From 33fa7c4ebe34d25b20d82e318b05cec8b0e32c9c Mon Sep 17 00:00:00 2001 From: Tero Rintaluoma Date: Mon, 11 Apr 2011 12:04:17 +0300 Subject: [PATCH] neon fast quantizer updated vp8_fast_quantize_b_neon function updated and further optimized. - match current C implementation of fast quantizer - updated to use asm_enc_offsets for structure members - updated ads2gas scripts to handle alignment issues Change-Id: I5cbad9c460ad8ddb35d2970a8684cc620711c56d --- build/make/ads2gas.pl | 7 +- build/make/ads2gas_apple.pl | 7 +- vp8/encoder/arm/arm_csystemdependent.c | 8 +- vp8/encoder/arm/neon/fastquantizeb_neon.asm | 136 +++++++++++--------- vp8/encoder/arm/quantize_arm.c | 79 ------------ vp8/encoder/arm/quantize_arm.h | 11 +- vp8/encoder/quantize.h | 2 + vp8/vp8cx_arm.mk | 1 - 8 files changed, 93 insertions(+), 158 deletions(-) delete mode 100644 vp8/encoder/arm/quantize_arm.c diff --git a/build/make/ads2gas.pl b/build/make/ads2gas.pl index 3dff048b5..efdfce7a1 100755 --- a/build/make/ads2gas.pl +++ b/build/make/ads2gas.pl @@ -23,6 +23,9 @@ print "\t.equ DO1STROUNDING, 0\n"; while () { + # Load and store alignment + s/@/,:/g; + # Comment character s/;/@/g; @@ -114,8 +117,8 @@ while () # put the colon at the end of the line in the macro s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/; - # Strip ALIGN - s/\sALIGN/@ ALIGN/g; + # ALIGN directive + s/ALIGN/.balign/g; # Strip ARM s/\sARM/@ ARM/g; diff --git a/build/make/ads2gas_apple.pl b/build/make/ads2gas_apple.pl index 5014c61fb..1b3039374 100755 --- a/build/make/ads2gas_apple.pl +++ b/build/make/ads2gas_apple.pl @@ -41,6 +41,9 @@ sub trim($) while () { + # Load and store alignment + s/@/,:/g; + # Comment character s/;/@/g; @@ -137,8 +140,8 @@ while () # put the colon at the end of the line in the macro s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/; - # Strip ALIGN - s/\sALIGN/@ ALIGN/g; + # ALIGN directive + s/ALIGN/.balign/g; # Strip ARM s/\sARM/@ ARM/g; diff --git a/vp8/encoder/arm/arm_csystemdependent.c b/vp8/encoder/arm/arm_csystemdependent.c index f148b8997..af2a5df98 100644 --- a/vp8/encoder/arm/arm_csystemdependent.c +++ b/vp8/encoder/arm/arm_csystemdependent.c @@ -121,12 +121,8 @@ void vp8_arch_arm_encoder_init(VP8_COMP *cpi) cpi->rtcd.encodemb.submby = vp8_subtract_mby_neon; cpi->rtcd.encodemb.submbuv = vp8_subtract_mbuv_neon; - /*cpi->rtcd.quantize.quantb = vp8_regular_quantize_b; - cpi->rtcd.quantize.fastquantb = vp8_fast_quantize_b_c;*/ - /* The neon quantizer has not been updated to match the new exact - * quantizer introduced in commit e04e2935 - */ - /*cpi->rtcd.quantize.fastquantb = vp8_fast_quantize_b_neon;*/ + /*cpi->rtcd.quantize.quantb = vp8_regular_quantize_b;*/ + cpi->rtcd.quantize.fastquantb = vp8_fast_quantize_b_neon; } #endif diff --git a/vp8/encoder/arm/neon/fastquantizeb_neon.asm b/vp8/encoder/arm/neon/fastquantizeb_neon.asm index ca1ea9c18..3dd92b12e 100644 --- a/vp8/encoder/arm/neon/fastquantizeb_neon.asm +++ b/vp8/encoder/arm/neon/fastquantizeb_neon.asm @@ -1,5 +1,5 @@ ; -; Copyright (c) 2010 The WebM project authors. All Rights Reserved. +; Copyright (c) 2011 The WebM project authors. All Rights Reserved. ; ; Use of this source code is governed by