Add ARGBToRAWRow_RVV, ARGBToRGB24Row_RVV, RGB24ToARGBRow_RVV

* Run on SiFive internal FPGA:

ARGBToRAW_Opt (~1.55x vs scalar)

ARGBToRGB24_Opt (~1.44x vs scalar)

RGB24ToARGB_Opt (~1.77x vs scalar)

LIBYUV_WIDTH=1280 LIBYUV_HEIGHT=720 LIBYUV_REPEAT=10

Bug: libyuv:956

Change-Id: I26722f6848cd68684d95d9a7ee06ce0416e7985d
Signed-off-by: Darren Hsieh <darren.hsieh@sifive.com>
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4413083
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>

1 files changed, 50 insertions, 8 deletions

diff --git a/source/row_rvv.cc b/source/row_rvv.cc
index 0f264d34..629eca46 100644
--- a/source/row_rvv.cc
+++ b/source/row_rvv.cc
@@ -30,33 +30,33 @@ extern "C" {
 void RAWToARGBRow_RVV(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   size_t vl = __riscv_vsetvl_e8m2(width);
   vuint8m2_t v_a = __riscv_vmv_v_x_u8m2(255u, vl);
-  while (width > 0) {
+  do {
     vuint8m2_t v_b, v_g, v_r;
-    vl = __riscv_vsetvl_e8m2(width);
     __riscv_vlseg3e8_v_u8m2(&v_r, &v_g, &v_b, src_raw, vl);
     __riscv_vsseg4e8_v_u8m2(dst_argb, v_b, v_g, v_r, v_a, vl);
     width -= vl;
     src_raw += (3 * vl);
     dst_argb += (4 * vl);
-  }
+    vl = __riscv_vsetvl_e8m2(width);
+  } while (width > 0);
 }
 
 void RAWToRGBARow_RVV(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
   size_t vl = __riscv_vsetvl_e8m2(width);
   vuint8m2_t v_a = __riscv_vmv_v_x_u8m2(255u, vl);
-  while (width > 0) {
+  do {
     vuint8m2_t v_b, v_g, v_r;
-    vl = __riscv_vsetvl_e8m2(width);
     __riscv_vlseg3e8_v_u8m2(&v_r, &v_g, &v_b, src_raw, vl);
     __riscv_vsseg4e8_v_u8m2(dst_rgba, v_a, v_b, v_g, v_r, vl);
     width -= vl;
     src_raw += (3 * vl);
     dst_rgba += (4 * vl);
-  }
+    vl = __riscv_vsetvl_e8m2(width);
+  } while (width > 0);
 }
 
 void RAWToRGB24Row_RVV(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
-  while (width > 0) {
+  do {
     vuint8m2_t v_b, v_g, v_r;
     size_t vl = __riscv_vsetvl_e8m2(width);
     __riscv_vlseg3e8_v_u8m2(&v_b, &v_g, &v_r, src_raw, vl);
@@ -64,7 +64,49 @@ void RAWToRGB24Row_RVV(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
     width -= vl;
     src_raw += (3 * vl);
     dst_rgb24 += (3 * vl);
-  }
+  } while (width > 0);
+}
+
+void ARGBToRAWRow_RVV(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
+  do {
+    vuint8m2_t v_b, v_g, v_r, v_a;
+    size_t vl = __riscv_vsetvl_e8m2(width);
+    __riscv_vlseg4e8_v_u8m2(&v_b, &v_g, &v_r, &v_a, src_argb, vl);
+    __riscv_vsseg3e8_v_u8m2(dst_raw, v_r, v_g, v_b, vl);
+    width -= vl;
+    src_argb += (4 * vl);
+    dst_raw += (3 * vl);
+  } while (width > 0);
+}
+
+void ARGBToRGB24Row_RVV(const uint8_t* src_argb,
+                        uint8_t* dst_rgb24,
+                        int width) {
+  do {
+    vuint8m2_t v_b, v_g, v_r, v_a;
+    size_t vl = __riscv_vsetvl_e8m2(width);
+    __riscv_vlseg4e8_v_u8m2(&v_b, &v_g, &v_r, &v_a, src_argb, vl);
+    __riscv_vsseg3e8_v_u8m2(dst_rgb24, v_b, v_g, v_r, vl);
+    width -= vl;
+    src_argb += (4 * vl);
+    dst_rgb24 += (3 * vl);
+  } while (width > 0);
+}
+
+void RGB24ToARGBRow_RVV(const uint8_t* src_rgb24,
+                        uint8_t* dst_argb,
+                        int width) {
+  size_t vl = __riscv_vsetvl_e8m2(width);
+  vuint8m2_t v_a = __riscv_vmv_v_x_u8m2(255u, vl);
+  do {
+    vuint8m2_t v_b, v_g, v_r;
+    __riscv_vlseg3e8_v_u8m2(&v_b, &v_g, &v_r, src_rgb24, vl);
+    __riscv_vsseg4e8_v_u8m2(dst_argb, v_b, v_g, v_r, v_a, vl);
+    width -= vl;
+    src_rgb24 += (3 * vl);
+    dst_argb += (4 * vl);
+    vl = __riscv_vsetvl_e8m2(width);
+  } while (width > 0);
 }
 
 #ifdef __cplusplus