int src_pixel_linear = get_global_id(1) * STRIDE_X;
- const int full_length = SRC_WIDTH + PAD_LEFT + PAD_RIGHT;
- const int max_initial_x = STRIDE_X * (((full_length - (KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1))) / STRIDE_X) + 1);
-
- const int src_x = -PAD_LEFT + src_pixel_linear % max_initial_x;
- const int src_y = -PAD_TOP + src_pixel_linear / max_initial_x * STRIDE_Y;
- const int src_z = get_global_id(2) / DEPTH_MULTIPLIER;
-
- __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + src_z * in_stride_z;
- __global DATA_TYPE *output_ptr = ((__global DATA_TYPE *)(dst.ptr));
-
- for(int y = src_y; y < src_y + KERNEL_HEIGHT + (KERNEL_HEIGHT - 1) * (DILATION_Y - 1); y += DILATION_Y)
- {
- for(int x = src_x; x < src_x + KERNEL_WIDTH + (KERNEL_WIDTH - 1) * (DILATION_X - 1); x += DILATION_X, ++output_ptr)
- {
- if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
- {
- *output_ptr = PAD_VALUE;
- }
- else
- {
- *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * in_stride_x + y * in_stride_y));
- }
- }
- }
-#if defined(HAS_BIAS)
- *output_ptr = (DATA_TYPE)(1);
-#endif // defined(HAS_BIAS)
-}
-
-#endif //defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(SRC_WIDTH) && defined(DATA_TYPE) && defined(PAD_VALUE) && defined(DEPTH_MULTIPLIER)
-
-#if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
-
-/** This kernel performs a reshaping of the output of the depthwise generic convolution.
- *
- * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
- * @note The convolution information must be passed at compile time using -DCONV_WIDTH, -DCONV_HEIGHT, e.g -DCONV_WIDTH=32, -DCONV_HEIGHT=42
- *
- * @param[in] src_ptr Pointer to the source tensor. Supported data types: F16/F32
- * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
- * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
- * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
- * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- */
-__kernel void depthwise_vector_to_tensor(
- VECTOR_DECLARATION(src),
- TENSOR3D_DECLARATION(dst))
-{
- Vector src = CONVERT_TO_VECTOR_STRUCT(src);
-
- const int patch_size = CONV_WIDTH * CONV_HEIGHT;
- const int id0 = get_global_id(0);
- const int z = id0 / patch_size;
- const int index2D = id0 - z * patch_size;
-
- __global uchar *out_ptr = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * out_stride_x + index2D / CONV_WIDTH * out_stride_y + z * out_stride_z;
- *((__global DATA_TYPE *)out_ptr) = *((__global DATA_TYPE *)src.ptr);
-}
-
-#endif //defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
-
#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(DEPTH_MULTIPLIER) && defined(DST_CHANNELS) && defined(IS_F16)
#if defined(CONV_STRIDE_X)
#if CONV_STRIDE_X == 1
@@ -1478,7 +1311,7 @@
//3x3 Convolution of elements starting in 0th row
pixels0 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 0, weights_addr, weights_stride_y);
//3x3 Convolution of elements starting in 2nd row
- pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
+ pixels1 = convolution_3x3_dilation_stridex2_stridey2_bifrost_f16(src_addr, src.stride_x, src.stride_y, 2, weights_addr, weights_stride_y);
#endif /* DILATION_X==1 && DILATION_Y==1 */
#ifdef HAS_BIAS
@@ -1556,23 +1389,17 @@
int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
int b = get_global_id(2) / (int)DST_DEPTH; // batch
#else // defined(DST_DEPTH)
- int z = get_global_id(2); // spatial coordinate y
+ int z = get_global_id(2); // spatial coordinate y
#endif // defined(DST_DEPTH)
- __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes +
- x * sizeof(DATA_TYPE) * (int)N0;
+ __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)N0;
- __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes +
- x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 +
- y * dst_stride_y +
- z * dst_stride_z;
+ __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z;
- __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes +
- x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+ __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
#if defined(HAS_BIAS)
- __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes +
- x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
+ __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(DATA_TYPE) * (int)DEPTH_MULTIPLIER * (int)N0;
#endif // defined(HAS_BIAS)
#if defined(DST_DEPTH)
@@ -1611,7 +1438,7 @@
#if GPU_ARCH == GPU_ARCH_MIDGARD
res += i * w;
#else // GPU_ARCH == GPU_ARCH_MIDGARD
- res = fma(i, w, res);
+ res = fma(i, w, res);
#endif // GPU_ARCH == GPU_ARCH_MIDGARD
}
x_coord_tmp += DILATION_X;
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index 8f2e441..10872d4 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -24,19 +24,30 @@
#include "helpers_asymm.h"
-#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
