aboutsummaryrefslogtreecommitdiff
path: root/files/source/row_lasx.cc
diff options
context:
space:
mode:
Diffstat (limited to 'files/source/row_lasx.cc')
-rw-r--r--files/source/row_lasx.cc370
1 files changed, 222 insertions, 148 deletions
diff --git a/files/source/row_lasx.cc b/files/source/row_lasx.cc
index 7dd18f40..1082ad80 100644
--- a/files/source/row_lasx.cc
+++ b/files/source/row_lasx.cc
@@ -775,40 +775,6 @@ void UYVYToUV422Row_LASX(const uint8_t* src_uyvy,
}
}
-void ARGBToYRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
- int x;
- int len = width / 32;
- __m256i src0, src1, src2, src3, vec0, vec1, vec2, vec3;
- __m256i tmp0, tmp1, dst0;
- __m256i const_19 = __lasx_xvldi(0x19);
- __m256i const_42 = __lasx_xvldi(0x42);
- __m256i const_81 = __lasx_xvldi(0x81);
- __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
- 0x1080108010801080, 0x1080108010801080};
- __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
- 0x0000000700000003};
-
- for (x = 0; x < len; x++) {
- DUP4_ARG2(__lasx_xvld, src_argb0, 0, src_argb0, 32, src_argb0, 64,
- src_argb0, 96, src0, src1, src2, src3);
- vec0 = __lasx_xvpickev_b(src1, src0);
- vec1 = __lasx_xvpickev_b(src3, src2);
- vec2 = __lasx_xvpickod_b(src1, src0);
- vec3 = __lasx_xvpickod_b(src3, src2);
- tmp0 = __lasx_xvmaddwev_h_bu(const_1080, vec0, const_19);
- tmp1 = __lasx_xvmaddwev_h_bu(const_1080, vec1, const_19);
- tmp0 = __lasx_xvmaddwev_h_bu(tmp0, vec2, const_81);
- tmp1 = __lasx_xvmaddwev_h_bu(tmp1, vec3, const_81);
- tmp0 = __lasx_xvmaddwod_h_bu(tmp0, vec0, const_42);
- tmp1 = __lasx_xvmaddwod_h_bu(tmp1, vec1, const_42);
- dst0 = __lasx_xvssrani_b_h(tmp1, tmp0, 8);
- dst0 = __lasx_xvperm_w(dst0, control);
- __lasx_xvst(dst0, dst_y, 0);
- src_argb0 += 128;
- dst_y += 32;
- }
-}
-
void ARGBToUVRow_LASX(const uint8_t* src_argb0,
int src_stride_argb,
uint8_t* dst_u,
@@ -1216,7 +1182,7 @@ void ARGBAttenuateRow_LASX(const uint8_t* src_argb,
void ARGBToRGB565DitherRow_LASX(const uint8_t* src_argb,
uint8_t* dst_rgb,
- const uint32_t dither4,
+ uint32_t dither4,
int width) {
int x;
int len = width / 16;
@@ -1811,48 +1777,6 @@ void RGB565ToUVRow_LASX(const uint8_t* src_rgb565,
}
}
-void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
- int x;
- int len = width / 32;
- __m256i src0, src1, src2;
- __m256i tmp0, tmp1, tmp2, tmp3;
- __m256i reg0, reg1, reg2, dst0;
- __m256i const_129 = __lasx_xvldi(129);
- __m256i const_br = {0x4219421942194219, 0x4219421942194219,
- 0x4219421942194219, 0x4219421942194219};
- __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
- 0x1080108010801080, 0x1080108010801080};
- __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
- 0x17151412110F0E0C};
- __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
- 0x0F0D0C0A09070604};
- __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
- 0x001600130010000D};
- __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
- 0x000E000B00080005};
-
- for (x = 0; x < len; x++) {
- reg0 = __lasx_xvld(src_rgb24, 0);
- reg1 = __lasx_xvld(src_rgb24, 32);
- reg2 = __lasx_xvld(src_rgb24, 64);
- src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
- src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
- src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
- tmp0 = __lasx_xvshuf_b(src1, src0, shuff0);
- tmp1 = __lasx_xvshuf_b(src1, src2, shuff1);
- tmp2 = __lasx_xvshuf_b(src1, src0, shuff2);
- tmp3 = __lasx_xvshuf_b(src1, src2, shuff3);
- reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmp2, const_129);
- reg1 = __lasx_xvmaddwev_h_bu(const_1080, tmp3, const_129);
- reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
- reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp1);
