Merge remote-tracking branch 'qatar/master'
authorMichael Niedermayer <michaelni@gmx.at>
Wed, 19 Mar 2014 01:51:38 +0000 (02:51 +0100)
committerMichael Niedermayer <michaelni@gmx.at>
Wed, 19 Mar 2014 01:52:08 +0000 (02:52 +0100)
* qatar/master:
  aarch64: float_dsp NEON assembler

Merged-by: Michael Niedermayer <michaelni@gmx.at>
libavutil/aarch64/Makefile
libavutil/aarch64/float_dsp_init.c [new file with mode: 0644]
libavutil/aarch64/float_dsp_neon.S [new file with mode: 0644]
libavutil/float_dsp.c
libavutil/float_dsp.h

index 13d26a1..5613813 100644 (file)
@@ -1 +1,4 @@
-OBJS += aarch64/cpu.o
+OBJS += aarch64/cpu.o                                                 \
+        aarch64/float_dsp_init.o                                      \
+
+NEON-OBJS += aarch64/float_dsp_neon.o
diff --git a/libavutil/aarch64/float_dsp_init.c b/libavutil/aarch64/float_dsp_init.c
new file mode 100644 (file)
index 0000000..4325071
--- /dev/null
@@ -0,0 +1,69 @@
+/*
+ * ARM NEON optimised Float DSP functions
+ * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdint.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/cpu.h"
+#include "libavutil/float_dsp.h"
+#include "cpu.h"
+
+void ff_vector_fmul_neon(float *dst, const float *src0, const float *src1,
+                         int len);
+
+void ff_vector_fmac_scalar_neon(float *dst, const float *src, float mul,
+                                int len);
+
+void ff_vector_fmul_scalar_neon(float *dst, const float *src, float mul,
+                                int len);
+
+void ff_vector_dmul_scalar_neon(double *dst, const double *src, double mul,
+                                int len);
+
+void ff_vector_fmul_window_neon(float *dst, const float *src0,
+                                const float *src1, const float *win, int len);
+
+void ff_vector_fmul_add_neon(float *dst, const float *src0, const float *src1,
+                             const float *src2, int len);
+
+void ff_vector_fmul_reverse_neon(float *dst, const float *src0,
+                                 const float *src1, int len);
+
+void ff_butterflies_float_neon(float *v1, float *v2, int len);
+
+float ff_scalarproduct_float_neon(const float *v1, const float *v2, int len);
+
+av_cold void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+        fdsp->butterflies_float   = ff_butterflies_float_neon;
+        fdsp->scalarproduct_float = ff_scalarproduct_float_neon;
+        fdsp->vector_dmul_scalar  = ff_vector_dmul_scalar_neon;
+        fdsp->vector_fmul         = ff_vector_fmul_neon;
+        fdsp->vector_fmac_scalar  = ff_vector_fmac_scalar_neon;
+        fdsp->vector_fmul_add     = ff_vector_fmul_add_neon;
+        fdsp->vector_fmul_reverse = ff_vector_fmul_reverse_neon;
+        fdsp->vector_fmul_scalar  = ff_vector_fmul_scalar_neon;
+        fdsp->vector_fmul_window  = ff_vector_fmul_window_neon;
+    }
+}
diff --git a/libavutil/aarch64/float_dsp_neon.S b/libavutil/aarch64/float_dsp_neon.S
new file mode 100644 (file)
index 0000000..02d790c
--- /dev/null
@@ -0,0 +1,202 @@
+/*
+ * ARM NEON optimised Float DSP functions
+ * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
+ * Copyright (c) 2014 Janne Grunau <janne-libav@jannau.net>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "config.h"
+#include "asm.S"
+
+function ff_vector_fmul_neon, export=1
+1:      subs            w3,  w3,  #16
+        ld1             {v0.4S, v1.4S}, [x1], #32
+        ld1             {v2.4S, v3.4S}, [x1], #32
+        ld1             {v4.4S, v5.4S}, [x2], #32
+        ld1             {v6.4S, v7.4S}, [x2], #32
+        fmul            v16.4S, v0.4S,  v4.4S
+        fmul            v17.4S, v1.4S,  v5.4S
+        fmul            v18.4S, v2.4S,  v6.4S
+        fmul            v19.4S, v3.4S,  v7.4S