+#ifndef VEC_SIZE
+#if defined(N0)
+#define VEC_SIZE N0
+#else /* defined(N0) */
+#define VEC_SIZE 8
+#endif /* defined(N0) */
+#endif /* VEC_SIZE */
#if defined(ACTIVATION_TYPE) && defined(CONST_0)
#define DATA_TYPE uchar
-#ifndef VEC_SIZE
-#define VEC_SIZE 8
-#endif /* VEC_SIZE */
#include "activation_layer_quant.cl"
#define ACTIVATION_FUNC(x) PERFORM_ACTIVATION_QUANT(ACTIVATION_TYPE, x)
#else /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
#define ACTIVATION_FUNC(x) (x)
#endif /* defined(ACTIVATION_TYPE) && defined(CONST_0) */
+#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
+#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
+#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
+#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE)
+#define VEC_SHORT VEC_DATA_TYPE(short, VEC_SIZE)
+
+#if defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
+
#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)
#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)
#define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), val);
@@ -635,11 +646,6 @@
#define asymm_mult_by_quant_multiplier_less_than_one(x, y, z) ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, y, z, VEC_SIZE)
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
-#define VEC_UCHAR VEC_DATA_TYPE(uchar, VEC_SIZE)
-#define VEC_USHORT VEC_DATA_TYPE(ushort, VEC_SIZE)
-
#define MULTIPLY_ADD(x, y, acc) acc += CONVERT(CONVERT(x, VEC_USHORT) * CONVERT(y, VEC_USHORT), VEC_INT)
#if WEIGHTS_OFFSET != 0
@@ -1375,3 +1381,144 @@
#endif // defined(VEC_SIZE) && defined(SRC_DIM_1) && defined(SRC_DIM_2) && defined(CONV_PAD_TOP) && defined(CONV_PAD_LEFT)
#endif // defined(WEIGHTS_OFFSET) && defined(INPUT_OFFSET) && defined(K_OFFSET) && ((defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)) || defined(REAL_MULTIPLIER))
+
+#if defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
+/** This function computes the depthwise convolution for NHWC data layout. This kernel assumes that the weights tensor is NOT reshaped
+ *
+ * @note The number of elements processed must be passed at compile time using -DN0 (e.g. -DN0=2)
+ * @note The depth multiplier must be passed at compile time using -DDEPTH_MULTIPLIER (e.g. -DDEPTH_MULTIPLIER=1)
+ * @note The first dimension of the input tensor must be passed at compile time using -DSRC_DIM1 (e.g. -DSRC_DIM1=112)
+ * @note The second dimension of the input tensor must be passed at compile time using -DSRC_DIM2 (e.g. -DSRC_DIM2=80)
+ * @note The kernel width must be passed at compile time using -DKERNEL_WIDTH (e.g. -DKERNEL_WIDTH=5)
+ * @note The kernel height must be passed at compile time using -DKERNEL_HEIGHT (e.g. -DKERNEL_HEIGHT=5)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_TOP (e.g. -DCONV_PAD_TOP=1)
+ * @note The convolution pad top must be passed at compile time using -DCONV_PAD_LEFT (e.g. -DCONV_PAD_LEFT=1)
+ * @note The convolution stride along the width must be passed at compile time using -DCONV_STRIDE_X (e.g. -DCONV_STRIDE_Y=X)
+ * @note The convolution stride along the height must be passed at compile time using -DCONV_STRIDE_Y (e.g. -DCONV_STRIDE_Y=1)
+ * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
+ *
+ * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8
+ * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
+ * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
+ * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z src_stride_y * number of elements along Z processed per workitem(in bytes)
+ * @param[in] src_stride_w Stride of the source tensor in W dimension (in bytes)
+ * @param[in] src_step_w src_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: same as src_ptr
+ * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
+ * @param[in] dst_step_w dst_stride_w * number of elements along W processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in] weights_ptr Pointer to the weights tensor. Supported data types: QASYMM8
+ * @param[in] weights_stride_x Stride of the weights tensor in X dimension (in bytes)
+ * @param[in] weights_step_x weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] weights_stride_y Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in] weights_step_y weights_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_stride_z Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in] weights_step_z weights_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in] biases_ptr (Optional) Pointer to the biases vector. Supported data types: same as src_ptr
+ * @param[in] biases_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
+ */
+__kernel void dwc_MxN_native_quantized8_nhwc(
+ TENSOR4D_DECLARATION(src),
+ TENSOR4D_DECLARATION(dst),
+ TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+ VECTOR_DECLARATION(biases)
+#endif // defined(HAS_BIAS)
