Add loongarch support and LSX SIMD optimizations

Enable LSX by default:
    ./configure && make

Disable LSX:
    ./configure --enable-loongarch-lsx=no && make

Signed-off-by: Cosmin Truta <ctruta@gmail.com>

Makefile.am

@@ -127,6 +127,15 @@ libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES += powerpc/powerpc_init.c\
         powerpc/filter_vsx_intrinsics.c
 endif
 
+if PNG_LOONGARCH_LSX
+noinst_LTLIBRARIES= libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx.la
+libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx_la_SOURCES = loongarch/loongarch_lsx_init.c\
+	loongarch/filter_lsx_intrinsics.c
+libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx_la_CFLAGS = -mlsx
+libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_LIBADD = libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx.la
+# libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_DEPENDENCIES = libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx.la
+endif
+
 nodist_libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_SOURCES = pnglibconf.h
 
 libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_LDFLAGS = -no-undefined -export-dynamic \
@@ -147,6 +156,10 @@ else
   libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_DEPENDENCIES = libpng.sym
 endif
 
+if PNG_LOONGARCH_LSX
+  libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@_la_DEPENDENCIES += libpng@PNGLIB_MAJOR@@PNGLIB_MINOR@lsx.la
+endif
+
 #distribute headers in /usr/include/libpng/*
 pkgincludedir= $(includedir)/$(PNGLIB_BASENAME)
 pkginclude_HEADERS= png.h pngconf.h

configure.ac

@@ -334,6 +334,9 @@ AC_ARG_ENABLE([hardware-optimizations],
          enable_intel_sse=no
          AC_DEFINE([PNG_INTEL_SSE_OPT], [0],
            [Disable INTEL_SSE optimizations])
+         enable_loongarch_lsx=no
+         AC_DEFINE([PNG_LOONGARCH_LSX_OPT], [0],
+           [Disable LOONGARCH_LSX optimizations])
          ;;
       *)
          # allow enabling hardware optimization on any system:
@@ -358,6 +361,11 @@ AC_ARG_ENABLE([hardware-optimizations],
               AC_DEFINE([PNG_POWERPC_VSX_OPT], [2],
                 [Enable POWERPC VSX optimizations])
               ;;
+            loongarch*)
+              enable_loongarch_lsx=yes
+              AC_DEFINE([PNG_LOONGARCH_LSX_OPT], [1],
+                [Enable LOONGARCH_LSX optimizations])
+              ;;
          esac
          ;;
    esac])
@@ -535,6 +543,69 @@ AM_CONDITIONAL([PNG_POWERPC_VSX],
       powerpc*|ppc64*) : ;;
     esac])
 
+# LOONGARCH
+# ===
+#
+# LOONGARCH LSX (SIMD) support
+
+if test "$LSX_CFLAGS" = ''; then
+    LSX_CFLAGS="-mlsx"
+fi
+
+compiler_support_loongarch_lsx=no
+AC_MSG_CHECKING(whether to use loongarch LSX intrinsics)
+save_CFLAGS=$CFLAGS
+CFLAGS="$CFLAGS $LSX_CFLAGS"
+AC_COMPILE_IFELSE([AC_LANG_SOURCE([[
+#include<lsxintrin.h>
+int main(){
+    __m128i a, b, c;
+    a = __lsx_vadd_w(b, c);
+    return 0;
+}]])],compiler_support_loongarch_lsx=yes)
+CFLAGS=$save_CFLAGS
+AC_MSG_RESULT($compiler_support_loongarch_lsx)
+
+AC_ARG_ENABLE([loongarch-lsx],
+   AS_HELP_STRING([[[--enable-loongarch-lsx]]],
+      [Enable LOONGARCH LSX optimizations: =no/off, yes/on:]
+      [no/off: disable the optimizations;]
+      [yes/on: turn on unconditionally.]
+      [If not specified: determined by the compiler.]),
+   [case "$enableval" in
+      no|off)
+         # disable the default enabling on __loongarch_simd systems:
+         AC_DEFINE([PNG_LOONGARCH_LSX_OPT], [0],
+                   [Disable LOONGARCH LSX optimizations])
+         # Prevent inclusion of the assembler files below:
+         enable_loongarch_lsx=no;;
+      yes|on)
+         AC_DEFINE([PNG_LOONGARCH_LSX_OPT], [1],
+                   [Enable LOONGARCH LSX optimizations])
+         ;;
+      *)
+         AC_MSG_ERROR([--enable-loongarch-lsx=${enable_loongarch_lsx}: invalid value])
+   esac])
+
+if test "$enable_loongarch_lsx" != 'no'; then
+   if test $compiler_support_loongarch_lsx = yes; then
+      AC_DEFINE([PNG_LOONGARCH_LSX_OPT], [1], [Enable LOONGARCH LSX optimizations])
+   else
+      AC_MSG_WARN([Compiler does not support loongarch LSX.])
+   fi
+fi
+
+# Add LOONGARCH specific files to all builds where the host_cpu is loongarch ('loongarch*') or
+# where LOONGARCH optimizations were explicitly requested (this allows a fallback if a
+# future host CPU does not match 'loongarch*')
+
+AM_CONDITIONAL([PNG_LOONGARCH_LSX],
+   [test "$enable_loongarch_lsx" != 'no' && test $compiler_support_loongarch_lsx = yes &&
+    case "$host_cpu" in
+      loongarch*) :;;
+      *)    test "$enable_loongarch_lsx" != '';;
+    esac])
+
 AC_MSG_NOTICE([[Extra options for compiler: $PNG_COPTS]])
 
