1. ffmpeg
  2. Commit 5e586e1befaab7b77883d73c0ebcb19f1d786ed8
Browse code

huffyuvencdsp: Add ff_diff_bytes_{sse2,avx2}

SSE2 version 4%-35% faster than MMX depending on the width.
AVX2 version 1%-13% faster than SSE2 depending on the width.

Timothy Gu authored on 2015/10/19 10:25:11
Showing 3 changed files
libavcodec/huffyuvenc.c
History View file @ 5e586e1
... ... 
@@ -60,12 +60,12 @@ static inline int sub_left_prediction(HYuvContext *s, uint8_t *dst,
60 60 
             }
61 61 
             return left;
62 62 
         } else {
63 
-            for (i = 0; i < 16; i++) {
63 
+            for (i = 0; i < 32; i++) {
64 64 
                 const int temp = src[i];
65 65 
                 dst[i] = temp - left;
66 66 
                 left   = temp;
67 67 
             }
68 
-            s->hencdsp.diff_bytes(dst + 16, src + 16, src + 15, w - 16);
68 
+            s->hencdsp.diff_bytes(dst + 32, src + 32, src + 31, w - 32);
69 69 
             return src[w-1];
70 70 
         }
71 71 
     } else {
libavcodec/x86/huffyuvencdsp.asm
History View file @ 5e586e1
... ... 
@@ -27,9 +27,9 @@
27 27
28 28 
 section .text
29 29
30 
-INIT_MMX mmx
31 30 
 ; void ff_diff_bytes_mmx(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
32 31 
 ;                        intptr_t w);
32 
+%macro DIFF_BYTES_PROLOGUE 0
33 33 
 %if ARCH_X86_32
34 34 
 cglobal diff_bytes, 3,5,2, dst, src1, src2
35 35 
 %define wq r4q
... ... 
@@ -40,34 +40,108 @@ cglobal diff_bytes, 4,5,2, dst, src1, src2, w
40 40 
     DECLARE_REG_TMP 4
41 41 
 %endif ; ARCH_X86_32
42 42 
 %define i t0q
43 
+%endmacro
44 
+
45 
+; label to jump to if w < regsize
46 
+%macro DIFF_BYTES_LOOP_PREP 1
43 47 
     mov                i, wq
44 
-    and                i, -2 * mmsize
45 
-        jz  .setup_loop2
48 
+    and                i, -2 * regsize
49 
+        jz            %1
46 50 
     add             dstq, i
47 51 
     add            src1q, i
48 52 
     add            src2q, i
49 53 
     neg                i
50 
-.loop:
51 
-    mova              m0, [src1q + i]
52 
-    mova              m1, [src1q + i + mmsize]
53 
-    psubb             m0, [src2q + i]
54 
-    psubb             m1, [src2q + i + mmsize]
55 
-    mova          [dstq + i], m0
56 
-    mova [mmsize + dstq + i], m1
57 
-    add                i, 2 * mmsize
58 
-        jl         .loop
59 
-.setup_loop2:
60 
-    and               wq, 2 * mmsize - 1
61 
-        jz          .end
54 
+%endmacro
55 
+
56 
+; mov type used for src1q, dstq, first reg, second reg
57 
+%macro DIFF_BYTES_LOOP_CORE 4
58 
+%if regsize != 16
59 
+    mov%1             %3, [src1q + i]
60 
+    mov%1             %4, [src1q + i + regsize]
61 
+    psubb             %3, [src2q + i]
62 
+    psubb             %4, [src2q + i + regsize]
63 
+    mov%2           [dstq + i], %3
64 
+    mov%2 [regsize + dstq + i], %4
65 
+%else
66 
+    ; SSE enforces alignment of psubb operand
67 
+    mov%1             %3, [src1q + i]
68 
+    movu              %4, [src2q + i]
69 
+    psubb             %3, %4
70 
+    mov%2     [dstq + i], %3
71 
+    mov%1             %3, [src1q + i + regsize]
72 
+    movu              %4, [src2q + i + regsize]
73 
+    psubb             %3, %4
74 
+    mov%2 [regsize + dstq + i], %3
75 
+%endif
76 
+%endmacro
77 
+
78 
+%macro DIFF_BYTES_BODY 2 ; mov type used for src1q, for dstq
79 
+    %define regsize mmsize
80 
+.loop_%1%2:
81 
+    DIFF_BYTES_LOOP_CORE %1, %2, m0, m1
82 