+)
+{
+ int x = get_global_id(0); // channels
+ int y = get_global_id(1); // spatial coordinate x
+#if defined(DST_DEPTH)
+ int z = get_global_id(2) % (int)DST_DEPTH; // spatial coordinate y
+ int b = get_global_id(2) / (int)DST_DEPTH; // batch
+#else // defined(DST_DEPTH)
+ int z = get_global_id(2); // spatial coordinate y
+#endif // defined(DST_DEPTH)
+
+ __global uchar *s_addr = src_ptr + src_offset_first_element_in_bytes + x * sizeof(uchar) * (int)N0;
+
+ __global uchar *d_addr = dst_ptr + dst_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0 + y * dst_stride_y + z * dst_stride_z;
+
+ __global uchar *w_addr = weights_ptr + weights_offset_first_element_in_bytes + x * sizeof(uchar) * (int)DEPTH_MULTIPLIER * (int)N0;
+
+#if defined(HAS_BIAS)
+ __global uchar *b_addr = biases_ptr + biases_offset_first_element_in_bytes + x * sizeof(int) * (int)DEPTH_MULTIPLIER * (int)N0;
+#endif // defined(HAS_BIAS)
+
+#if defined(DST_DEPTH)
+ s_addr += b * src_stride_w;
+ d_addr += b * dst_stride_w;
+#endif // defined(DST_DEPTH)
+
+#if DEPTH_MULTIPLIER > 1
+ for(int d = 0; d < (int)DEPTH_MULTIPLIER; ++d)
+ {
+#endif // DEPTH_MULTIPLIER > 1
+ // Each work-item computes N0x1x1 elements
+ VEC_SHORT res = 0;
+
+ int x_coord = y * CONV_STRIDE_X - (int)CONV_PAD_LEFT;
+ int y_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;
+
+ for(int yk = 0; yk < KERNEL_HEIGHT; ++yk)
+ {
+ if(y_coord >= 0 && y_coord < SRC_DIM2)
+ {
+ int x_coord_tmp = x_coord;
+
+ for(int xk = 0; xk < KERNEL_WIDTH; ++xk)
+ {
+ if(x_coord_tmp >= 0 && x_coord_tmp < SRC_DIM1)
+ {
+ int s_offset = x_coord_tmp * (int)src_stride_y + y_coord * (int)src_stride_z;
+ int w_offset = xk * weights_stride_y + yk * weights_stride_z;
+
+ // Load input and weights values
+ VEC_SHORT i = CONVERT(VLOAD(N0)(0, (__global uchar *)(s_addr + s_offset)), VEC_SHORT);
+ VEC_SHORT w = CONVERT(VLOAD(N0)(0, (__global uchar *)(w_addr + w_offset)), VEC_SHORT);
+
+ res += (i + (VEC_SHORT)INPUT_OFFSET) * (w + (VEC_SHORT)WEIGHTS_OFFSET);
+ }
+ x_coord_tmp += DILATION_X;
+ }
+ }
+ y_coord += DILATION_Y;
+ }
+
+#if defined(HAS_BIAS)
+ VEC_SHORT bias = CONVERT(VLOAD(N0)(0, (__global int *)(b_addr)), VEC_SHORT);
+ res += bias;
+#endif // defined(HAS_BIAS)
+
+ res = CONVERT(ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(CONVERT(res, VEC_INT), OUTPUT_MULTIPLIER, OUTPUT_SHIFT, N0), VEC_SHORT);
+ res += (VEC_SHORT)OUTPUT_OFFSET;
+
+ VEC_UCHAR res1 = CONVERT_SAT(res, VEC_UCHAR);
+
+ VSTORE(N0)
+ (ACTIVATION_FUNC(res1), 0, (__global uchar *)(d_addr));
+
+#if DEPTH_MULTIPLIER > 1
+ w_addr += sizeof(uchar);
+ d_addr += sizeof(uchar);
+#if defined(HAS_BIAS)
+ b_addr += sizeof(int);
+#endif // defined(HAS_BIAS)
+ }
+#endif // DEPTH_MULTIPLIER > 1
+}
+#endif // defined(SRC_DIM1) && defined(SRC_DIM2) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defiend(N0) && defined(DILATION_X) && defined(DILATION_Y) && defined(CONV_STRIDE_X) && defined(CONV_STRIDE_Y) && defined(CONV_PAD_LEFT) && defined(CONV_PAD_TOP) && defined(INPUT_OFFSET) && defined(WEIGHTS_OFFSET) && defined(OUTPUT_OFFSET) && defined(OUTPUT_MULTIPLIER) && defined(OUTPUT_SHIFT)
diff --git a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
index 83da767..5ad9afb 100644
--- a/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
+++ b/src/core/CL/cl_kernels/direct_convolution_1x1_3x3_5x5_quantized.cl
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -247,79 +247,3 @@
vstore8(convert_uchar8_sat(pixels0), 0, (__global uchar *)dst.ptr);
}
#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)
-
-#if defined(VEC_SIZE)
-
-#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE)
-#define CONVERT_SAT_UCHAR_STR(x, size) (convert_uchar##size##_sat((x)))
-#define CONVERT_SAT_UCHAR(x, size) CONVERT_SAT_UCHAR_STR(x, size)
-
-/** This function computes the output stage of a depthwise convolution.
- *
- * @param[in] src_ptr Pointer to the source image. Supported data types: QASYMM8
- * @param[in] src_stride_x Stride of the source image in X dimension (in bytes)
- * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] src_stride_y Stride of the source image in Y dimension (in bytes)
- * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source image
- * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
- * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_ptr Pointer to the destination tensor. Supported data types: QASYMM8
- * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
- * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
- * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
- * @param[in] dst_step_z dst_stride_z * number of elements along Y processed per workitem(in bytes)
- * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
- * @param[in] bias_ptr (Optional) Pointer to the biases vector. Supported data types: S32
