Enable {J400/I400}ToARGBRow_RVV

Run on SiFive internal FPGA*:

I400ToARGB_Opt (~8x vs scalar)
J400ToARGB_Opt (~10x vs scalar)

LIBYUV_WIDTH=1280 LIBYUV_HEIGHT=720 LIBYUV_REPEAT=10

Bug: libyuv:956, libyuv:961
Change-Id: If4e21ec85c4ff79083ec16a6faae0e457129a8de
Signed-off-by: Bruce Lai <bruce.lai@sifive.com>
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4544972
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Wan-Teh Chang <wtc@google.com>

3 files changed, 74 insertions, 0 deletions

diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index a390a68d..3488d256 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -807,6 +807,7 @@ extern "C" {
 #define HAS_ABGRTOYROW_RVV
 #define HAS_ABGRTOYJROW_RVV
 #define HAS_BGRATOYROW_RVV
+#define HAS_I400TOARGBROW_RVV
 #define HAS_I422ALPHATOARGBROW_RVV
 #define HAS_I422TOARGBROW_RVV
 #define HAS_I422TORGB24ROW_RVV
@@ -814,6 +815,7 @@ extern "C" {
 #define HAS_I444ALPHATOARGBROW_RVV
 #define HAS_I444TOARGBROW_RVV
 #define HAS_I444TORGB24ROW_RVV
+#define HAS_J400TOARGBROW_RVV
 #define HAS_MERGEARGBROW_RVV
 #define HAS_MERGERGBROW_RVV
 #define HAS_MERGEXRGBROW_RVV
@@ -3531,6 +3533,7 @@ void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_RVV(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
@@ -4460,6 +4463,10 @@ void I400ToARGBRow_LSX(const uint8_t* src_y,
                        uint8_t* dst_argb,
                        const struct YuvConstants* yuvconstants,
                        int width);
+void I400ToARGBRow_RVV(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
 void I400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* param,
diff --git a/source/convert_argb.cc b/source/convert_argb.cc
index 99bbb96d..f16c368d 100644
--- a/source/convert_argb.cc
+++ b/source/convert_argb.cc
@@ -2897,6 +2897,11 @@ int I400ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I400TOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I400ToARGBRow = I400ToARGBRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I400ToARGBRow(src_y, dst_argb, yuvconstants, width);
@@ -2984,6 +2989,12 @@ int J400ToARGB(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_J400TOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    J400ToARGBRow = J400ToARGBRow_RVV;
+  }
+#endif
+
   for (y = 0; y < height; ++y) {
     J400ToARGBRow(src_y, dst_argb, width);
     src_y += src_stride_y;
diff --git a/source/row_rvv.cc b/source/row_rvv.cc
index 904cbf7d..89333321 100644
--- a/source/row_rvv.cc
+++ b/source/row_rvv.cc
@@ -474,6 +474,62 @@ void I422ToRGB24Row_RVV(const uint8_t* src_y,
   } while (w > 0);
 }
 
+void I400ToARGBRow_RVV(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  size_t w = (size_t)width;
+  size_t vl = __riscv_vsetvl_e8m2(w);
+  const bool is_yb_positive = (yuvconstants->kRGBCoeffBias[4] >= 0);
+  vuint8m2_t v_a = __riscv_vmv_v_x_u8m2(255u, vl);
+  vuint16m4_t v_yb;
+  vuint16m4_t v_yg = __riscv_vmv_v_x_u16m4(yuvconstants->kRGBCoeffBias[0], vl);
+  // To match behavior on other platforms, vxrm (fixed-point rounding mode
+  // register) sets to round-down mode(2).
+  asm volatile("csrwi vxrm, 2");
+  if (is_yb_positive) {
+    v_yb = __riscv_vmv_v_x_u16m4(yuvconstants->kRGBCoeffBias[4], vl);
+  } else {
+    v_yb = __riscv_vmv_v_x_u16m4(-yuvconstants->kRGBCoeffBias[4], vl);
+  }
+  do {
+    vuint8m2_t v_y, v_out;
+    vuint16m4_t v_y_16, v_tmp0, v_tmp1, v_tmp2;
+    vuint32m8_t v_y1;
+    vl = __riscv_vsetvl_e8m2(w);
+    v_y = __riscv_vle8_v_u8m2(src_y, vl);
+    v_y_16 = __riscv_vwaddu_vx_u16m4(v_y, 0, vl);
+    v_tmp0 = __riscv_vmul_vx_u16m4(v_y_16, 0x0101, vl);  // 257 * v_y
+    v_y1 = __riscv_vwmulu_vv_u32m8(v_tmp0, v_yg, vl);
+    v_tmp1 = __riscv_vnsrl_wx_u16m4(v_y1, 16, vl);
+    if (is_yb_positive) {
+      v_tmp2 = __riscv_vsaddu_vv_u16m4(v_tmp1, v_yb, vl);
+    } else {
+      v_tmp2 = __riscv_vssubu_vv_u16m4(v_tmp1, v_yb, vl);
+    }
+    v_out = __riscv_vnclipu_wx_u8m2(v_tmp2, 6, vl);
+    __riscv_vsseg4e8_v_u8m2(dst_argb, v_out, v_out, v_out, v_a, vl);
+    w -= vl;
+    src_y += vl;
+    dst_argb += vl * 4;
+  } while (w > 0);
+}
+
+void J400ToARGBRow_RVV(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  size_t w = (size_t)width;
+  size_t vl = __riscv_vsetvl_e8m2(w);
+  vuint8m2_t v_a = __riscv_vmv_v_x_u8m2(255u, vl);
+  do {
+    vuint8m2_t v_y;
+    v_y = __riscv_vle8_v_u8m2(src_y, vl);
+    __riscv_vsseg4e8_v_u8m2(dst_argb, v_y, v_y, v_y, v_a, vl);
+    w -= vl;
+    src_y += vl;
+    dst_argb += vl * 4;
+    vl = __riscv_vsetvl_e8m2(w);
+  } while (w > 0);
+}
+
 void SplitRGBRow_RVV(const uint8_t* src_rgb,
                      uint8_t* dst_r,
                      uint8_t* dst_g,