RPP Fog - Enhancement of Fog Effect with Gray Tone Integration by sampath1117 · Pull Request #340 · r-abishek/rpp

sampath1117 · 2024-09-19T12:00:16Z

A gray tone has been incorporated into the fog effect to achieve a more realistic and balanced visual output

sampath1117 · 2024-09-19T12:01:32Z

@r-abishek
Performance needs to be checked on this PR and needs review

Please dont merge this PR to opensource PR till both of the above is checked

sampath1117 · 2024-09-20T10:38:35Z

Checked the performance without and without this additional changes and it looks similar

sampath1117 · 2024-09-20T10:45:48Z

src/include/cpu/rpp_cpu_common.hpp

+{
+    __m256 pAlphaMaskFactor[2], pIntensityMaskFactor[2], pGray[2], pOneMinusGrayFactor;
+    pGray[0] = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(0.299f), p[0]), _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(0.587f), p[2]), _mm256_mul_ps(_mm256_set1_ps(0.114f), p[4]))); 
+    pGray[1] = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(0.299f), p[1]), _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(0.587f), p[3]), _mm256_mul_ps(_mm256_set1_ps(0.114f), p[5])));


Couple of changes

Please store this 3 __m256 const values in an array in fog kernel here, instead of doing set1_ps operation each time
https://github.com/r-abishek/rpp/blob/sr/fog_pr_changes/src/modules/cpu/kernel/fog.hpp#L84

__m256 pConversionFactor[3]; pConversionFactor[0] = _mm256_set1_ps(0.299f); pConversionFactor[1] = _mm256_set1_ps(0.587f); pConversionFactor[2] = _mm256_set1_ps(0.114f);

Pass this array to compute_fog_48_host and use it

Use fmadd operation for L3141 and L3142 since we are doing multiplication followed by addition

is this done?

r-abishek · 2024-09-23T22:02:22Z

@sampath1117 The diff should also show us removing the fog image from ROCm docs and adding the new improved image?

r-abishek

@sampath1117 I have added some more comments on this PR to address before merge. Some comments have multiple instances to be changed, pls check

r-abishek · 2024-09-24T22:53:16Z

utilities/test_suite/HIP/Tensor_hip.cpp

                    for (i = 0; i < batchSize; i++)
+                    {
                        intensityFactor[i] = 0;
+                        greyFactor[i] = 0;


Why is greyFactor 0 for hip and 0.3 for host? Please match this

r-abishek · 2024-09-24T22:57:22Z

docs/data/doxygenOutputs/effects_augmentations_fog_img640x480.png

Does this have a grayFactor of 0? Please consistently add the 0.3 or perhaps 0.35 everywhere. On the HOST/HIP test suites, as well as this ROCm docs image

r-abishek · 2024-09-24T22:58:54Z

include/rppt_tensor_effects_augmentations.h

 * \param [out] dstPtr destination tensor in HOST memory
 * \param [in] dstDescPtr destination tensor descriptor (Restrictions - numDims = 4, offsetInBytes >= 0, dataType = U8/F16/F32/I8, layout = NCHW/NHWC, c = same as that of srcDescPtr)
 * \param [in] intensityFactor intensity factor values for fog calculation (1D tensor in HOST memory, of size batchSize, with 0 <= intensityFactor <= 0.5 for each image in batch)
+ * \param [in] grayFactor gray factor values for fog calculation (1D tensor in HOST memory, of size batchSize, with 0 <= grayFactor <= 1 for each image in batch)


Change comment for both HOST and HIP to something like:

\param [in] grayFactor gray factor values to introduce grayness in the image for fog calculation

r-abishek · 2024-09-24T23:00:37Z

src/modules/hip/kernel/fog.hpp

+    float4 bMultiplier_f4 = static_cast<float4>(0.114f);
+    grey_f4[0] = r_f8->f4[0] * rMultiplier_f4 + g_f8->f4[0] * gMultiplier_f4 + b_f8->f4[0] * bMultiplier_f4;
+    grey_f4[1] = r_f8->f4[1] * rMultiplier_f4 + g_f8->f4[1] * gMultiplier_f4 + b_f8->f4[1] * bMultiplier_f4;
+    float4 oneMinusGreyFactor = static_cast<float4>(1.0f) - *greyFactor_f4;


oneMinusGreyFactor_f4

r-abishek · 2024-09-24T23:14:00Z