+    add                i, 2 * regsize
83 
+        jl    .loop_%1%2
84 
+.skip_main_%1%2:
85 
+    and               wq, 2 * regsize - 1
86 
+        jz     .end_%1%2
87 
+%if mmsize > 16
88 
+    ; fall back to narrower xmm
89 
+    %define regsize mmsize / 2
90 
+    DIFF_BYTES_LOOP_PREP .setup_loop_gpr_aa
91 
+.loop2_%1%2:
92 
+    DIFF_BYTES_LOOP_CORE %1, %2, xm0, xm1
93 
+    add                i, 2 * regsize
94 
+        jl   .loop2_%1%2
95 
+.setup_loop_gpr_%1%2:
96 
+    and               wq, 2 * regsize - 1
97 
+        jz     .end_%1%2
98 
+%endif
62 99 
     add             dstq, wq
63 100 
     add            src1q, wq
64 101 
     add            src2q, wq
65 102 
     neg               wq
66 
-.loop2:
103 
+.loop_gpr_%1%2:
67 104 
     mov              t0b, [src1q + wq]
68 105 
     sub              t0b, [src2q + wq]
69 106 
     mov      [dstq + wq], t0b
70 107 
     inc               wq
71 
-        jl        .loop2
72 
-.end:
108 
+        jl .loop_gpr_%1%2
109 
+.end_%1%2:
73 110 
     REP_RET
111 
+%endmacro
112 
+
113 
+%if ARCH_X86_32
114 
+INIT_MMX mmx
115 
+DIFF_BYTES_PROLOGUE
116 
+    %define regsize mmsize
117 
+    DIFF_BYTES_LOOP_PREP .skip_main_aa
118 
+    DIFF_BYTES_BODY    a, a
119 
+%endif
120 
+
121 
+INIT_XMM sse2
122 
+DIFF_BYTES_PROLOGUE
123 
+    %define regsize mmsize
124 
+    DIFF_BYTES_LOOP_PREP .skip_main_aa
125 
+    test            dstq, regsize - 1
126 
+        jnz     .loop_uu
127 
+    test           src1q, regsize - 1
128 
+        jnz     .loop_ua
129 
+    DIFF_BYTES_BODY    a, a
130 
+    DIFF_BYTES_BODY    u, a
131 
+    DIFF_BYTES_BODY    u, u
132 
+
133 
+%if HAVE_AVX2_EXTERNAL
134 
+INIT_YMM avx2
135 
+DIFF_BYTES_PROLOGUE
136 
+    %define regsize mmsize
137 
+    ; Directly using unaligned SSE2 version is marginally faster than
138 
+    ; branching based on arguments.
139 
+    DIFF_BYTES_LOOP_PREP .skip_main_uu
140 
+    test            dstq, regsize - 1
141 
+        jnz     .loop_uu
142 
+    test           src1q, regsize - 1
143 
+        jnz     .loop_ua
144 
+    DIFF_BYTES_BODY    a, a
145 
+    DIFF_BYTES_BODY    u, a
146 
+    DIFF_BYTES_BODY    u, u
147 
+%endif
libavcodec/x86/huffyuvencdsp_mmx.c
History View file @ 5e586e1
... ... 
@@ -31,6 +31,10 @@
31 31
32 32 
 void ff_diff_bytes_mmx(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
33 33 
                        intptr_t w);
34 
+void ff_diff_bytes_sse2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
35 
+                        intptr_t w);
36 
+void ff_diff_bytes_avx2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
37 
+                        intptr_t w);
34 38
35 39 
 #if HAVE_INLINE_ASM
36 40
... ... 
@@ -80,7 +84,7 @@ av_cold void ff_huffyuvencdsp_init_x86(HuffYUVEncDSPContext *c)
80 80 
 {
81 81 
     av_unused int cpu_flags = av_get_cpu_flags();
82 82
83 
-    if (EXTERNAL_MMX(cpu_flags)) {
83 
+    if (ARCH_X86_32 && EXTERNAL_MMX(cpu_flags)) {
84 84 
         c->diff_bytes = ff_diff_bytes_mmx;
85 85 
     }
86 86
... ... 
@@ -89,4 +93,12 @@ av_cold void ff_huffyuvencdsp_init_x86(HuffYUVEncDSPContext *c)
89 89 
         c->sub_hfyu_median_pred = sub_hfyu_median_pred_mmxext;
90 90 
     }
91 91 
 #endif /* HAVE_INLINE_ASM */
92 
+
93 
+    if (EXTERNAL_SSE2(cpu_flags)) {
94 
+        c->diff_bytes = ff_diff_bytes_sse2;
95 
+    }
96 
+
97 
+    if (EXTERNAL_AVX2(cpu_flags)) {
98 
+        c->diff_bytes = ff_diff_bytes_avx2;
99 
+    }
92 100 
 }