- dst0 = __lasx_xvpickod_b(reg1, reg0);
- __lasx_xvst(dst0, dst_y, 0);
- dst_y += 32;
- src_rgb24 += 96;
- }
-}
-
void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
int src_stride_rgb24,
uint8_t* dst_u,
@@ -1916,48 +1840,6 @@ void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
}
}
-void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
- int x;
- int len = width / 32;
- __m256i src0, src1, src2;
- __m256i tmp0, tmp1, tmp2, tmp3;
- __m256i reg0, reg1, reg2, dst0;
- __m256i const_129 = __lasx_xvldi(129);
- __m256i const_br = {0x1942194219421942, 0x1942194219421942,
- 0x1942194219421942, 0x1942194219421942};
- __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
- 0x1080108010801080, 0x1080108010801080};
- __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
- 0x17151412110F0E0C};
- __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
- 0x0F0D0C0A09070604};
- __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
- 0x001600130010000D};
- __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
- 0x000E000B00080005};
-
- for (x = 0; x < len; x++) {
- reg0 = __lasx_xvld(src_raw, 0);
- reg1 = __lasx_xvld(src_raw, 32);
- reg2 = __lasx_xvld(src_raw, 64);
- src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
- src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
- src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
- tmp0 = __lasx_xvshuf_b(src1, src0, shuff0);
- tmp1 = __lasx_xvshuf_b(src1, src2, shuff1);
- tmp2 = __lasx_xvshuf_b(src1, src0, shuff2);
- tmp3 = __lasx_xvshuf_b(src1, src2, shuff3);
- reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmp2, const_129);
- reg1 = __lasx_xvmaddwev_h_bu(const_1080, tmp3, const_129);
- reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
- reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp1);
- dst0 = __lasx_xvpickod_b(reg1, reg0);
- __lasx_xvst(dst0, dst_y, 0);
- dst_y += 32;
- src_raw += 96;
- }
-}
-
void RAWToUVRow_LASX(const uint8_t* src_raw,
int src_stride_raw,
uint8_t* dst_u,
@@ -2118,36 +2000,228 @@ void NV21ToARGBRow_LASX(const uint8_t* src_y,
}
}
-void ARGBToYJRow_LASX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
- int x;
- int len = width / 32;
- __m256i src0, src1, src2, src3, dst0;
- __m256i tmp0, tmp1, tmp2, tmp3;
- __m256i reg0, reg1;
- __m256i const_128 = __lasx_xvldi(0x480);
- __m256i const_150 = __lasx_xvldi(0x96);
- __m256i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D,
- 0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
- __m256i shuff = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
- 0x0000000700000003};
+struct RgbConstants {
+ uint8_t kRGBToY[4];
+ uint16_t kAddY;
+ uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+ 128,
+ 0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+ 0x1080,
+ 0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+ 0x1080,
+ 0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+static void ARGBToYMatrixRow_LASX(const uint8_t* src_argb,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
+ int32_t shuff[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+ asm volatile(
+ "xvldrepl.b $xr0, %3, 0 \n\t" // load rgbconstants
+ "xvldrepl.b $xr1, %3, 1 \n\t" // load rgbconstants
+ "xvldrepl.b $xr2, %3, 2 \n\t" // load rgbconstants
+ "xvldrepl.h $xr3, %3, 4 \n\t" // load rgbconstants
+ "xvld $xr20, %4, 0 \n\t" // load shuff
+ "1: \n\t"
+ "xvld $xr4, %0, 0 \n\t"
+ "xvld $xr5, %0, 32 \n\t"
+ "xvld $xr6, %0, 64 \n\t"
+ "xvld $xr7, %0, 96 \n\t" // load 32 pixels of
+ // ARGB
+ "xvor.v $xr12, $xr3, $xr3 \n\t"
+ "xvor.v $xr13, $xr3, $xr3 \n\t"
+ "addi.d %2, %2, -32 \n\t" // 32 processed per
+ // loop.