 # Config files, substituting as above

loongarch/filter_lsx_intrinsics.c

@@ -0,0 +1,412 @@
+/* filter_lsx_intrinsics.c - LSX optimized filter functions
+ *
+ * Copyright (c) 2021 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2016 Glenn Randers-Pehrson
+ * Contributed by Jin Bo (jinbo@loongson.cn)
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+#if PNG_LOONGARCH_LSX_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
+
+#include <lsxintrin.h>
+
+#define LSX_LD(psrc) __lsx_vld((psrc), 0)
+
+#define LSX_LD_2(psrc, stride, out0, out1) \
+{                                          \
+   out0 = LSX_LD(psrc);                    \
+   out1 = LSX_LD(psrc + stride);           \
+}
+
+#define LSX_LD_4(psrc, stride, out0, out1, out2, out3) \
+{                                                      \
+   LSX_LD_2(psrc, stride, out0, out1);                 \
+   LSX_LD_2(psrc + stride * 2, stride, out2, out3);    \
+}
+
+#define LSX_ST(in, pdst) __lsx_vst(in, (pdst), 0)
+
+#define LSX_ST_2(in0, in1, pdst, stride) \
+{                                        \
+   LSX_ST(in0, pdst);                    \
+   LSX_ST(in1, pdst + stride);           \
+}
+
+#define LSX_ST_4(in0, in1, in2, in3, pdst, stride) \
+{                                                  \
+   LSX_ST_2(in0, in1, pdst, stride);               \
+   LSX_ST_2(in2, in3, pdst + stride * 2, stride);  \
+}
+
+#define LSX_ADD_B(in0, in1, out0) \
+{                                 \
+   out0 = __lsx_vadd_b(in0, in1); \
+}
+
+#define LSX_ADD_B_2(in0, in1, in2, in3, out0, out1) \
+{                                                   \
+   LSX_ADD_B(in0, in1, out0);                       \
+   LSX_ADD_B(in2, in3, out1);                       \
+}
+
+#define LSX_ADD_B_4(in0, in1, in2, in3, in4, in5,     \
+                    in6, in7, out0, out1, out2, out3) \
+{                                                     \
+   LSX_ADD_B_2(in0, in1, in2, in3, out0, out1);       \
+   LSX_ADD_B_2(in4, in5, in6, in7, out2, out3);       \
+}
+
+#define LSX_ABS_B_3(in0, in1, in2, out0, out1, out2) \
+{                                                    \
+   out0 = __lsx_vadda_h(in0, zero);                  \
+   out1 = __lsx_vadda_h(in1, zero);                  \
+   out2 = __lsx_vadda_h(in2, zero);                  \
+}
+
+#define LSX_ILVL_B(in_h, in_l, out0)  \
+{                                     \
+   out0 = __lsx_vilvl_b(in_h, in_l);  \
+}
+
+#define LSX_ILVL_B_2(in0_h, in0_l, in1_h, in1_l, out0, out1) \
+{                                                            \
+   LSX_ILVL_B(in0_h, in0_l, out0);                           \
+   LSX_ILVL_B(in1_h, in1_l, out1);                           \
+}
+
+#define LSX_HSUB_HU_BU_2(in0, in1, out0, out1) \
+{                                              \
+   out0 = __lsx_vhsubw_hu_bu(in0, in0);        \
