sse2 version of vp8_regular_quantize_b

Added sse2 version of vp8_regular_quantize_b which improved encode
performance(for the clip used) by ~10% for 32 bit builds and ~3% for
64 bit builds.

Also updated SHADOW_ARGS_TO_STACK to allow for more than 9 arguments.

Change-Id: I62f78eabc8040b39f3ffdf21be175811e96b39af

diff --git a/vp8/encoder/quantize.h b/vp8/encoder/quantize.h
index f5ee9d7..ca073ef 100644 (file)
--- a/vp8/encoder/quantize.h
+++ b/vp8/encoder/quantize.h
@@ -17,6 +17,10 @@
 #define prototype_quantize_block(sym) \
     void (sym)(BLOCK *b,BLOCKD *d)
 
+#if ARCH_X86 || ARCH_X86_64
+#include "x86/quantize_x86.h"
+#endif
+
 #if ARCH_ARM
 #include "arm/quantize_arm.h"
 #endif

diff --git a/vp8/encoder/x86/quantize_sse2.asm b/vp8/encoder/x86/quantize_sse2.asm
new file mode 100644 (file)
index 0000000..c64a8ba
--- /dev/null
+++ b/vp8/encoder/x86/quantize_sse2.asm
@@ -0,0 +1,254 @@
+;
+;  Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license and patent
+;  grant that can be found in the LICENSE file in the root of the source
+;  tree. All contributing project authors may be found in the AUTHORS
+;  file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;int vp8_regular_quantize_b_impl_sse2(short *coeff_ptr, short *zbin_ptr,
+;               short *qcoeff_ptr,short *dequant_ptr,
+;               const int *default_zig_zag, short *round_ptr,
+;               short *quant_ptr, short *dqcoeff_ptr,
+;               unsigned short zbin_oq_value,
+;               short *zbin_boost_ptr);
+;
+global sym(vp8_regular_quantize_b_impl_sse2)
+sym(vp8_regular_quantize_b_impl_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 10
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+
+    %define abs_minus_zbin_lo 0
+    %define abs_minus_zbin_hi 16
+    %define temp_qcoeff_lo 32
+    %define temp_qcoeff_hi 48
+    %define save_xmm6 64
+    %define save_xmm7 80
+    %define eob 96
+
+    %define vp8_regularquantizeb_stack_size eob + 16
+
+    sub         rsp, vp8_regularquantizeb_stack_size
+
+    movdqa      DQWORD PTR[rsp + save_xmm6], xmm6
+    movdqa      DQWORD PTR[rsp + save_xmm7], xmm7
+
+    mov         rdx, arg(0)                 ;coeff_ptr
+    mov         eax, arg(8)                 ;zbin_oq_value
+
+    mov         rcx, arg(1)                 ;zbin_ptr
+    movd        xmm7, eax
+
+    movdqa      xmm0, DQWORD PTR[rdx]
+    movdqa      xmm4, DQWORD PTR[rdx + 16]
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm5, xmm4
+
+    psraw       xmm0, 15                    ;sign of z (aka sz)
+    psraw       xmm4, 15                    ;sign of z (aka sz)
+
+    pxor        xmm1, xmm0
+    pxor        xmm5, xmm4
+
+    movdqa      xmm2, DQWORD PTR[rcx]       ;load zbin_ptr
+    movdqa      xmm3, DQWORD PTR[rcx + 16]  ;load zbin_ptr
+
+    pshuflw     xmm7, xmm7, 0
+    psubw       xmm1, xmm0                  ;x = abs(z)
+
+    punpcklwd   xmm7, xmm7                  ;duplicated zbin_oq_value
+    psubw       xmm5, xmm4                  ;x = abs(z)
+
+    paddw       xmm2, xmm7
+    paddw       xmm3, xmm7
+
+    psubw       xmm1, xmm2                  ;sub (zbin_ptr + zbin_oq_value)
+    psubw       xmm5, xmm3                  ;sub (zbin_ptr + zbin_oq_value)
+
+    mov         rdi, arg(5)                 ;round_ptr
+    mov         rsi, arg(6)                 ;quant_ptr
+
+    movdqa      DQWORD PTR[rsp + abs_minus_zbin_lo], xmm1
+    movdqa      DQWORD PTR[rsp + abs_minus_zbin_hi], xmm5
+
+    paddw       xmm1, xmm2                  ;add (zbin_ptr + zbin_oq_value) back
+    paddw       xmm5, xmm3                  ;add (zbin_ptr + zbin_oq_value) back
+
+    movdqa      xmm2, DQWORD PTR[rdi]
+    movdqa      xmm3, DQWORD PTR[rsi]
+
+    movdqa      xmm6, DQWORD PTR[rdi + 16]
+    movdqa      xmm7, DQWORD PTR[rsi + 16]
+
+    paddw       xmm1, xmm2
+    paddw       xmm5, xmm6
+
+    pmulhw      xmm1, xmm3
+    pmulhw      xmm5, xmm7
+
+    mov         rsi, arg(2)                 ;qcoeff_ptr
+    pxor        xmm6, xmm6
+
+    pxor        xmm1, xmm0
+    pxor        xmm5, xmm4
+
+    psubw       xmm1, xmm0
+    psubw       xmm5, xmm4
+
+    movdqa      DQWORD PTR[rsp + temp_qcoeff_lo], xmm1
+    movdqa      DQWORD PTR[rsp + temp_qcoeff_hi], xmm5
+
+    movdqa      DQWORD PTR[rsi], xmm6       ;zero qcoeff