+        st1             {v16.4S, v17.4S}, [x0], #32
+        st1             {v18.4S, v19.4S}, [x0], #32
+        b.ne            1b
+        ret
+endfunc
+
+function ff_vector_fmac_scalar_neon, export=1
+        mov             x3,  #-32
+1:      subs            w2,  w2,  #16
+        ld1             {v16.4S, v17.4S}, [x0], #32
+        ld1             {v18.4S, v19.4S}, [x0], x3
+        ld1             {v4.4S,  v5.4S},  [x1], #32
+        ld1             {v6.4S,  v7.4S},  [x1], #32
+        fmla            v16.4S, v4.4S,  v0.S[0]
+        fmla            v17.4S, v5.4S,  v0.S[0]
+        fmla            v18.4S, v6.4S,  v0.S[0]
+        fmla            v19.4S, v7.4S,  v0.S[0]
+        st1             {v16.4S, v17.4S}, [x0], #32
+        st1             {v18.4S, v19.4S}, [x0], #32
+        b.ne            1b
+        ret
+endfunc
+
+function ff_vector_fmul_scalar_neon, export=1
+        mov             w4,  #15
+        bics            w3,  w2,  w4
+        dup             v16.4S, v0.S[0]
+        b.eq            3f
+        ld1             {v0.4S, v1.4S}, [x1], #32
+1:      subs            w3,  w3,  #16
+        fmul            v0.4S,  v0.4S,  v16.4S
+        ld1             {v2.4S, v3.4S}, [x1], #32
+        fmul            v1.4S,  v1.4S,  v16.4S
+        fmul            v2.4S,  v2.4S,  v16.4S
+        st1             {v0.4S, v1.4S}, [x0], #32
+        fmul            v3.4S,  v3.4S,  v16.4S
+        b.eq            2f
+        ld1             {v0.4S, v1.4S}, [x1], #32
+        st1             {v2.4S, v3.4S}, [x0], #32
+        b               1b
+2:      ands            w2,  w2,  #15
+        st1             {v2.4S, v3.4S}, [x0], #32
+        b.eq            4f
+3:      ld1             {v0.4S}, [x1], #16
+        fmul            v0.4S,  v0.4S,  v16.4S
+        st1             {v0.4S}, [x0], #16
+        subs            w2,  w2,  #4
+        b.gt            3b
+4:      ret
+endfunc
+
+function ff_vector_dmul_scalar_neon, export=1
+        dup             v16.2D, v0.D[0]
+        ld1             {v0.2D, v1.2D}, [x1], #32
+1:      subs            w2,  w2,  #8
+        fmul            v0.2D,  v0.2D,  v16.2D
+        ld1             {v2.2D, v3.2D}, [x1], #32
+        fmul            v1.2D,  v1.2D,  v16.2D
+        fmul            v2.2D,  v2.2D,  v16.2D
+        st1             {v0.2D, v1.2D}, [x0], #32
+        fmul            v3.2D,  v3.2D,  v16.2D
+        ld1             {v0.2D, v1.2D}, [x1], #32
+        st1             {v2.2D, v3.2D}, [x0], #32
+        b.gt            1b
+        ret
+endfunc
+
+function ff_vector_fmul_window_neon, export=1
+        sxtw            x4,  w4                 // len
+        sub             x2,  x2,  #8
+        sub             x5,  x4,  #2
+        add             x2,  x2,  x5, lsl #2    // src1 + 4 * (len - 4)
+        add             x6,  x3,  x5, lsl #3    // win  + 8 * (len - 2)
+        add             x5,  x0,  x5, lsl #3    // dst  + 8 * (len - 2)
+        mov             x7,  #-16
+        ld1             {v0.4S},  [x1], #16     // s0
+        ld1             {v2.4S},  [x3], #16     // wi
+        ld1             {v1.4S},  [x2], x7      // s1
+1:      ld1             {v3.4S},  [x6], x7      // wj
+        subs            x4,  x4,  #4
+        fmul            v17.4S, v0.4S,  v2.4S   // s0 * wi
+        rev64           v4.4S,  v1.4S
+        rev64           v5.4S,  v3.4S
+        rev64           v17.4S, v17.4S
+        ext             v4.16B,  v4.16B,  v4.16B,  #8 // s1_r
+        ext             v5.16B,  v5.16B,  v5.16B,  #8 // wj_r
+        ext             v17.16B, v17.16B, v17.16B, #8 // (s0 * wi)_rev
+        fmul            v16.4S, v0.4S,  v5.4S  // s0 * wj_r
+        fmla            v17.4S, v1.4S,  v3.4S  // (s0 * wi)_rev + s1 * wj
+        b.eq            2f