+   out1 = __lsx_vhsubw_hu_bu(in1, in1);        \
+}
+
+#define LSX_CMP_PICK_SMALLER(in0, in1, in2, in3, in4, in5, out0) \
+{                                                                \
+   __m128i _cmph, _cmpb, _in0, _in3;                             \
+   _cmph = __lsx_vslt_h(in1, in0);                               \
+   _cmpb = __lsx_vpickev_b(_cmph, _cmph);                        \
+   _in0  = __lsx_vmin_bu(in0,in1);                               \
+   _in3  = __lsx_vbitsel_v(in3, in4, _cmpb);                     \
+   _cmph = __lsx_vslt_h(in2, _in0);                              \
+   _cmpb = __lsx_vpickev_b(_cmph, _cmph);                        \
+   _in3  = __lsx_vbitsel_v(_in3, in5, _cmpb);                    \
+   out0  = __lsx_vadd_b(out0, _in3);                             \
+}
+
+void png_read_filter_row_up_lsx(png_row_infop row_info, png_bytep row,
+                                png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   __m128i vec_0, vec_1, vec_2, vec_3;
+   __m128i vec_4, vec_5, vec_6, vec_7;
+
+   while (n >= 64)
+   {
+      LSX_LD_4(rp, 16, vec_0, vec_1, vec_2, vec_3);
+      LSX_LD_4(pp, 16, vec_4, vec_5, vec_6, vec_7);
+      pp += 64;
+      LSX_ADD_B_4(vec_0 ,vec_4, vec_1, vec_5, vec_2, vec_6,
+                  vec_3, vec_7, vec_0, vec_1, vec_2, vec_3);
+      LSX_ST_4(vec_0, vec_1, vec_2, vec_3, rp, 16);
+      rp += 64;
+      n -= 64;
+   }
+   if (n & 63)
+   {
+      if (n >= 32)
+      {
+         LSX_LD_2(rp, 16, vec_0, vec_1);
+         LSX_LD_2(pp, 16, vec_2, vec_3);
+         pp += 32;
+         LSX_ADD_B_2(vec_0, vec_2, vec_1, vec_3, vec_0, vec_1);
+         LSX_ST_2(vec_0, vec_1, rp, 16);
+         rp += 32;
+         n -= 32;
+      }
+      if (n & 31)
+      {
+         if (n >= 16)
+         {
+            vec_0 = LSX_LD(rp);
+            vec_1 = LSX_LD(pp);
+            pp += 16;
+            LSX_ADD_B(vec_0, vec_1, vec_0);
+            LSX_ST(vec_0, rp);
+            rp += 16;
+            n -= 16;
+         }
+         if (n >= 8)
+         {
+            vec_0 = __lsx_vldrepl_d(rp, 0);
+            vec_1 = __lsx_vldrepl_d(pp, 0);
+            vec_0 = __lsx_vadd_b(vec_0, vec_1);
+            __lsx_vstelm_d(vec_0, rp, 0, 0);
+            rp += 8;
+            pp += 8;
+            n -= 8;
+         }
+         while (n--)
+         {
+            *rp = *rp + *pp++;
+            rp++;
+         }
+      }
+   }
+}
+
+void png_read_filter_row_sub3_lsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   png_uint_32 tmp;
+   png_bytep nxt = row;
+   __m128i vec_0, vec_1;
+
+   PNG_UNUSED(prev_row);
+
+   vec_0 = __lsx_vldrepl_w(nxt, 0);
+   nxt += 3;
+   n -= 3;
+
+   while (n >= 3)
+   {
+      vec_1 = __lsx_vldrepl_w(nxt, 0);
+      vec_1 = __lsx_vadd_b(vec_1, vec_0);
+      __lsx_vstelm_h(vec_1, nxt, 0, 0);
+      vec_0 = vec_1;
+      nxt += 2;