- * @param[in] bias_stride_x (Optional) Stride of the biases vector in X dimension (in bytes)
- * @param[in] bias_step_x (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
- * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
- * @param[in] output_offset Quantized offset of zero point of the output tensor data range
- * @param[in] output_multiplier Output scale multiplier
- * @param[in] output_shift Output scale divisor exponent
- */
-__kernel void output_stage_quantized(
- TENSOR3D_DECLARATION(src),
- TENSOR3D_DECLARATION(dst),
-#if defined(HAS_BIAS)
- VECTOR_DECLARATION(bias),
-#endif //defined(HAS_BIAS)
- int output_offset,
- int output_multiplier,
- int output_shift)
-{
- Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
- Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
-#if defined(HAS_BIAS)
- Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
-#endif //defined(HAS_BIAS)
-
- // Load input
- VEC_INT vals = VLOAD(VEC_SIZE)(0, (__global int *)(src.ptr));
-
-#if defined(HAS_BIAS)
- // Load and add bias
-#if defined(NCHW)
- int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2))));
-#else // defined(NCHW)
- VEC_INT bias_value = VLOAD(VEC_SIZE)(0, ((__global int *)(vector_offset(&bias, get_global_id(0) * VEC_SIZE))));
-#endif // defined(NCHW)
-
- vals += (VEC_INT)(bias_value);
-#endif //defined(HAS_BIAS)
-
- vals = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(vals, output_multiplier, output_shift, VEC_SIZE);
- vals = vals + output_offset;
-
- // Store result in dst
- VSTORE(VEC_SIZE)
- (CONVERT_SAT_UCHAR(vals, VEC_SIZE), 0, (__global uchar *)dst.ptr);
-}
-
-#undef VEC_INT
-#undef CONVERT_SAT_UCHAR_STR
-#undef CONVERT_SAT_UCHAR
-
-#endif // defined(VEC_SIZE)
diff --git a/src/core/CL/cl_kernels/helpers.h b/src/core/CL/cl_kernels/helpers.h
index f7f2085..8117c1e 100644
--- a/src/core/CL/cl_kernels/helpers.h
+++ b/src/core/CL/cl_kernels/helpers.h
@@ -150,6 +150,28 @@
#define convert_half8_sat convert_half8
#define convert_half16_sat convert_half16
+#define convert_float1 convert_float
+#define convert_half1 convert_half
+#define convert_char1 convert_char
+#define convert_uchar1 convert_uchar
+#define convert_short1 convert_short
+#define convert_ushort1 convert_ushort
+#define convert_int1 convert_int
+#define convert_uint1 convert_uint
+#define convert_long1 convert_long
+#define convert_ulong1 convert_ulong
+#define convert_double1 convert_double
+
+#define convert_char1_sat convert_char_sat
+#define convert_uchar1_sat convert_uchar_sat
+#define convert_short1_sat convert_short_sat
+#define convert_ushort1_sat convert_ushort_sat
+#define convert_int1_sat convert_int_sat
+#define convert_uint1_sat convert_uint_sat
+#define convert_long1_sat convert_long_sat
+#define convert_ulong1_sat convert_ulong_sat
+#define convert_double1_sat convert_double_sat
+
#define VEC_DATA_TYPE_STR(type, size) type##size
#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h
index 53e6719..57ecccc 100644
--- a/src/core/CL/cl_kernels/helpers_asymm.h
+++ b/src/core/CL/cl_kernels/helpers_asymm.h
@@ -381,11 +381,13 @@
DEQUANTIZE_IMPL(ushort, 4)
DEQUANTIZE_IMPL(short, 4)
+ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
+ASYMM_MULT_IMPL(1)
ASYMM_MULT_IMPL(2)
ASYMM_MULT_IMPL(4)
ASYMM_MULT_IMPL(8)
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
index 8b624bb..2115fc6 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp
@@ -35,6 +35,7 @@
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
namespace arm_compute
{
@@ -46,7 +47,7 @@
ARM_COMPUTE_UNUSED(dwc_info);
ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1 && dwc_weights_info.n0 != 1);
ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
@@ -59,8 +60,16 @@
if(biases != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+
+ if(is_data_type_quantized(input->data_type()))
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+ }
}
if(output->total_size() != 0)
@@ -137,6 +146,7 @@
const size_t idx_h = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT);
const size_t weights_width = weights->info()->dimension(idx_w);
const size_t weights_height = weights->info()->dimension(idx_h);
+ const bool is_quantized = is_data_type_quantized(input->info()->data_type());
CLBuildOptions build_opts;
build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
@@ -155,10 +165,46 @@
build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
- build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
- build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
- std::string kernel_name("dwc_MxN_native_fp_nhwc");
+ std::string kernel_name = (is_quantized) ? "dwc_MxN_native_quantized8_nhwc" : "dwc_MxN_native_fp_nhwc";
+
+ if(is_quantized)
+ {
+ const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
+ const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
