From f37352abfccb76e649b8b04f14dc2cfe903fdf31 Mon Sep 17 00:00:00 2001
From: Martin Pulec <martin.pulec@cesnet.cz>
Date: Tue, 24 May 2022 14:24:47 +0200
Subject: [PATCH] NDI cap, disp.: mark some conversions form optimization

---
 src/video_capture/ndi.cpp | 6 ++++--
 src/video_display/ndi.c   | 6 ++++--
 2 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/src/video_capture/ndi.cpp b/src/video_capture/ndi.cpp
index d62a08872..2493d00f0 100644
--- a/src/video_capture/ndi.cpp
+++ b/src/video_capture/ndi.cpp
@@ -363,7 +363,8 @@ static void convert_BGRA_RGBA(struct video_frame *out, const uint8_t *data, int
         auto *out_p = reinterpret_cast<uint32_t *>(out->tiles[0].data) + field_idx * out->tiles[0].width;
         for (unsigned int i = 0; i < out->tiles[0].height; i += total_fields) {
                 const auto *in_p = reinterpret_cast<const uint32_t *>(data + i * in_stride);
-                for (unsigned int j = 0; j < out->tiles[0].width; j++) {
+                unsigned int width = out->tiles[0].width;
+                OPTIMIZED_FOR (unsigned int j = 0; j < width; j++) {
                         uint32_t argb = *in_p++;
                         *out_p++ = (argb & 0xFF000000U) | ((argb & 0xFFU) << 16U) | (argb & 0xFF00U) | ((argb & 0xFF0000U) >> 16U);
                 }
@@ -377,7 +378,8 @@ static void convert_P216_Y216(struct video_frame *out, const uint8_t *data, [[ma
         const auto *in_cb_cr = reinterpret_cast<const uint16_t *>(data) + out->tiles[0].width * out->tiles[0].height;
         auto *out_p = reinterpret_cast<uint16_t *>(out->tiles[0].data) + 2 * field_idx * out->tiles[0].width;
         for (unsigned int i = 0; i < out->tiles[0].height;  i += total_fields) {
-                for (unsigned int j = 0; j < out->tiles[0].width; j += 2) {
+                unsigned int width = out->tiles[0].width;
+                OPTIMIZED_FOR (unsigned int j = 0; j < (width + 1) / 2; j += 1) {
                         *out_p++ = *in_y++;
                         *out_p++ = *in_cb_cr++;
                         *out_p++ = *in_y++;
diff --git a/src/video_display/ndi.c b/src/video_display/ndi.c
index f5b482956..75c769f89 100644
--- a/src/video_display/ndi.c
+++ b/src/video_display/ndi.c
@@ -258,7 +258,8 @@ static void ndi_disp_convert_Y216_to_P216(const struct video_frame *f, char *out
         uint16_t *out_cb_cr = (uint16_t *)(void *) out + f->tiles[0].width * f->tiles[0].height;
 
         for (unsigned int i = 0; i < f->tiles[0].height; ++i) {
-                for (unsigned int j = 0; j < f->tiles[0].width; j += 2) {
+                unsigned int width = f->tiles[0].width;
+                OPTIMIZED_FOR (unsigned int j = 0; j < (width + 1) / 2; j += 1) {
                         *out_y++ = *in++;
                         *out_cb_cr++ = *in++;
                         *out_y++ = *in++;
@@ -276,7 +277,8 @@ static void ndi_disp_convert_Y416_to_P216(const struct video_frame *f, char *out
         uint16_t *out_cb_cr = (uint16_t *)(void *) out + f->tiles[0].width * f->tiles[0].height;
 
         for (unsigned int i = 0; i < f->tiles[0].height; ++i) {
-                OPTIMIZED_FOR (unsigned int j = 0; j < f->tiles[0].width; j += 2) {
+                unsigned int width = f->tiles[0].width;
+                OPTIMIZED_FOR (unsigned int j = 0; j < (width + 1) / 2; j += 1) {
                         *out_y++ = in[1];
                         *out_cb_cr++ = (in[2] + in[6]) / 2;
                         *out_y++ = in[5];