+      __lsx_vstelm_b(vec_1, nxt, 0, 2);
+      nxt += 1;
+      n -= 3;
+   }
+
+   row = nxt - 3;
+   while (n--)
+   {
+      *nxt = *nxt + *row++;
+      nxt++;
+   }
+}
+
+void png_read_filter_row_sub4_lsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   __m128i vec_0, vec_1;
+
+   PNG_UNUSED(prev_row);
+
+   vec_0 = __lsx_vldrepl_w(row, 0);
+   row += 4;
+   n -= 4;
+
+   while (n >= 4)
+   {
+      vec_1 = __lsx_vldrepl_w(row, 0);
+      vec_1 = __lsx_vadd_b(vec_1, vec_0);
+      __lsx_vstelm_w(vec_1, row, 0, 0);
+      vec_0 = vec_1;
+      row += 4;
+      n -= 4;
+   }
+}
+
+void png_read_filter_row_avg3_lsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   png_bytep nxt = row;
+   png_const_bytep prev_nxt = prev_row;
+   __m128i vec_0, vec_1, vec_2;
+
+   vec_0 = __lsx_vldrepl_w(nxt, 0);
+   vec_1 = __lsx_vldrepl_w(prev_nxt, 0);
+   prev_nxt += 3;
+   vec_1 = __lsx_vsrli_b(vec_1, 1);
+   vec_1 = __lsx_vadd_b(vec_1, vec_0);
+   __lsx_vstelm_h(vec_1, nxt, 0, 0);
+   nxt += 2;
+   __lsx_vstelm_b(vec_1, nxt, 0, 2);
+   nxt += 1;
+   n -= 3;
+
+   while (n >= 3)
+   {
+      vec_2 = vec_1;
+      vec_0 = __lsx_vldrepl_w(nxt, 0);
+      vec_1 = __lsx_vldrepl_w(prev_nxt, 0);
+      prev_nxt += 3;
+
+      vec_1 = __lsx_vavg_bu(vec_1, vec_2);
+      vec_1 = __lsx_vadd_b(vec_1, vec_0);
+
+      __lsx_vstelm_h(vec_1, nxt, 0, 0);
+      nxt += 2;
+      __lsx_vstelm_b(vec_1, nxt, 0, 2);
+      nxt += 1;
+      n -= 3;
+   }
+
+   row = nxt - 3;
+   while (n--)
+   {
+      vec_2 = __lsx_vldrepl_b(row, 0);
+      row++;
+      vec_0 = __lsx_vldrepl_b(nxt, 0);
+      vec_1 = __lsx_vldrepl_b(prev_nxt, 0);
+      prev_nxt++;
+
+      vec_1 = __lsx_vavg_bu(vec_1, vec_2);
+      vec_1 = __lsx_vadd_b(vec_1, vec_0);
+
+      __lsx_vstelm_b(vec_1, nxt, 0, 0);
+      nxt++;
+   }
+}
+
+void png_read_filter_row_avg4_lsx(png_row_infop row_info, png_bytep row,
+                                  png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   __m128i vec_0, vec_1, vec_2;
+
+   vec_0 = __lsx_vldrepl_w(row, 0);
+   vec_1 = __lsx_vldrepl_w(prev_row, 0);
+   prev_row += 4;
+   vec_1 = __lsx_vsrli_b(vec_1, 1);
+   vec_1 = __lsx_vadd_b(vec_1, vec_0);
+   __lsx_vstelm_w(vec_1, row, 0, 0);
+   row += 4;
+   n -= 4;
+
+   while (n >= 4)
+   {
+      vec_2 = vec_1;
+      vec_0 = __lsx_vldrepl_w(row, 0);
+      vec_1 = __lsx_vldrepl_w(prev_row, 0);
+      prev_row += 4;
+
+      vec_1 = __lsx_vavg_bu(vec_1, vec_2);
+      vec_1 = __lsx_vadd_b(vec_1, vec_0);
+
+      __lsx_vstelm_w(vec_1, row, 0, 0);
+      row += 4;
+      n -= 4;
+   }
+}
+
+void png_read_filter_row_paeth3_lsx(png_row_infop row_info,
+                                    png_bytep row,
+                                    png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   png_bytep nxt = row;
+   png_const_bytep prev_nxt = prev_row;
