Enable I{422,444}AlphaToARGBRow_RVV & ARGBAttentuateRow_RVV

Run on SiFive internal FPGA:

I444AlphaToARGB_Opt (~16x vs scalar)
I422AlphaToARGB_Opt (~10x vs scalar)
ARGBAttenuate_Opt (~3x vs scalar)

LIBYUV_WIDTH=1280 LIBYUV_HEIGHT=720 LIBYUV_REPEAT=10

Change-Id: I0046eb7af8104bc8e13cee1cb91a19f90940d5b0
Signed-off-by: Bruce Lai <bruce.lai@sifive.com>
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4535657
Commit-Queue: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>

3 files changed, 170 insertions, 2 deletions

diff --git a/source/convert_argb.cc b/source/convert_argb.cc
index 62884e5d..4d953fae 100644
--- a/source/convert_argb.cc
+++ b/source/convert_argb.cc
@@ -11,7 +11,6 @@
 #include "libyuv/convert_argb.h"
 
 #include <assert.h>
-
 #include "libyuv/convert_from_argb.h"
 #include "libyuv/cpu_id.h"
 #ifdef HAVE_JPEG
@@ -2059,6 +2058,11 @@ int I420AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422ALPHATOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_RVV;
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -2091,6 +2095,11 @@ int I420AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -2192,6 +2201,11 @@ int I422AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422ALPHATOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_RVV;
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -2224,6 +2238,11 @@ int I422AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -2307,6 +2326,11 @@ int I444AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I444ALPHATOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_RVV;
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -2339,6 +2363,11 @@ int I444AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I444AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -2567,6 +2596,11 @@ int I010AlphaToARGBMatrix(const uint16_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -2668,6 +2702,11 @@ int I210AlphaToARGBMatrix(const uint16_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -2767,6 +2806,11 @@ int I410AlphaToARGBMatrix(const uint16_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I410AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -6703,6 +6747,11 @@ static int I420AlphaToARGBMatrixBilinear(
     }
   }
 #endif
+#if defined(HAS_I444ALPHATOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_RVV;
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -6735,6 +6784,11 @@ static int I420AlphaToARGBMatrixBilinear(
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
 #if defined(HAS_SCALEROWUP2_BILINEAR_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
@@ -6895,6 +6949,11 @@ static int I422AlphaToARGBMatrixLinear(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I444ALPHATOARGBROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_RVV;
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -6927,6 +6986,12 @@ static int I422AlphaToARGBMatrixLinear(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
+
 #if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ScaleRowUp2_Linear = ScaleRowUp2_Linear_Any_SSE2;
@@ -7060,6 +7125,11 @@ static int I010AlphaToARGBMatrixBilinear(
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
 #if defined(HAS_SCALEROWUP2_BILINEAR_12_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
@@ -7221,6 +7291,11 @@ static int I210AlphaToARGBMatrixLinear(const uint16_t* src_y,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
 #if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
diff --git a/source/planar_functions.cc b/source/planar_functions.cc
index 32424fb9..fd12718d 100644
--- a/source/planar_functions.cc
+++ b/source/planar_functions.cc
@@ -3566,6 +3566,11 @@ int ARGBAttenuate(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBATTENUATEROW_RVV)
+  if (TestCpuFlag(kCpuHasRVV)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_RVV;
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBAttenuateRow(src_argb, dst_argb, width);
diff --git a/source/row_rvv.cc b/source/row_rvv.cc
index 99f23165..ad131924 100644
--- a/source/row_rvv.cc
+++ b/source/row_rvv.cc
@@ -16,7 +16,6 @@
  */
 
 #include <assert.h>
-
 #include "libyuv/row.h"
 
 #if !defined(LIBYUV_DISABLE_RVV) && defined(__riscv_vector)
@@ -293,6 +292,38 @@ void I444ToARGBRow_RVV(const uint8_t* src_y,
   } while (w > 0);
 }
 
