diff options
Diffstat (limited to 'chapters')
-rw-r--r-- | chapters/image.adoc | 28 |
1 files changed, 18 insertions, 10 deletions
diff --git a/chapters/image.adoc b/chapters/image.adoc index b560b5c..3bf4109 100644 --- a/chapters/image.adoc +++ b/chapters/image.adoc @@ -53,16 +53,24 @@ None [source,c++] ---- -// shift assert prevents int32_t accumulator overflow for in_t==int8_t -assert((resize_t == float_t && shift == 0)||(0 < shift && shift <= 11)); -// stride asserts set consistent lower limit of 1/16 downscale -// independent of the shift value to simplify implementations -assert(0 < stride_x && stride_x < (16 << shift)); -assert(0 < stride_y && stride_y < (16 << shift)); -// offset range is similarly limited to maximum 16 pixels irrespective -// of shift. Both stride and offset fit in int16_t when shift=11. -assert((-16 << shift) < offset_x && offset_x < (16 << shift)); -assert((-16 << shift) < offset_y && offset_y < (16 << shift)); +// Ensure image size is supported by GPU APIs and that for integer +// implementations, position * stride does not overflow int32_t. +assert(max(OH,OW,IH,IW) < 16384); +if (resize_t == float_t) { + // The shift attribute is not used for floating point + assert(shift == 0); +} else { + // if in_t=int8_t ensure that an int32_t accumulator can be used + assert(0 < shift && shift <= 11); + // set a consistent lower limit of 1/16 downscale + // independent of the shift value to simplify implementations + assert(0 < stride_x && stride_x < (16 << shift)); + assert(0 < stride_y && stride_y < (16 << shift)); + // offset range is similarly limited to maximum 16 pixels irrespective + // of shift. Both stride and offset fit in int16_t when shift=11. + assert((-16 << shift) < offset_x && offset_x < (16 << shift)); + assert((-16 << shift) < offset_y && offset_y < (16 << shift)); +} for_each(0 <= n < N, 0 <= oy < OH, 0 <= ox < OW; 0 <= c < C) { unit = (resize_t == float_t) ? 1.0 : (1 << shift); y = oy * stride_y + offset_y; |