+    movdqa      DQWORD PTR[rsi + 16], xmm6  ;zero qcoeff
+
+    xor         rax, rax
+    mov         rcx, -1
+
+    mov         [rsp + eob], rcx
+    mov         rsi, arg(9)                 ;zbin_boost_ptr
+
+    mov         rbx, arg(4)                 ;default_zig_zag
+
+rq_zigzag_loop:
+    movsxd      rcx, DWORD PTR[rbx + rax*4] ;now we have rc
+    movsx       edi, WORD PTR [rsi]         ;*zbin_boost_ptr aka zbin
+    lea         rsi, [rsi + 2]              ;zbin_boost_ptr++
+
+    movsx       edx, WORD PTR[rsp + abs_minus_zbin_lo + rcx *2]
+
+    sub         edx, edi                    ;x - zbin
+    jl          rq_zigzag_1
+
+    mov         rdi, arg(2)                 ;qcoeff_ptr
+
+    movsx       edx, WORD PTR[rsp + temp_qcoeff_lo + rcx *2]
+
+    cmp         edx, 0
+    je          rq_zigzag_1
+
+    mov         WORD PTR[rdi + rcx * 2], dx ;qcoeff_ptr[rc] = temp_qcoeff[rc]
+
+    mov         rsi, arg(9)                 ;zbin_boost_ptr
+    mov         [rsp + eob], rax            ;eob = i
+
+rq_zigzag_1:
+    movsxd      rcx, DWORD PTR[rbx + rax*4 + 4]
+    movsx       edi, WORD PTR [rsi]         ;*zbin_boost_ptr aka zbin
+    lea         rsi, [rsi + 2]              ;zbin_boost_ptr++
+
+    movsx       edx, WORD PTR[rsp + abs_minus_zbin_lo + rcx *2]
+    lea         rax, [rax + 1]
+
+    sub         edx, edi                    ;x - zbin
+    jl          rq_zigzag_1a
+
+    mov         rdi, arg(2)                 ;qcoeff_ptr
+
+    movsx       edx, WORD PTR[rsp + temp_qcoeff_lo + rcx *2]
+
+    cmp         edx, 0
+    je          rq_zigzag_1a
+
+    mov         WORD PTR[rdi + rcx * 2], dx ;qcoeff_ptr[rc] = temp_qcoeff[rc]
+
+    mov         rsi, arg(9)                 ;zbin_boost_ptr
+    mov         [rsp + eob], rax            ;eob = i
+
+rq_zigzag_1a:
+    movsxd      rcx, DWORD PTR[rbx + rax*4 + 4]
+    movsx       edi, WORD PTR [rsi]         ;*zbin_boost_ptr aka zbin
+    lea         rsi, [rsi + 2]              ;zbin_boost_ptr++
+
+    movsx       edx, WORD PTR[rsp + abs_minus_zbin_lo + rcx *2]
+    lea         rax, [rax + 1]
+
+    sub         edx, edi                    ;x - zbin
+    jl          rq_zigzag_1b
+
+    mov         rdi, arg(2)                 ;qcoeff_ptr
+
+    movsx       edx, WORD PTR[rsp + temp_qcoeff_lo + rcx *2]
+
+    cmp         edx, 0
+    je          rq_zigzag_1b
+
+    mov         WORD PTR[rdi + rcx * 2], dx ;qcoeff_ptr[rc] = temp_qcoeff[rc]
+
+    mov         rsi, arg(9)                 ;zbin_boost_ptr
+    mov         [rsp + eob], rax            ;eob = i
+
+rq_zigzag_1b:
+    movsxd      rcx, DWORD PTR[rbx + rax*4 + 4]
+    movsx       edi, WORD PTR [rsi]         ;*zbin_boost_ptr aka zbin
+    lea         rsi, [rsi + 2]              ;zbin_boost_ptr++
+
+    movsx       edx, WORD PTR[rsp + abs_minus_zbin_lo + rcx *2]
+    lea         rax, [rax + 1]
+
+    sub         edx, edi                    ;x - zbin
+    jl          rq_zigzag_1c
+
+    mov         rdi, arg(2)                 ;qcoeff_ptr
+
+    movsx       edx, WORD PTR[rsp + temp_qcoeff_lo + rcx *2]
+
+    cmp         edx, 0
+    je          rq_zigzag_1c
+
+    mov         WORD PTR[rdi + rcx * 2], dx ;qcoeff_ptr[rc] = temp_qcoeff[rc]
+
+    mov         rsi, arg(9)                 ;zbin_boost_ptr
+    mov         [rsp + eob], rax            ;eob = i
+
+rq_zigzag_1c:
+    lea         rax, [rax + 1]
+
+    cmp         rax, 16
+    jl          rq_zigzag_loop
+
+    mov         rdi, arg(2)                 ;qcoeff_ptr
+    mov         rcx, arg(3)                 ;dequant_ptr
+    mov         rsi, arg(7)                 ;dqcoeff_ptr
+
+    movdqa      xmm2, DQWORD PTR[rdi]
+    movdqa      xmm3, DQWORD PTR[rdi + 16]
+
+    movdqa      xmm0, DQWORD PTR[rcx]
+    movdqa      xmm1, DQWORD PTR[rcx + 16]
+
+    pmullw      xmm0, xmm2
+    pmullw      xmm1, xmm3
+
+    movdqa      DQWORD PTR[rsi], xmm0       ;store dqcoeff
+    movdqa      DQWORD PTR[rsi + 16], xmm1  ;store dqcoeff
+
+    mov         rax, [rsp + eob]
+
+    movdqa      xmm6, DQWORD PTR[rsp + save_xmm6]
+    movdqa      xmm7, DQWORD PTR[rsp + save_xmm7]
+
+    add         rax, 1
+
+    add         rsp, vp8_regularquantizeb_stack_size
+    pop         rsp
+
+    ; begin epilog
+    pop         rbx
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret

diff --git a/vp8/encoder/x86/quantize_x86.h b/vp8/encoder/x86/quantize_x86.h
new file mode 100644 (file)
index 0000000..37d69a8
--- /dev/null
+++ b/vp8/encoder/x86/quantize_x86.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2010 The VP8 project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license and patent
+ *  grant that can be found in the LICENSE file in the root of the source
+ *  tree. All contributing project authors may be found in the AUTHORS
+ *  file in the root of the source tree.
+ */
+
+#ifndef QUANTIZE_X86_H
+#define QUANTIZE_X86_H
+
+
+/* Note:
+ *
+ * This platform is commonly built for runtime CPU detection. If you modify
+ * any of the function mappings present in this file, be sure to also update
+ * them in the function pointer initialization code
+ */
+#if HAVE_MMX
+
+#endif
+
+
+#if HAVE_SSE2
+extern prototype_quantize_block(vp8_regular_quantize_b_sse2);
+
+#if !CONFIG_RUNTIME_CPU_DETECT
+
+#undef vp8_quantize_quantb
+#define vp8_quantize_quantb vp8_regular_quantize_b_sse2
+
+#endif
+
+#endif
+
+
+#endif

diff --git a/vp8/encoder/x86/x86_csystemdependent.c b/vp8/encoder/x86/x86_csystemdependent.c
index ad10a9e..f6123a8 100644 (file)
--- a/vp8/encoder/x86/x86_csystemdependent.c
+++ b/vp8/encoder/x86/x86_csystemdependent.c
@@ -121,6 +121,40 @@ void vp8_fast_quantize_b_sse(BLOCK *b, BLOCKD *d)
              );
 }
 
+int vp8_regular_quantize_b_impl_sse2(short *coeff_ptr, short *zbin_ptr,
+                               short *qcoeff_ptr,short *dequant_ptr,
+                               const int *default_zig_zag, short *round_ptr,
+                               short *quant_ptr, short *dqcoeff_ptr,
+                               unsigned short zbin_oq_value,
+                               short *zbin_boost_ptr);
+
+void vp8_regular_quantize_b_sse2(BLOCK *b,BLOCKD *d)
+{
+    short *zbin_boost_ptr = &b->zrun_zbin_boost[0];
+    short *coeff_ptr  = &b->coeff[0];
+    short *zbin_ptr   = &b->zbin[0][0];
+    short *round_ptr  = &b->round[0][0];
+    short *quant_ptr  = &b->quant[0][0];
+    short *qcoeff_ptr = d->qcoeff;
+    short *dqcoeff_ptr = d->dqcoeff;
+    short *dequant_ptr = &d->dequant[0][0];
+    short zbin_oq_value = b->zbin_extra;
+
+    d->eob = vp8_regular_quantize_b_impl_sse2(
+        coeff_ptr,
+        zbin_ptr,
+        qcoeff_ptr,
+        dequant_ptr,
+        vp8_default_zig_zag1d,
+
+        round_ptr,
+        quant_ptr,
+        dqcoeff_ptr,
+        zbin_oq_value,
+        zbin_boost_ptr
+        );
+}
+
 int vp8_mbblock_error_xmm_impl(short *coeff_ptr, short *dcoef_ptr, int dc);
 int vp8_mbblock_error_xmm(MACROBLOCK *mb, int dc)
 {
@@ -251,6 +285,7 @@ void vp8_arch_x86_encoder_init(VP8_COMP *cpi)
         /* cpi->rtcd.encodemb.sub* not implemented for wmt */
 
         cpi->rtcd.quantize.fastquantb            = vp8_fast_quantize_b_sse;
+        cpi->rtcd.quantize.quantb            = vp8_regular_quantize_b_sse2;
     }
 
 #endif

diff --git a/vp8/vp8cx.mk b/vp8/vp8cx.mk
index 9496ef0..971a175 100644 (file)
--- a/vp8/vp8cx.mk
+++ b/vp8/vp8cx.mk
@@ -98,6 +98,7 @@ VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/variance_impl_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/sad_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/fwalsh_sse2.asm
+VP8_CX_SRCS-$(HAVE_SSE2) += encoder/x86/quantize_sse2.asm
 VP8_CX_SRCS-$(HAVE_SSE3) += encoder/x86/sad_sse3.asm
 VP8_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/sad_ssse3.asm
 VP8_CX_SRCS-$(ARCH_X86)$(ARCH_X86_64) += encoder/x86/quantize_mmx.asm

diff --git a/vpx_ports/x86_abi_support.asm b/vpx_ports/x86_abi_support.asm
index 7840e35..a1622e6 100644 (file)
--- a/vpx_ports/x86_abi_support.asm
+++ b/vpx_ports/x86_abi_support.asm
@@ -199,16 +199,13 @@
         push r9
     %endif
     %if %1 > 6
-        mov rax,[rbp+16]
-        push rax
-    %endif
-    %if %1 > 7
-        mov rax,[rbp+24]
-        push rax
-    %endif
-    %if %1 > 8
-        mov rax,[rbp+32]
+      %assign i %1-6
+      %assign off 16
+      %rep i
+        mov rax,[rbp+off]
         push rax
+        %assign off off+8
+      %endrep
     %endif
   %endm
 %endif