+ const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
+
+ float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
+ int output_multiplier = 0;
+ int output_shift = 0;
+ quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+ build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
+ build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
+ build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
+ build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+ build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+
+ if(dwc_info.activation_info.enabled())
+ {
+ const int a_val = quantize_qasymm8(dwc_info.activation_info.a(), oq_info);
+ const int b_val = quantize_qasymm8(dwc_info.activation_info.b(), oq_info);
+ const int o1 = oq_info.offset;
+
+ build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
+ build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
+ build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
+
+ const float s1 = iq_info.scale;
+ build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
+ build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
+ }
+ }
+ else
+ {
+ build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.a()));
+ build_opts.add_option_if(dwc_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(dwc_info.activation_info.b()));
+ }
ICLKernel::configure_internal(win_config.second);
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
deleted file mode 100644
index 582c600..0000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
+++ /dev/null
@@ -1,147 +0,0 @@
-/*
- * Copyright (c) 2017-2019 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "support/ToolchainSupport.h"
-
-using namespace arm_compute;
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
-{
- const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
- const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
- const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && (biases != nullptr));
- ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_c) != output->dimension(1));
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (input->dimension(idx_w) * input->dimension(idx_h) + ((biases != nullptr) ? 1 : 0)));
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-
- if(biases != nullptr)
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
- ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != input->dimension(idx_c));
- ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
- }
-
- return Status{};
-}
-} // namespace
-
-CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel()
- : _input(nullptr), _biases(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), (biases != nullptr) ? biases->info() : nullptr));
-
- _input = input;
- _biases = biases;
- _output = output;
-
- const size_t idx_w = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH);
-
- // Create kernel
- std::set<std::string> build_opts;
-
- build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
- build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
- build_opts.emplace("-D" + string_from_data_layout(input->info()->data_layout()));
- if(_biases != nullptr)
- {
- build_opts.emplace("-DHAS_BIAS");
- }
-
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_reshape_weights_generic", build_opts));
-
- // Configure kernel window
- Window win = calculate_max_window(*input->info(), Steps());
- // The CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel doesn't need padding so update_window_and_padding() can be skipped
- output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
- ICLKernel::configure_internal(win);
-}
-
-Status CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
-{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, biases));
- return Status{};
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::run(const Window &window, cl::CommandQueue &queue)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
- Window slice = window.first_slice_window_3D();
- Window slice_out = window.first_slice_window_2D();
-
- const size_t idx_w = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::WIDTH);
- const size_t idx_h = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::HEIGHT);
- const size_t idx_c = get_data_layout_dimension_index(_input->info()->data_layout(), DataLayoutDimension::CHANNEL);
-
- // Setup slice
- slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(idx_w), _input->info()->dimension(idx_w)));
- slice.set(Window::DimY, Window::Dimension(0, _input->info()->dimension(idx_h), 1));
- slice.set(Window::DimZ, Window::Dimension(0, _input->info()->dimension(idx_c), 1));
-
- // Setup output slice
- // The first two dimensions of the output are increased by the inner loops
- slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
- slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