+        ld1             {v0.4S},  [x1], #16
+        fmls            v16.4S, v4.4S,  v2.4S  // s0 * wj_r - s1_r * wi
+        st1             {v17.4S}, [x5], x7
+        ld1             {v2.4S},  [x3], #16
+        ld1             {v1.4S},  [x2], x7
+        st1             {v16.4S}, [x0], #16
+        b               1b
+2:
+        fmls            v16.4S, v4.4S,  v2.4S  // s0 * wj_r - s1_r * wi
+        st1             {v17.4S}, [x5], x7
+        st1             {v16.4S}, [x0], #16
+        ret
+endfunc
+
+function ff_vector_fmul_add_neon, export=1
+        ld1             {v0.4S, v1.4S},  [x1], #32
+        ld1             {v2.4S, v3.4S},  [x2], #32
+        ld1             {v4.4S, v5.4S},  [x3], #32
+1:      subs            w4,  w4,  #8
+        fmla            v4.4S,  v0.4S,  v2.4S
+        fmla            v5.4S,  v1.4S,  v3.4S
+        b.eq            2f
+        ld1             {v0.4S, v1.4S},  [x1], #32
+        ld1             {v2.4S, v3.4S},  [x2], #32
+        st1             {v4.4S, v5.4S},  [x0], #32
+        ld1             {v4.4S, v5.4S},  [x3], #32
+        b               1b
+2:      st1             {v4.4S, v5.4S},  [x0], #32
+        ret
+endfunc
+
+function ff_vector_fmul_reverse_neon, export=1
+        sxtw            x3,  w3
+        add             x2,  x2,  x3,  lsl #2
+        sub             x2,  x2,  #32
+        mov             x4, #-32
+        ld1             {v2.4S, v3.4S},  [x2], x4
+        ld1             {v0.4S, v1.4S},  [x1], #32
+1:      subs            x3,  x3,  #8
+        rev64           v3.4S,  v3.4S
+        rev64           v2.4S,  v2.4S
+        ext             v3.16B, v3.16B, v3.16B,  #8
+        ext             v2.16B, v2.16B, v2.16B,  #8
+        fmul            v16.4S, v0.4S,  v3.4S
+        fmul            v17.4S, v1.4S,  v2.4S
+        b.eq            2f
+        ld1             {v2.4S, v3.4S},  [x2], x4
+        ld1             {v0.4S, v1.4S},  [x1], #32
+        st1             {v16.4S, v17.4S},  [x0], #32
+        b               1b
+2:      st1             {v16.4S, v17.4S},  [x0], #32
+        ret
+endfunc
+
+function ff_butterflies_float_neon, export=1
+1:      ld1             {v0.4S}, [x0]
+        ld1             {v1.4S}, [x1]
+        subs            w2,  w2,  #4
+        fsub            v2.4S,   v0.4S,  v1.4S
+        fadd            v3.4S,   v0.4S,  v1.4S
+        st1             {v2.4S}, [x1],   #16
+        st1             {v3.4S}, [x0],   #16
+        b.gt            1b
+        ret
+endfunc
+
+function ff_scalarproduct_float_neon, export=1
+        movi            v2.4S,  #0
+1:      ld1             {v0.4S}, [x0],   #16
+        ld1             {v1.4S}, [x1],   #16
+        subs            w2,      w2,     #4
+        fmla            v2.4S,   v0.4S,  v1.4S
+        b.gt            1b
+        faddp           v0.4S,   v2.4S,  v2.4S
+        faddp           s0,      v0.2S
+        ret
+endfunc
index e50dab0..be85d20 100644 (file)
@@ -127,7 +127,9 @@ av_cold void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int bit_exact)
     fdsp->butterflies_float = butterflies_float_c;
     fdsp->scalarproduct_float = avpriv_scalarproduct_float_c;
 
-#if ARCH_ARM
+#if   ARCH_AARCH64
+    ff_float_dsp_init_aarch64(fdsp);
+#elif ARCH_ARM
     ff_float_dsp_init_arm(fdsp);
 #elif ARCH_PPC
     ff_float_dsp_init_ppc(fdsp, bit_exact);
index 3424cb5..7fc851b 100644 (file)
@@ -179,6 +179,7 @@ float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len);
 void avpriv_float_dsp_init(AVFloatDSPContext *fdsp, int strict);
 
 
+void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp);
 void ff_float_dsp_init_arm(AVFloatDSPContext *fdsp);
 void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int strict);
 void ff_float_dsp_init_x86(AVFloatDSPContext *fdsp);