+   __m128i vec_a, vec_b, vec_c, vec_d;
+   __m128i vec_pa, vec_pb, vec_pc;
+   __m128i zero = {0};
+
+   vec_a = __lsx_vldrepl_w(nxt, 0);
+   vec_b = __lsx_vldrepl_w(prev_nxt, 0);
+   prev_nxt += 3;
+   vec_d = __lsx_vadd_b(vec_a, vec_b);
+   __lsx_vstelm_h(vec_d, nxt, 0, 0);
+   nxt += 2;
+   __lsx_vstelm_b(vec_d, nxt, 0, 2);
+   nxt += 1;
+   n -= 3;
+
+   while (n >= 3)
+   {
+      vec_a = vec_d;
+      vec_c = vec_b;
+      vec_b = __lsx_vldrepl_w(prev_nxt, 0);
+      prev_nxt += 3;
+      vec_d = __lsx_vldrepl_w(nxt, 0);
+
+      LSX_ILVL_B_2(vec_b, vec_c, vec_a, vec_c, vec_pa, vec_pb);
+      LSX_HSUB_HU_BU_2(vec_pa, vec_pb, vec_pa, vec_pb);
+      vec_pc = __lsx_vadd_h(vec_pa, vec_pb);
+      LSX_ABS_B_3(vec_pa, vec_pb, vec_pc, vec_pa, vec_pb, vec_pc);
+      LSX_CMP_PICK_SMALLER(vec_pa, vec_pb, vec_pc, vec_a, vec_b, vec_c, vec_d);
+
+      __lsx_vstelm_h(vec_d, nxt, 0, 0);
+      nxt += 2;
+      __lsx_vstelm_b(vec_d, nxt, 0, 2);
+      nxt += 1;
+      n -= 3;
+   }
+
+   prev_row = prev_nxt - 3;
+   row = nxt - 3;
+   while (n--)
+   {
+      vec_a = __lsx_vldrepl_b(row, 0);
+      row++;
+      vec_b = __lsx_vldrepl_b(prev_nxt, 0);
+      prev_nxt++;
+      vec_c = __lsx_vldrepl_b(prev_row, 0);
+      prev_row++;
+      vec_d = __lsx_vldrepl_b(nxt, 0);
+
+      LSX_ILVL_B_2(vec_b, vec_c, vec_a, vec_c, vec_pa, vec_pb);
+      LSX_HSUB_HU_BU_2(vec_pa, vec_pb, vec_pa, vec_pb);
+      vec_pc = __lsx_vadd_h(vec_pa, vec_pb);
+      LSX_ABS_B_3(vec_pa, vec_pb, vec_pc, vec_pa, vec_pb, vec_pc);
+      LSX_CMP_PICK_SMALLER(vec_pa, vec_pb, vec_pc, vec_a, vec_b, vec_c, vec_d);
+
+      __lsx_vstelm_b(vec_d, nxt, 0, 0);
+      nxt++;
+   }
+}
+
+void png_read_filter_row_paeth4_lsx(png_row_infop row_info,
+                                    png_bytep row,
+                                    png_const_bytep prev_row)
+{
+   size_t n = row_info->rowbytes;
+   __m128i vec_a, vec_b, vec_c, vec_d;
+   __m128i vec_pa, vec_pb, vec_pc;
+   __m128i zero = {0};
+
+   vec_a = __lsx_vldrepl_w(row, 0);
+   vec_b = __lsx_vldrepl_w(prev_row, 0);
+   prev_row += 4;
+   vec_d = __lsx_vadd_b(vec_a, vec_b);
+   __lsx_vstelm_w(vec_d, row, 0, 0);
+   row += 4;
+   n -= 4;
+
+   while (n >= 4)
+   {
+      vec_a = vec_d;
+      vec_c = vec_b;
+      vec_b = __lsx_vldrepl_w(prev_row, 0);
+      prev_row += 4;
+      vec_d = __lsx_vldrepl_w(row, 0);
+
+      LSX_ILVL_B_2(vec_b, vec_c, vec_a, vec_c, vec_pa, vec_pb);
+      LSX_HSUB_HU_BU_2(vec_pa, vec_pb, vec_pa, vec_pb);
+      vec_pc = __lsx_vadd_h(vec_pa, vec_pb);
+      LSX_ABS_B_3(vec_pa, vec_pb, vec_pc, vec_pa, vec_pb, vec_pc);
+      LSX_CMP_PICK_SMALLER(vec_pa, vec_pb, vec_pc, vec_a, vec_b, vec_c, vec_d);
+
+      __lsx_vstelm_w(vec_d, row, 0, 0);
+      row += 4;
+      n -= 4;
+   }
+}
+
+#endif /* PNG_LOONGARCH_LSX_IMPLEMENTATION == 1 (intrinsics) */
+#endif /* PNG_READ_SUPPORTED */

