diff options
Diffstat (limited to 'intern/cycles/kernel/filter/filter_features_sse.h')
| -rw-r--r-- | intern/cycles/kernel/filter/filter_features_sse.h | 80 |
1 files changed, 34 insertions, 46 deletions
diff --git a/intern/cycles/kernel/filter/filter_features_sse.h b/intern/cycles/kernel/filter/filter_features_sse.h index 3185330994c..3ddd8712266 100644 --- a/intern/cycles/kernel/filter/filter_features_sse.h +++ b/intern/cycles/kernel/filter/filter_features_sse.h @@ -16,7 +16,7 @@ CCL_NAMESPACE_BEGIN -#define ccl_get_feature_sse(pass) _mm_loadu_ps(buffer + (pass)*pass_stride) +#define ccl_get_feature_sse(pass) load_float4(buffer + (pass)*pass_stride) /* Loop over the pixels in the range [low.x, high.x) x [low.y, high.y), 4 at a time. * pixel_buffer always points to the first of the 4 current pixel in the first pass. @@ -24,25 +24,25 @@ CCL_NAMESPACE_BEGIN #define FOR_PIXEL_WINDOW_SSE pixel_buffer = buffer + (low.y - rect.y)*buffer_w + (low.x - rect.x); \ for(pixel.y = low.y; pixel.y < high.y; pixel.y++) { \ - __m128 y4 = _mm_set1_ps(pixel.y); \ + float4 y4 = make_float4(pixel.y); \ for(pixel.x = low.x; pixel.x < high.x; pixel.x += 4, pixel_buffer += 4) { \ - __m128 x4 = _mm_add_ps(_mm_set1_ps(pixel.x), _mm_set_ps(3.0f, 2.0f, 1.0f, 0.0f)); \ - __m128 active_pixels = _mm_cmplt_ps(x4, _mm_set1_ps(high.x)); + float4 x4 = make_float4(pixel.x) + make_float4(0.0f, 1.0f, 2.0f, 3.0f); \ + int4 active_pixels = x4 < make_float4(high.x); #define END_FOR_PIXEL_WINDOW_SSE } \ pixel_buffer += buffer_w - (pixel.x - low.x); \ } -ccl_device_inline void filter_get_features_sse(__m128 x, __m128 y, - __m128 active_pixels, +ccl_device_inline void filter_get_features_sse(float4 x, float4 y, + int4 active_pixels, const float *ccl_restrict buffer, - __m128 *features, - const __m128 *ccl_restrict mean, + float4 *features, + const float4 *ccl_restrict mean, int pass_stride) { features[0] = x; features[1] = y; - features[2] = _mm_fabs_ps(ccl_get_feature_sse(0)); + features[2] = fabs(ccl_get_feature_sse(0)); features[3] = ccl_get_feature_sse(1); features[4] = ccl_get_feature_sse(2); features[5] = ccl_get_feature_sse(3); @@ -52,53 +52,41 @@ ccl_device_inline void filter_get_features_sse(__m128 x, __m128 y, features[9] = ccl_get_feature_sse(7); if(mean) { for(int i = 0; i < DENOISE_FEATURES; i++) - features[i] = _mm_sub_ps(features[i], mean[i]); + features[i] = features[i] - mean[i]; } for(int i = 0; i < DENOISE_FEATURES; i++) - features[i] = _mm_mask_ps(features[i], active_pixels); + features[i] = mask(active_pixels, features[i]); } -ccl_device_inline void filter_get_feature_scales_sse(__m128 x, __m128 y, - __m128 active_pixels, +ccl_device_inline void filter_get_feature_scales_sse(float4 x, float4 y, + int4 active_pixels, const float *ccl_restrict buffer, - __m128 *scales, - const __m128 *ccl_restrict mean, + float4 *scales, + const float4 *ccl_restrict