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.

libavcodec/huffyuvenc.c

@@ -60,12 +60,12 @@ static inline int sub_left_prediction(HYuvContext *s, uint8_t *dst,
             }
             return left;
         } else {
-            for (i = 0; i < 16; i++) {
+            for (i = 0; i < 32; i++) {
                 const int temp = src[i];
                 dst[i] = temp - left;
                 left   = temp;
             }
-            s->hencdsp.diff_bytes(dst + 16, src + 16, src + 15, w - 16);
+            s->hencdsp.diff_bytes(dst + 32, src + 32, src + 31, w - 32);
             return src[w-1];
         }
     } else {

libavcodec/x86/huffyuvencdsp.asm

@@ -27,9 +27,9 @@
 
 section .text
 
-INIT_MMX mmx
 ; void ff_diff_bytes_mmx(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
 ;                        intptr_t w);
+%macro DIFF_BYTES_PROLOGUE 0
 %if ARCH_X86_32
 cglobal diff_bytes, 3,5,2, dst, src1, src2
 %define wq r4q
@@ -40,34 +40,108 @@ cglobal diff_bytes, 4,5,2, dst, src1, src2, w
     DECLARE_REG_TMP 4
 %endif ; ARCH_X86_32
 %define i t0q
+%endmacro
+
+; label to jump to if w < regsize
+%macro DIFF_BYTES_LOOP_PREP 1
     mov                i, wq
-    and                i, -2 * mmsize
-        jz  .setup_loop2
+    and                i, -2 * regsize
+        jz            %1
     add             dstq, i
     add            src1q, i
     add            src2q, i
     neg                i
-.loop:
-    mova              m0, [src1q + i]
-    mova              m1, [src1q + i + mmsize]
-    psubb             m0, [src2q + i]
-    psubb             m1, [src2q + i + mmsize]
-    mova          [dstq + i], m0
-    mova [mmsize + dstq + i], m1
-    add                i, 2 * mmsize
-        jl         .loop
-.setup_loop2:
-    and               wq, 2 * mmsize - 1
-        jz          .end
+%endmacro
+
+; mov type used for src1q, dstq, first reg, second reg
+%macro DIFF_BYTES_LOOP_CORE 4
+%if regsize != 16
+    mov%1             %3, [src1q + i]
+    mov%1             %4, [src1q + i + regsize]
+    psubb             %3, [src2q + i]
+    psubb             %4, [src2q + i + regsize]
+    mov%2           [dstq + i], %3
+    mov%2 [regsize + dstq + i], %4
+%else
+    ; SSE enforces alignment of psubb operand
+    mov%1             %3, [src1q + i]
+    movu              %4, [src2q + i]
+    psubb             %3, %4
+    mov%2     [dstq + i], %3
+    mov%1             %3, [src1q + i + regsize]
+    movu              %4, [src2q + i + regsize]
+    psubb             %3, %4
+    mov%2 [regsize + dstq + i], %3
+%endif
+%endmacro
+
+%macro DIFF_BYTES_BODY 2 ; mov type used for src1q, for dstq
+    %define regsize mmsize
+.loop_%1%2:
+    DIFF_BYTES_LOOP_CORE %1, %2, m0, m1
+    add                i, 2 * regsize
+        jl    .loop_%1%2
+.skip_main_%1%2:
+    and               wq, 2 * regsize - 1
+        jz     .end_%1%2
+%if mmsize > 16
+    ; fall back to narrower xmm
+    %define regsize mmsize / 2
+    DIFF_BYTES_LOOP_PREP .setup_loop_gpr_aa
+.loop2_%1%2:
+    DIFF_BYTES_LOOP_CORE %1, %2, xm0, xm1
+    add                i, 2 * regsize
+        jl   .loop2_%1%2
+.setup_loop_gpr_%1%2:
+    and               wq, 2 * regsize - 1
+        jz     .end_%1%2
+%endif
     add             dstq, wq
     add            src1q, wq
     add            src2q, wq
     neg               wq
-.loop2:
+.loop_gpr_%1%2:
     mov              t0b, [src1q + wq]
     sub              t0b, [src2q + wq]
     mov      [dstq + wq], t0b
     inc               wq
-        jl        .loop2
-.end:
+        jl .loop_gpr_%1%2
+.end_%1%2:
     REP_RET
+%endmacro
+
+%if ARCH_X86_32
+INIT_MMX mmx
+DIFF_BYTES_PROLOGUE
+    %define regsize mmsize
+    DIFF_BYTES_LOOP_PREP .skip_main_aa
+    DIFF_BYTES_BODY    a, a
+%endif
+
+INIT_XMM sse2
+DIFF_BYTES_PROLOGUE
+    %define regsize mmsize
+    DIFF_BYTES_LOOP_PREP .skip_main_aa
+    test            dstq, regsize - 1
+        jnz     .loop_uu
+    test           src1q, regsize - 1
+        jnz     .loop_ua
+    DIFF_BYTES_BODY    a, a
+    DIFF_BYTES_BODY    u, a
+    DIFF_BYTES_BODY    u, u
+
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+DIFF_BYTES_PROLOGUE
+    %define regsize mmsize
+    ; Directly using unaligned SSE2 version is marginally faster than
+    ; branching based on arguments.
+    DIFF_BYTES_LOOP_PREP .skip_main_uu
+    test            dstq, regsize - 1
+        jnz     .loop_uu
+    test           src1q, regsize - 1
+        jnz     .loop_ua
+    DIFF_BYTES_BODY    a, a
+    DIFF_BYTES_BODY    u, a
+    DIFF_BYTES_BODY    u, u
+%endif

libavcodec/x86/huffyuvencdsp_mmx.c

@@ -31,6 +31,10 @@
 
 void ff_diff_bytes_mmx(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
                        intptr_t w);
+void ff_diff_bytes_sse2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
+                        intptr_t w);
+void ff_diff_bytes_avx2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2,
+                        intptr_t w);
 
 #if HAVE_INLINE_ASM
 
@@ -80,7 +84,7 @@ av_cold void ff_huffyuvencdsp_init_x86(HuffYUVEncDSPContext *c)
 {
     av_unused int cpu_flags = av_get_cpu_flags();
 
-    if (EXTERNAL_MMX(cpu_flags)) {
+    if (ARCH_X86_32 && EXTERNAL_MMX(cpu_flags)) {
         c->diff_bytes = ff_diff_bytes_mmx;
     }
 
@@ -89,4 +93,12 @@ av_cold void ff_huffyuvencdsp_init_x86(HuffYUVEncDSPContext *c)
         c->sub_hfyu_median_pred = sub_hfyu_median_pred_mmxext;
     }
 #endif /* HAVE_INLINE_ASM */
+
+    if (EXTERNAL_SSE2(cpu_flags)) {
+        c->diff_bytes = ff_diff_bytes_sse2;
+    }
+
+    if (EXTERNAL_AVX2(cpu_flags)) {
+        c->diff_bytes = ff_diff_bytes_avx2;
+    }
 }