loongarch/loongarch_lsx_init.c

@@ -0,0 +1,65 @@
+/* loongarch_lsx_init.c - LSX optimized filter functions
+ *
+ * Copyright (c) 2021 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Contributed by Jin Bo <jinbo@loongson.cn>
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_LOONGARCH_LSX_IMPLEMENTATION == 1
+
+#include <sys/auxv.h>
+
+#define LA_HWCAP_LSX    (1<<4)
+static int png_has_lsx(void)
+{
+    int flags = 0;
+    int flag  = (int)getauxval(AT_HWCAP);
+
+    if (flag & LA_HWCAP_LSX)
+        return 1;
+
+    return 0;
+}
+
+void
+png_init_filter_functions_lsx(png_structp pp, unsigned int bpp)
+{
+   /* IMPORTANT: any new external functions used here must be declared using
+    * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
+    * 'prefix' option to configure works:
+    *
+    *    ./configure --with-libpng-prefix=foobar_
+    *
+    * Verify you have got this right by running the above command, doing a build
+    * and examining pngprefix.h; it must contain a #define for every external
+    * function you add.  (Notice that this happens automatically for the
+    * initialization function.)
+    */
+
+   if (png_has_lsx())
+   {
+      pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_lsx;
+      if (bpp == 3)
+      {
+         pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_lsx;
+         pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_lsx;
+         pp->read_filter[PNG_FILTER_VALUE_PAETH-1] = png_read_filter_row_paeth3_lsx;
+      }
+      else if (bpp == 4)
+      {
+         pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_lsx;
+         pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_lsx;
+         pp->read_filter[PNG_FILTER_VALUE_PAETH-1] = png_read_filter_row_paeth4_lsx;
+      }
+   }
+}
+
+#endif /* PNG_LOONGARCH_LSX_IMPLEMENTATION == 1 */
+#endif /* PNG_READ_SUPPORTED */

pngpriv.h

@@ -276,6 +276,12 @@
 #  define PNG_POWERPC_VSX_IMPLEMENTATION 0
 #endif
 
+#if PNG_LOONGARCH_LSX_OPT > 0
+#   define PNG_FILTER_OPTIMIZATIONS png_init_filter_functions_lsx
+#   define PNG_LOONGARCH_LSX_IMPLEMENTATION 1
+#else
+#   define PNG_LOONGARCH_LSX_IMPLEMENTATION 0
+#endif
 
 /* Is this a build of a DLL where compilation of the object modules requires
  * different preprocessor settings to those required for a simple library?  If
@@ -1355,6 +1361,23 @@ PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_sse2,(png_row_infop
     row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
 #endif
 
+#if PNG_LOONGARCH_LSX_IMPLEMENTATION == 1
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_up_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub3_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_sub4_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg3_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_avg4_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth3_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+PNG_INTERNAL_FUNCTION(void,png_read_filter_row_paeth4_lsx,(png_row_infop
+    row_info, png_bytep row, png_const_bytep prev_row),PNG_EMPTY);
+#endif
+
 /* Choose the best filter to use and filter the row data */
 PNG_INTERNAL_FUNCTION(void,png_write_find_filter,(png_structrp png_ptr,
     png_row_infop row_info),PNG_EMPTY);
@@ -2105,6 +2128,11 @@ PNG_INTERNAL_FUNCTION(void, png_init_filter_functions_sse2,
 #  endif
 #endif
 
+#if PNG_LOONGARCH_LSX_OPT > 0
+PNG_INTERNAL_FUNCTION(void, png_init_filter_functions_lsx,
+    (png_structp png_ptr, unsigned int bpp), PNG_EMPTY);
+#endif
+
 PNG_INTERNAL_FUNCTION(png_uint_32, png_check_keyword, (png_structrp png_ptr,
    png_const_charp key, png_bytep new_key), PNG_EMPTY);