+ "xvpickev.b $xr8, $xr5, $xr4 \n\t" // BR
+ "xvpickev.b $xr10, $xr7, $xr6 \n\t"
+ "xvpickod.b $xr9, $xr5, $xr4 \n\t" // GA
+ "xvpickod.b $xr11, $xr7, $xr6 \n\t"
+ "xvmaddwev.h.bu $xr12, $xr8, $xr0 \n\t" // B
+ "xvmaddwev.h.bu $xr13, $xr10, $xr0 \n\t"
+ "xvmaddwev.h.bu $xr12, $xr9, $xr1 \n\t" // G
+ "xvmaddwev.h.bu $xr13, $xr11, $xr1 \n\t"
+ "xvmaddwod.h.bu $xr12, $xr8, $xr2 \n\t" // R
+ "xvmaddwod.h.bu $xr13, $xr10, $xr2 \n\t"
+ "addi.d %0, %0, 128 \n\t"
+ "xvpickod.b $xr10, $xr13, $xr12 \n\t"
+ "xvperm.w $xr11, $xr10, $xr20 \n\t"
+ "xvst $xr11, %1, 0 \n\t"
+ "addi.d %1, %1, 32 \n\t"
+ "bnez %2, 1b \n\t"
+ : "+&r"(src_argb), // %0
+ "+&r"(dst_y), // %1
+ "+&r"(width) // %2
+ : "r"(rgbconstants), "r"(shuff)
+ : "memory");
+}
- for (x = 0; x < len; x++) {
- DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
- 96, src0, src1, src2, src3);
- tmp0 = __lasx_xvpickev_b(src1, src0);
- tmp1 = __lasx_xvpickod_b(src1, src0);
- tmp2 = __lasx_xvpickev_b(src3, src2);
- tmp3 = __lasx_xvpickod_b(src3, src2);
- reg0 = __lasx_xvmaddwev_h_bu(const_128, tmp1, const_150);
- reg1 = __lasx_xvmaddwev_h_bu(const_128, tmp3, const_150);
- reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
- reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp2);
- dst0 = __lasx_xvpickod_b(reg1, reg0);
- dst0 = __lasx_xvperm_w(dst0, shuff);
- __lasx_xvst(dst0, dst_y, 0);
- dst_y += 32;
- src_argb += 128;
- }
+void ARGBToYRow_LASX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_LASX(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_LASX(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+ ARGBToYMatrixRow_LASX(src_argb, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void ABGRToYRow_LASX(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+ ARGBToYMatrixRow_LASX(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+void ABGRToYJRow_LASX(const uint8_t* src_abgr, uint8_t* dst_yj, int width) {
+ ARGBToYMatrixRow_LASX(src_abgr, dst_yj, width, &kRawJPEGConstants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+static void RGBAToYMatrixRow_LASX(const uint8_t* src_rgba,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
+ int32_t shuff[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+ asm volatile(
+ "xvldrepl.b $xr0, %3, 0 \n\t" // load rgbconstants
+ "xvldrepl.b $xr1, %3, 1 \n\t" // load rgbconstants
+ "xvldrepl.b $xr2, %3, 2 \n\t" // load rgbconstants
+ "xvldrepl.h $xr3, %3, 4 \n\t" // load rgbconstants
+ "xvld $xr20, %4, 0 \n\t" // load shuff
+ "1: \n\t"
+ "xvld $xr4, %0, 0 \n\t"
+ "xvld $xr5, %0, 32 \n\t"
+ "xvld $xr6, %0, 64 \n\t"
+ "xvld $xr7, %0, 96 \n\t" // load 32 pixels of
+ // RGBA
+ "xvor.v $xr12, $xr3, $xr3 \n\t"
+ "xvor.v $xr13, $xr3, $xr3 \n\t"
+ "addi.d %2, %2, -32 \n\t" // 32 processed per
+ // loop.
+ "xvpickev.b $xr8, $xr5, $xr4 \n\t" // AG
+ "xvpickev.b $xr10, $xr7, $xr6 \n\t"
+ "xvpickod.b $xr9, $xr5, $xr4 \n\t" // BR
+ "xvpickod.b $xr11, $xr7, $xr6 \n\t"
+ "xvmaddwev.h.bu $xr12, $xr9, $xr0 \n\t" // B
+ "xvmaddwev.h.bu $xr13, $xr11, $xr0 \n\t"
+ "xvmaddwod.h.bu $xr12, $xr8, $xr1 \n\t" // G
+ "xvmaddwod.h.bu $xr13, $xr10, $xr1 \n\t"
+ "xvmaddwod.h.bu $xr12, $xr9, $xr2 \n\t" // R
+ "xvmaddwod.h.bu $xr13, $xr11, $xr2 \n\t"
+ "addi.d %0, %0, 128 \n\t"
+ "xvpickod.b $xr10, $xr13, $xr12 \n\t"
+ "xvperm.w $xr11, $xr10, $xr20 \n\t"
+ "xvst $xr11, %1, 0 \n\t"