+void I444AlphaToARGBRow_RVV(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  size_t vl;
+  size_t w = (size_t)width;
+  vuint8m1_t v_u, v_v;
+  vuint8m1_t v_ub, v_vr, v_ug, v_vg;
+  vuint8m1_t v_b, v_g, v_r, v_a;
+  vuint16m2_t v_yg, v_bb, v_bg, v_br;
+  vuint16m2_t v_y_16, v_g_16, v_b_16, v_r_16;
+  YUVTORGB_SETUP(yuvconstants, vl, v_ub, v_vr, v_ug, v_vg, v_yg, v_bb, v_bg,
+                 v_br);
+  do {
+    READYUV444(vl, v_u, v_v, v_y_16);
+    v_a = __riscv_vle8_v_u8m1(src_a, vl);
+    YUVTORGB(vl, v_u, v_v, v_ub, v_vr, v_ug, v_vg, v_yg, v_bb, v_bg, v_br,
+             v_y_16, v_g_16, v_b_16, v_r_16);
+    RGBTORGB8(vl, v_g_16, v_b_16, v_r_16, v_g, v_b, v_r);
+    __riscv_vsseg4e8_v_u8m1(dst_argb, v_b, v_g, v_r, v_a, vl);
+    w -= vl;
+    src_y += vl;
+    src_a += vl;
+    src_u += vl;
+    src_v += vl;
+    dst_argb += vl * 4;
+  } while (w > 0);
+}
+
 void I444ToRGB24Row_RVV(const uint8_t* src_y,
                         const uint8_t* src_u,
                         const uint8_t* src_v,
@@ -352,6 +383,38 @@ void I422ToARGBRow_RVV(const uint8_t* src_y,
   } while (w > 0);
 }
 
+void I422AlphaToARGBRow_RVV(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  size_t vl;
+  size_t w = (size_t)width;
+  vuint8m1_t v_u, v_v;
+  vuint8m1_t v_ub, v_vr, v_ug, v_vg;
+  vuint8m1_t v_b, v_g, v_r, v_a;
+  vuint16m2_t v_yg, v_bb, v_bg, v_br;
+  vuint16m2_t v_y_16, v_g_16, v_b_16, v_r_16;
+  YUVTORGB_SETUP(yuvconstants, vl, v_ub, v_vr, v_ug, v_vg, v_yg, v_bb, v_bg,
+                 v_br);
+  do {
+    READYUV422(vl, v_u, v_v, v_y_16);
+    v_a = __riscv_vle8_v_u8m1(src_a, vl);
+    YUVTORGB(vl, v_u, v_v, v_ub, v_vr, v_ug, v_vg, v_yg, v_bb, v_bg, v_br,
+             v_y_16, v_g_16, v_b_16, v_r_16);
+    RGBTORGB8(vl, v_g_16, v_b_16, v_r_16, v_g, v_b, v_r);
+    __riscv_vsseg4e8_v_u8m1(dst_argb, v_b, v_g, v_r, v_a, vl);
+    w -= vl;
+    src_y += vl;
+    src_a += vl;
+    src_u += vl / 2;
+    src_v += vl / 2;
+    dst_argb += vl * 4;
+  } while (w > 0);
+}
+
 void I422ToRGBARow_RVV(const uint8_t* src_y,
                        const uint8_t* src_u,
                        const uint8_t* src_v,
@@ -710,6 +773,31 @@ void RAWToYRow_RVV(const uint8_t* src_raw, uint8_t* dst_y, int width) {
   RGBToYMatrixRow_RVV(src_raw, dst_y, width, &kRawI601Constants);
 }
 
+void ARGBAttenuateRow_RVV(const uint8_t* src_argb,
+                          uint8_t* dst_argb,
+                          int width) {
+  size_t w = (size_t)width;
+  // To match behavior on other platforms, vxrm (fixed-point rounding mode
+  // register) is set to round-to-nearest-up(0).
+  asm volatile("csrwi vxrm, 0");
+  do {
+    vuint8m2_t v_b, v_g, v_r, v_a;
+    vuint16m4_t v_ba_16, v_ga_16, v_ra_16;
+    size_t vl = __riscv_vsetvl_e8m2(w);
+    __riscv_vlseg4e8_v_u8m2(&v_b, &v_g, &v_r, &v_a, src_argb, vl);
+    v_ba_16 = __riscv_vwmulu_vv_u16m4(v_b, v_a, vl);
+    v_ga_16 = __riscv_vwmulu_vv_u16m4(v_g, v_a, vl);
+    v_ra_16 = __riscv_vwmulu_vv_u16m4(v_r, v_a, vl);
+    v_b = __riscv_vnclipu_wx_u8m2(v_ba_16, 8, vl);
+    v_g = __riscv_vnclipu_wx_u8m2(v_ga_16, 8, vl);
+    v_r = __riscv_vnclipu_wx_u8m2(v_ra_16, 8, vl);
+    __riscv_vsseg4e8_v_u8m2(dst_argb, v_b, v_g, v_r, v_a, vl);
+    w -= vl;
+    src_argb += vl * 4;
+    dst_argb += vl * 4;
+  } while (w > 0);
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv