Tfloat patch 4: bugfixes for AVX2 FAST_FLOAT Extract8+16 implementations #3494
Conversation
Up to now Tesseract used double for training and recognition with "best" models. This commit replaces double by a new data type TFloat which is double by default, but float if FAST_FLOAT is defined. Ideally this should allow faster training. Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Weil <[email protected]>
…6d scale01234567 = _mm256_loadu_ps(scales)`, i.e. loading float vectors into double vector types. Extract from tesseract-ocr#3490.
|
Hi, You are doing patches wrong. |
…ation: for TFloat to work, we don't need to duplicate the integer work functions as it's only the ExtractResults16[8,16] functions that need different implementations for float vs. double. These are therefor common to both implementations:
```
static void PartialMatrixDotVector64(const int8_t *wi, const TFloat *scales, const int8_t *u,
int num_in, TFloat *v) {
static void PartialMatrixDotVector32(const int8_t *wi, const TFloat *scales, const int8_t *u,
int num_in, TFloat *v) {
static void PartialMatrixDotVector16(const int8_t *wi, const TFloat *scales, const int8_t *u,
int num_in, TFloat *v) {
static inline void PartialMatrixDotVector8(const int8_t *wi, const TFloat *scales, const int8_t *u,
int num_in, TFloat *v) {
static void matrixDotVector(int dim1, int dim2, const int8_t *wi, const TFloat *scales,
const int8_t *u, TFloat *v) {
```
|
Note: #3495 is this one (#3494) PLUS FAST_FLOAT condition only applied to the ExtractXYZ calls, as the others are good to go with only their prototype adjusted from double --> TFloat. Hence #3495 is only moving code compared to this one, no code change. (I don't know what diff tools you use, but thus this one (#3494) would be easier to diff/review, and then verify that #3495 is only copy/cut/paste work, resulting in a much larger diff) |
Crap. Yep, seen it. 😊 Discard. Will re-issue. |
|
🤔 I used the github link and didn't watch carefully that the bugger ref'd against mainline master instead of stweil/tfloat. Checked against my own visual commit graph and these were correct, but definitely wholly wrong to submit against master. ugh. |
|
Re-issued as stweil#4. |
Extract from #3490 - bugfixing the AVX2 Extract8+16 codes, where there's lines like
__m256d scale01234567 = _mm256_loadu_ps(scales), i.e. loading float vectors into double vector types. Extract from #3490.Note: next pullreq is a reduced version of this: less code duplication for bleeding edge tfloat branch.