mean, int pass_stride) { - scales[0] = _mm_mask_ps(_mm_fabs_ps(_mm_sub_ps(x, mean[0])), active_pixels); - scales[1] = _mm_mask_ps(_mm_fabs_ps(_mm_sub_ps(y, mean[1])), active_pixels); - - scales[2] = _mm_mask_ps(_mm_fabs_ps(_mm_sub_ps(_mm_fabs_ps(ccl_get_feature_sse(0)), mean[2])), active_pixels); - - __m128 diff, scale; - diff = _mm_sub_ps(ccl_get_feature_sse(1), mean[3]); - scale = _mm_mul_ps(diff, diff); - diff = _mm_sub_ps(ccl_get_feature_sse(2), mean[4]); - scale = _mm_add_ps(scale, _mm_mul_ps(diff, diff)); - diff = _mm_sub_ps(ccl_get_feature_sse(3), mean[5]); - scale = _mm_add_ps(scale, _mm_mul_ps(diff, diff)); - scales[3] = _mm_mask_ps(scale, active_pixels); - - scales[4] = _mm_mask_ps(_mm_fabs_ps(_mm_sub_ps(ccl_get_feature_sse(4), mean[6])), active_pixels); - - diff = _mm_sub_ps(ccl_get_feature_sse(5), mean[7]); - scale = _mm_mul_ps(diff, diff); - diff = _mm_sub_ps(ccl_get_feature_sse(6), mean[8]); - scale = _mm_add_ps(scale, _mm_mul_ps(diff, diff)); - diff = _mm_sub_ps(ccl_get_feature_sse(7), mean[9]); - scale = _mm_add_ps(scale, _mm_mul_ps(diff, diff)); - scales[5] = _mm_mask_ps(scale, active_pixels); + scales[0] = fabs(x - mean[0]); + scales[1] = fabs(y - mean[1]); + scales[2] = fabs(fabs(ccl_get_feature_sse(0)) - mean[2]); + scales[3] = sqr(ccl_get_feature_sse(1) - mean[3]) + + sqr(ccl_get_feature_sse(2) - mean[4]) + + sqr(ccl_get_feature_sse(3) - mean[5]); + scales[4] = fabs(ccl_get_feature_sse(4) - mean[6]); + scales[5] = sqr(ccl_get_feature_sse(5) - mean[7]) + + sqr(ccl_get_feature_sse(6) - mean[8]) + + sqr(ccl_get_feature_sse(7) - mean[9]); + for(int i = 0; i < 6; i++) + scales[i] = mask(active_pixels, scales[i]); } -ccl_device_inline void filter_calculate_scale_sse(__m128 *scale) +ccl_device_inline void filter_calculate_scale_sse(float4 *scale) { - scale[0] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(scale[0]), _mm_set1_ps(0.01f))); - scale[1] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(scale[1]), _mm_set1_ps(0.01f))); - scale[2] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(scale[2]), _mm_set1_ps(0.01f))); - scale[6] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(scale[4]), _mm_set1_ps(0.01f))); - - scale[7] = scale[8] = scale[9] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(_mm_sqrt_ps(scale[5])), _mm_set1_ps(0.01f))); - scale[3] = scale[4] = scale[5] = _mm_rcp_ps(_mm_max_ps(_mm_hmax_ps(_mm_sqrt_ps(scale[3])), _mm_set1_ps(0.01f))); + scale[0] = rcp(max(reduce_max(scale[0]), make_float4(0.01f))); + scale[1] = rcp(max(reduce_max(scale[1]), make_float4(0.01f))); + scale[2] = rcp(max(reduce_max(scale[2]), make_float4(0.01f))); + scale[6] = rcp(max(reduce_max(scale[4]), make_float4(0.01f))); + scale[7] = scale[8] = scale[9] = rcp(max(reduce_max(sqrt(scale[5])), make_float4(0.01f))); + scale[3] = scale[4] = scale[5] = rcp(max(reduce_max(sqrt(scale[3])), make_float4(0.01f))); } |