-
- // Set biases
- if(_biases != nullptr)
- {
- unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
- Window slice_biases;
- slice_biases.use_tensor_dimensions(_biases->info()->tensor_shape());
- add_1D_tensor_argument(idx, _biases, slice_biases);
- }
-
- do
- {
- unsigned int idx = 0;
- add_3D_tensor_argument(idx, _input, slice);
- add_2D_tensor_argument(idx, _output, slice_out);
- enqueue(queue, *this, slice, lws_hint());
- }
- while(window.slide_window_slice_3D(slice) && window.slide_window_slice_2D(slice_out));
-}
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
deleted file mode 100644
index 0312a57..0000000
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ /dev/null
@@ -1,148 +0,0 @@
-/*
- * Copyright (c) 2017-2019 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "support/ToolchainSupport.h"
-
-#include <tuple>
-
-using namespace arm_compute;
-
-CLDepthwiseIm2ColKernel::CLDepthwiseIm2ColKernel()
- : _input(nullptr), _output(nullptr)
-{
-}
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
- const Size2D &dilation)
-{
- const size_t idx_c = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
-
- ARM_COMPUTE_UNUSED(conv_info);
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(input->data_type()) && has_bias);
- ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(idx_c) * depth_multiplier) != output->dimension(2));
- ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
- ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || dilation.y() < 1);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
-
- return Status{};
-}
-} // namespace
-
-void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
- const Size2D &dilation)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
-
- _input = input;
- _output = output;
-
- const DataLayout data_layout = input->info()->data_layout();
- const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
- const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
- const UniformQuantizationInfo qinfo = input->info()->quantization_info().uniform();
-
- // Create kernel
- CLBuildOptions build_opts;
-
- build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
- build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
- build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
- build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
- build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
- build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
- build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
- build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(idx_w)));
- build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(idx_h)));
- build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
- build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
- build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(depth_multiplier));
- build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
- build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
- build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
- build_opts.add_option_if(has_bias, "-DHAS_BIAS");
- build_opts.add_option_if_else(is_data_type_quantized_asymmetric(input->info()->data_type()),
- "-DPAD_VALUE=" + support::cpp11::to_string(qinfo.offset),
- "-DPAD_VALUE=0");
-
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts.options()));
-
- // Configure kernel window
- Window win = calculate_max_window(*output->info(), Steps());
- // CLDepthwiseIm2ColKernel doesn't need padding so update_window_and_padding() can be skipped
- output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
- ICLKernel::configure_internal(win);
-}
-
-Status CLDepthwiseIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier,
- const Size2D &dilation)
-{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, depth_multiplier, dilation));
-
- return Status{};
-}
-
-void CLDepthwiseIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
- Window slice = window.first_slice_window_3D();
- Window slice_in = window.first_slice_window_3D();
-
- // Setup slice
- slice.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _output->info()->dimension(0)));
- slice.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
- slice.set(Window::DimZ, Window::Dimension(0, _output->info()->dimension(2), 1));
-
- // Setup input slice
- // The first three dimensions of the input are increased by the inner loops
- slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
- slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
- slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
- do
- {
- unsigned int idx = 0;
- add_3D_tensor_argument(idx, _input, slice_in);
- add_3D_tensor_argument(idx, _output, slice);
- enqueue(queue, *this, slice, lws_hint());
- }
- while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in));