+ "addi.d %1, %1, 32 \n\t"
+ "bnez %2, 1b \n\t"
+ : "+&r"(src_rgba), // %0
+ "+&r"(dst_y), // %1
+ "+&r"(width) // %2
+ : "r"(rgbconstants), "r"(shuff)
+ : "memory");
+}
+
+void RGBAToYRow_LASX(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+ RGBAToYMatrixRow_LASX(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_LASX(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+ RGBAToYMatrixRow_LASX(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_LASX(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+ RGBAToYMatrixRow_LASX(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+static void RGBToYMatrixRow_LASX(const uint8_t* src_rgba,
+ uint8_t* dst_y,
+ int width,
+ const struct RgbConstants* rgbconstants) {
+ int8_t shuff[128] = {
+ 0, 2, 3, 5, 6, 8, 9, 11, 12, 14, 15, 17, 18, 20, 21, 23,
+ 0, 2, 3, 5, 6, 8, 9, 11, 12, 14, 15, 17, 18, 20, 21, 23,
+ 24, 26, 27, 29, 30, 0, 1, 3, 4, 6, 7, 9, 10, 12, 13, 15,
+ 24, 26, 27, 29, 30, 0, 1, 3, 4, 6, 7, 9, 10, 12, 13, 15,
+ 1, 0, 4, 0, 7, 0, 10, 0, 13, 0, 16, 0, 19, 0, 22, 0,
+ 1, 0, 4, 0, 7, 0, 10, 0, 13, 0, 16, 0, 19, 0, 22, 0,
+ 25, 0, 28, 0, 31, 0, 2, 0, 5, 0, 8, 0, 11, 0, 14, 0,
+ 25, 0, 28, 0, 31, 0, 2, 0, 5, 0, 8, 0, 11, 0, 14, 0};
+ asm volatile(
+ "xvldrepl.b $xr0, %3, 0 \n\t" // load rgbconstants
+ "xvldrepl.b $xr1, %3, 1 \n\t" // load rgbconstants
+ "xvldrepl.b $xr2, %3, 2 \n\t" // load rgbconstants
+ "xvldrepl.h $xr3, %3, 4 \n\t" // load rgbconstants
+ "xvld $xr4, %4, 0 \n\t" // load shuff
+ "xvld $xr5, %4, 32 \n\t"
+ "xvld $xr6, %4, 64 \n\t"
+ "xvld $xr7, %4, 96 \n\t"
+ "1: \n\t"
+ "xvld $xr8, %0, 0 \n\t"
+ "xvld $xr9, %0, 32 \n\t"
+ "xvld $xr10, %0, 64 \n\t" // load 32 pixels of
+ // RGB
+ "xvor.v $xr12, $xr3, $xr3 \n\t"
+ "xvor.v $xr13, $xr3, $xr3 \n\t"
+ "xvor.v $xr11, $xr9, $xr9 \n\t"
+ "addi.d %2, %2, -32 \n\t" // 32 processed per
+ // loop.
+ "xvpermi.q $xr9, $xr8, 0x30 \n\t" // src0
+ "xvpermi.q $xr8, $xr10, 0x03 \n\t" // src1
+ "xvpermi.q $xr10, $xr11, 0x30 \n\t" // src2
+ "xvshuf.b $xr14, $xr8, $xr9, $xr4 \n\t"
+ "xvshuf.b $xr15, $xr8, $xr10, $xr5 \n\t"
+ "xvshuf.b $xr16, $xr8, $xr9, $xr6 \n\t"
+ "xvshuf.b $xr17, $xr8, $xr10, $xr7 \n\t"
+ "xvmaddwev.h.bu $xr12, $xr16, $xr1 \n\t" // G
+ "xvmaddwev.h.bu $xr13, $xr17, $xr1 \n\t"
+ "xvmaddwev.h.bu $xr12, $xr14, $xr0 \n\t" // B
+ "xvmaddwev.h.bu $xr13, $xr15, $xr0 \n\t"
+ "xvmaddwod.h.bu $xr12, $xr14, $xr2 \n\t" // R
+ "xvmaddwod.h.bu $xr13, $xr15, $xr2 \n\t"
+ "addi.d %0, %0, 96 \n\t"
+ "xvpickod.b $xr10, $xr13, $xr12 \n\t"
+ "xvst $xr10, %1, 0 \n\t"
+ "addi.d %1, %1, 32 \n\t"
+ "bnez %2, 1b \n\t"
+ : "+&r"(src_rgba), // %0
+ "+&r"(dst_y), // %1
+ "+&r"(width) // %2
+ : "r"(rgbconstants), // %3
+ "r"(shuff) // %4
+ : "memory");
+}
+
+void RGB24ToYJRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+ RGBToYMatrixRow_LASX(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_LASX(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+ RGBToYMatrixRow_LASX(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
+void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_LASX(src_rgb24, dst_y, width, &kRgb24I601Constants);
+}
+
+void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+ RGBToYMatrixRow_LASX(src_raw, dst_y, width, &kRawI601Constants);
}
void ARGBToUVJRow_LASX(const uint8_t* src_argb,
generated by cgit v1.2.3 (git 2.25.1) at 2024-05-27 00:05:10 +0000