28 #if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) 31 #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) 33 #define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) 35 #error "STRIDE_X larger than 2 is not supported" 38 #define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \ 40 VEC_DATA_TYPE(DATA_TYPE, 4) \ 41 weights_values0 = vload4(0, weights_row_ptr); \ 42 DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \ 43 VEC_DATA_TYPE(DATA_TYPE, 8) \ 44 src0 = vload8(0, src_row_ptr); \ 45 VEC_DATA_TYPE(DATA_TYPE, 4) \ 46 src1 = vload4(0, src_row_ptr + 8); \ 48 acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ 49 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ 50 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ 51 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ 52 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s45, src0.s67, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ 55 #define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \ 57 VEC_DATA_TYPE(DATA_TYPE, 4) \ 58 weights_values0 = vload4(0, weights_row_ptr); \ 59 DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \ 60 VEC_DATA_TYPE(DATA_TYPE, 16) \ 61 src0 = vload16(0, src_row_ptr); \ 62 VEC_DATA_TYPE(DATA_TYPE, 4) \ 63 src1 = vload4(0, src_row_ptr + 16); \ 64 acc += src0.even * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \ 65 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \ 66 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \ 68 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \ 69 acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \ 108 __kernel
void direct_convolution5x5(
115 unsigned int weights_stride_w)
124 __global uchar *weights_addr = (__global uchar *)
tensor3D_offset(&weights, 0, 0, 0);
125 __global uchar *src_addr = (__global uchar *)
offset(&src, 0, 0);
127 const
int kernel_index = get_global_id(2);
128 weights_addr += kernel_index * weights_stride_w;
130 for(volatile
int d = 0; d < WEIGHTS_DEPTH; ++d)
132 CONVOLUTION1x5(pixels0, (__global
DATA_TYPE *)src_addr, (__global
DATA_TYPE *)weights_addr);
133 CONVOLUTION1x5(pixels0, (__global
DATA_TYPE *)(src_addr + 1 * src_stride_y), (__global
DATA_TYPE *)(weights_addr + 1 * weights_stride_y));
134 CONVOLUTION1x5(pixels0, (__global
DATA_TYPE *)(src_addr + 2 * src_stride_y), (__global
DATA_TYPE *)(weights_addr + 2 * weights_stride_y));
135 CONVOLUTION1x5(pixels0, (__global
DATA_TYPE *)(src_addr + 3 * src_stride_y), (__global
DATA_TYPE *)(weights_addr + 3 * weights_stride_y));
136 CONVOLUTION1x5(pixels0, (__global
DATA_TYPE *)(src_addr + 4 * src_stride_y), (__global
DATA_TYPE *)(weights_addr + 4 * weights_stride_y));
138 src_addr += src_stride_z;
139 weights_addr += weights_stride_z;
150 #endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH) 152 #if defined(WEIGHTS_DEPTH) 154 #define CONVOLUTION1x5_BIFROST(acc, src0, weights_row00, weights_row01) \ 156 acc.s0 = mad(src0.s0, weights_row00.s0, acc.s0); \ 157 acc.s1 = mad(src0.s1, weights_row00.s0, acc.s1); \ 158 acc.s2 = mad(src0.s2, weights_row00.s0, acc.s2); \ 159 acc.s3 = mad(src0.s3, weights_row00.s0, acc.s3); \ 160 acc.s0 = mad(src0.s1, weights_row00.s1, acc.s0); \ 161 acc.s1 = mad(src0.s2, weights_row00.s1, acc.s1); \ 162 acc.s2 = mad(src0.s3, weights_row00.s1, acc.s2); \ 163 acc.s3 = mad(src0.s4, weights_row00.s1, acc.s3); \ 164 acc.s0 = mad(src0.s2, weights_row00.s2, acc.s0); \ 165 acc.s1 = mad(src0.s3, weights_row00.s2, acc.s1); \ 166 acc.s2 = mad(src0.s4, weights_row00.s2, acc.s2); \ 167 acc.s3 = mad(src0.s5, weights_row00.s2, acc.s3); \ 168 acc.s0 = mad(src0.s3, weights_row00.s3, acc.s0); \ 169 acc.s1 = mad(src0.s4, weights_row00.s3, acc.s1); \ 170 acc.s2 = mad(src0.s5, weights_row00.s3, acc.s2); \ 171 acc.s3 = mad(src0.s6, weights_row00.s3, acc.s3); \ 172 acc.s0 = mad(src0.s4, weights_row01, acc.s0); \ 173 acc.s1 = mad(src0.s5, weights_row01, acc.s1); \ 174 acc.s2 = mad(src0.s6, weights_row01, acc.s2); \ 175 acc.s3 = mad(src0.s7, weights_row01, acc.s3); \ 214 __kernel
void direct_convolution5x5_f32_bifrost(
221 unsigned int weights_stride_w)
224 const int kernel_index = get_global_id(2);
229 float4 pixels0 = 0.0f;
230 float4 pixels1 = 0.0f;
232 __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w);
233 __global uchar *src_addr = (__global uchar *)
offset(&src, 0, 0);
237 for(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d)
240 float4 weights_row00 = vload4(0, (__global
float *)(weights_addr + 0 * weights_stride_y));
241 float weights_row01 = *((__global
float *)(weights_addr + 0 * weights_stride_y) + 4);
242 float4 weights_row10 = vload4(0, (__global
float *)(weights_addr + 1 * weights_stride_y));
243 float weights_row11 = *((__global
float *)(weights_addr + 1 * weights_stride_y) + 4);
247 src0 = vload8(0, (__global
float *)(src_addr + 0 * src_stride_y));
250 CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
253 src0 = vload8(0, (__global
float *)(src_addr + 1 * src_stride_y));
256 CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row10, weights_row11);
257 CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
260 src0 = vload8(0, (__global
float *)(src_addr + 2 * src_stride_y));
263 weights_row00 = vload4(0, (__global
float *)(weights_addr + 2 * weights_stride_y));
264 weights_row01 = *((__global
float *)(weights_addr + 2 * weights_stride_y) + 4);
267 CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
268 CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row10, weights_row11);
271 src0 = vload8(0, (__global
float *)(src_addr + 3 * src_stride_y));
274 weights_row10 = vload4(0, (__global
float *)(weights_addr + 3 * weights_stride_y));
275 weights_row11 = *((__global
float *)(weights_addr + 3 * weights_stride_y) + 4);
278 CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row10, weights_row11);
279 CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
282 src0 = vload8(0, (__global
float *)(src_addr + 4 * src_stride_y));
285 weights_row00 = vload4(0, (__global
float *)(weights_addr + 4 * weights_stride_y));
286 weights_row01 = *((__global
float *)(weights_addr + 4 * weights_stride_y) + 4);
288 CONVOLUTION1x5_BIFROST(pixels0, src0, weights_row00, weights_row01);
289 CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row10, weights_row11);
292 src0 = vload8(0, (__global
float *)(src_addr + 5 * src_stride_y));
295 CONVOLUTION1x5_BIFROST(pixels1, src0, weights_row00, weights_row01);
297 src_addr += src_stride_z;
298 weights_addr += weights_stride_z;
304 float4 bias = (float4) * ((__global
float *)(
vector_offset(&biases, kernel_index)));
310 vstore4(pixels0, 0, (__global
float *)(dst.
ptr + 0 * dst_stride_y));
311 vstore4(pixels1, 0, (__global
float *)(dst.
ptr + 1 * dst_stride_y));
313 #endif // defined(WEIGHTS_DEPTH) Structure to hold Vector information.
#define CONVERT_TO_TENSOR3D_STRUCT(name)
Structure to hold 3D tensor information.
__global uchar * offset(const Image *img, int x, int y)
Get the pointer position of a Image.
#define CONVERT_TO_IMAGE_STRUCT(name)
__global const uchar * tensor3D_offset(const Tensor3D *tensor, int x, int y, int z)
Get the pointer position of a Tensor3D.
#define VECTOR_DECLARATION(name)
Structure to hold Image information.
#define TENSOR3D_DECLARATION(name)
#define VEC_DATA_TYPE(type, size)
#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name)
#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name)
__global const uchar * vector_offset(const Vector *vec, int x)
Get the pointer position of a Vector.
__global uchar * ptr
Pointer to the starting postion of the buffer.
convolution configure & src