-}
diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
deleted file mode 100644
index 0f029fd..0000000
--- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
+++ /dev/null
@@ -1,125 +0,0 @@
-/*
- * Copyright (c) 2017-2019 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/CL/OpenCL.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "support/ToolchainSupport.h"
-
-using namespace arm_compute;
-using namespace arm_compute::misc::shape_calculator;
-
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h)
-{
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16, DataType::F32);
-
- if(output->total_size() != 0)
- {
- TensorShape output_shape = compute_vector_to_tensor_output_shape(input->tensor_shape(), conv_w, conv_h, output->data_layout());
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
- }
-
- return Status{};
-}
-} // namespace
-
-CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel()
- : _input(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
- // Output auto inizialitation if not yet initialized
- TensorShape output_shape = compute_vector_to_tensor_output_shape(input->info()->tensor_shape(), conv_w, conv_h, output->info()->data_layout());
- auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
-
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), conv_w, conv_h));
-
- _input = input;
- _output = output;
-
- // Create kernel
- CLBuildOptions build_opts;
- build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
- build_opts.add_option("-DCONV_WIDTH=" + support::cpp11::to_string(conv_w));
- build_opts.add_option("-DCONV_HEIGHT=" + support::cpp11::to_string(conv_h));
- build_opts.add_option("-D" + string_from_data_layout(output->info()->data_layout()));
-
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_vector_to_tensor", build_opts.options()));
-
- // Configure kernel window
- Window win = calculate_max_window(*input->info(), Steps());
- // The CLDepthwisevectorToTensorKernel doesn't need padding so update_window_and_padding() can be skipped
- output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
- ICLKernel::configure_internal(win);
-}
-
-Status CLDepthwiseVectorToTensorKernel::validate(const ITensorInfo *input, const ITensorInfo *output, size_t conv_w, size_t conv_h)
-{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, conv_w, conv_h));
- return Status{};
-}
-
-void CLDepthwiseVectorToTensorKernel::run(const Window &window, cl::CommandQueue &queue)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
- Window slice = window.first_slice_window_1D();
- Window slice_out = window.first_slice_window_3D();
-
- // Setup slice
- slice.set(Window::DimX, Window::Dimension(0, _input->info()->dimension(0), 1));
-
- // Setup output slice
- // The first three dimensions of the output are increased by the inner loops
- slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
- slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
- slice_out.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
- do
- {
- unsigned int idx = 0;
- add_1D_tensor_argument(idx, _input, slice);
- add_3D_tensor_argument(idx, _output, slice_out);
- enqueue(queue, *this, slice, lws_hint());
- }
- while(window.slide_window_slice_1D(slice) && window.slide_window_slice_3D(slice_out));
-}
diff --git a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
deleted file mode 100644
index 515321c..0000000
--- a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
+++ /dev/null
@@ -1,209 +0,0 @@
-/*
- * Copyright (c) 2018-2019 ARM Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
-
-#include "arm_compute/core/AccessWindowStatic.h"
-#include "arm_compute/core/CL/CLValidate.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Error.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/Types.h"
-#include "arm_compute/core/Window.h"
-
-#include <cstddef>
-#include <cstdint>
-
-namespace arm_compute
-{
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
-{
- ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32, DataType::F16,
- DataType::F32);
-
- if(bias != nullptr)
- {
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(bias);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32, DataType::F16, DataType::F32);
-
- if(is_data_type_quantized_asymmetric(input->data_type()))
- {
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
- }
-
- ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()),
- "Calling output stage kernel with floating point arguments");
- }
-
- // Checks performed on output
- if(input->data_type() == DataType::S32)
- {
- // Quantized configuration checks
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
- }
- else
- {
- // In case of out-of-place computation (supported for non-quantized configurations)
- if((output != nullptr) && (output->total_size() != 0))
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- }
- }
-
- return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
-{
- bool window_changed = false;
- const unsigned int num_elems_processed_per_iteration = 16 / element_size_from_data_type(input->data_type());
-
- // Configure kernel window
- Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
-
- // Input window
- AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
- window_changed = window_changed || update_window_and_padding(win, input_access);
-
- // Bias window
- if(bias != nullptr)
- {
- AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->dimension(1));
- window_changed = window_changed || update_window_and_padding(win, bias_access);
- }
-
- // Output window
- if(output != nullptr && (output->total_size() != 0))
- {
- AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
- window_changed = window_changed || update_window_and_padding(win, output_access);
- output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
- }
- else
- {
- input_access.set_valid_region(win, ValidRegion(Coordinates(), input->tensor_shape()));
- }
-
- Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
- return std::make_pair(err, win);
-}
-} // namespace
-
-CLDirectConvolutionLayerOutputStageKernel::CLDirectConvolutionLayerOutputStageKernel()
- : _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0)
-{
-}
-
-void CLDirectConvolutionLayerOutputStageKernel::configure(ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
- int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input);
-
- // Auto-initialize output if required
- if(output != nullptr)
- {
- // Work out expected output data type
- const DataType output_dt = (input->info()->data_type() == DataType::S32) ? DataType::QASYMM8 : input->info()->data_type();
- // Output tensor auto initialization if not yet initialized
- auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_dt));
- }
-
- // Perform validation step
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info()));
-
- _bias = bias;
- _input = input;
- _output = output;
- _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
- _result_shift = result_shift;
- _result_offset_after_shift = result_offset_after_shift;
-
- const unsigned int num_elems_accessed_per_iteration = 16 / element_size_from_data_type(input->info()->data_type());
-
- // Create kernel
- CLBuildOptions build_opts;
- build_opts.add_option_if(bias != nullptr, "-DHAS_BIAS");
- build_opts.add_option("-D" + string_from_data_layout(input->info()->data_layout()));
- build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("output_stage_quantized", build_opts.options()));
-
- // Set static kernel arguments
- int idx = 2 * num_arguments_per_3D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
- _kernel.setArg<int>(idx++, _result_offset_after_shift);
- _kernel.setArg<int>(idx++, _result_fixedpoint_multiplier);
- _kernel.setArg<int>(idx++, _result_shift);
-
- // Configure kernel window
- auto win_config = validate_and_configure_window(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info());
- ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
- ICLKernel::configure_internal(win_config.second);
-}
-
-Status CLDirectConvolutionLayerOutputStageKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
-{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output));
- ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), bias == nullptr ? nullptr : bias->clone().get(), output == nullptr ? nullptr : output->clone().get()).first);
-
- return Status{};
-}
-
-void CLDirectConvolutionLayerOutputStageKernel::run(const Window &window, cl::CommandQueue &queue)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-
- Window slice = window.first_slice_window_3D();
-
- // Set bias vector
- if(_bias != nullptr)
- {
- unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
- Window slice_biases;
- slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
- add_1D_tensor_argument(idx1, _bias, slice_biases);
- }
-
- // Run kernel
- do
- {
- // Set arguments
- unsigned int idx = 0;
- add_3D_tensor_argument(idx, _input, slice);
- add_3D_tensor_argument(idx, _output, slice);
- enqueue(queue, *this, slice, lws_hint());
- }
- while(window.slide_window_slice_3D(slice));
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
index 2967a73..ae096f2 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.cpp
@@ -35,8 +35,6 @@
#include "support/ToolchainSupport.h"
-using namespace arm_compute;
-
namespace arm_compute
{
namespace
@@ -101,8 +99,6 @@
} // namespace
class Coordinates;
-} // namespace arm_compute
-
CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel()
: _input(nullptr), _bias(nullptr), _output(nullptr)
{
@@ -177,3 +173,4 @@
}
while(collapsed.slide_window_slice_3D(slice));
}
+} // namespace arm_compute