a BSD-style license ; that can be found in the LICENSE file in the root of the source @@ -9,110 +9,122 @@ ; - EXPORT |vp8_fast_quantize_b_neon_func| + EXPORT |vp8_fast_quantize_b_neon| + + INCLUDE asm_enc_offsets.asm ARM REQUIRE8 PRESERVE8 - AREA ||.text||, CODE, READONLY, ALIGN=2 + AREA ||.text||, CODE, READONLY, ALIGN=4 -; r0 short *coeff_ptr -; r1 short *zbin_ptr -; r2 short *qcoeff_ptr -; r3 short *dqcoeff_ptr -; stack short *dequant_ptr -; stack short *scan_mask -; stack short *round_ptr -; stack short *quant_ptr -; return int * eob -|vp8_fast_quantize_b_neon_func| PROC - vld1.16 {q0, q1}, [r0] ;load z - vld1.16 {q10, q11}, [r1] ;load zbin +;void vp8_fast_quantize_b_c(BLOCK *b, BLOCKD *d) +|vp8_fast_quantize_b_neon| PROC - vabs.s16 q4, q0 ;calculate x = abs(z) - vabs.s16 q5, q1 + stmfd sp!, {r4-r7} - vcge.s16 q10, q4, q10 ;x>=zbin - vcge.s16 q11, q5, q11 + ldr r3, [r0, #vp8_block_coeff] + ldr r4, [r0, #vp8_block_quant_fast] + ldr r5, [r0, #vp8_block_round] - ;if x> 16 - vqdmulh.s16 q5, q9 + vqdmulh.s16 q12, q8 ; y = ((Round+abs(z)) * Quant) >> 16 + vqdmulh.s16 q13, q9 - vld1.16 {q0, q1}, [r0] ;load rvsplus1_scan_order - vceq.s16 q8, q8 ;set q8 to all 1 + vld1.16 {q10, q11}, [r0@128]; load inverse scan order - vshr.s16 q4, #1 ;right shift 1 after vqdmulh - vshr.s16 q5, #1 + vceq.s16 q8, q8 ; set q8 to all 1 - ;modify data to have its original sign - veor.s16 q4, q2 ; y^sz - veor.s16 q5, q3 + ldr r4, [r1, #vp8_blockd_dequant] + + vshr.s16 q12, #1 ; right shift 1 after vqdmulh + vshr.s16 q13, #1 - ldr r12, [sp] ;load dequant_ptr + orr r2, r2, r3 ; check if all zero (step 4) + cmp r2, #0 ; check if all zero (step 5) + beq zero_output ; check if all zero (step 6) + + ;modify data to have its original sign + veor.s16 q12, q2 ; y^sz + veor.s16 q13, q3 - vsub.s16 q4, q2 ; x1 = (y^sz) - sz = (y^sz) - (-1) (two's complement) - vsub.s16 q5, q3 + vsub.s16 q12, q2 ; x1=(y^sz)-sz = (y^sz)-(-1) (2's complement) + vsub.s16 q13, q3 - vand.s16 q4, q10 ;mask off x1 elements - vand.s16 q5, q11 + vld1.s16 {q2, q3}, [r4@128] ; load dequant_ptr[i] - vld1.s16 {q6, q7}, [r12] ;load dequant_ptr[i] + vtst.16 q14, q12, q8 ; now find eob + vtst.16 q15, q13, q8 ; non-zero element is set to all 1 - vtst.16 q14, q4, q8 ;now find eob - vtst.16 q15, q5, q8 ;non-zero element is set to all 1 in q4, q5 + vst1.s16 {q12, q13}, [r6@128]; store: qcoeff = x1 - vst1.s16 {q4, q5}, [r2] ;store: qcoeff = x1 + vand q10, q10, q14 ; get all valid numbers from scan array + vand q11, q11, q15 - vand q0, q0, q14 ;get all valid number from rvsplus1_scan_order array - vand q1, q1, q15 - vmax.u16 q0, q0, q1 ;find maximum value in q0, q1 + vmax.u16 q0, q10, q11 ; find maximum value in q0, q1 vmax.u16 d0, d0, d1 vmovl.u16 q0, d0 - vmul.s16 q6, q4 ;x * Dequant - vmul.s16 q7, q5 + vmul.s16 q2, q12 ; x * Dequant + vmul.s16 q3, q13 vmax.u32 d0, d0, d1 vpmax.u32 d0, d0, d0 - vst1.s16 {q6, q7}, [r3] ;store dqcoeff = x * Dequant + vst1.s16 {q2, q3}, [r7@128] ; store dqcoeff = x * Dequant - vmov.32 r0, d0[0] + vmov.32 r0, d0[0] ; this instruction takes 1+13 cycles + ; if we have vfp, we could use + ; vstr s0, [r1, #vp8_blockd_eob] + str r0, [r1, #vp8_blockd_eob] + + ldmfd sp!, {r4-r7} bx lr zero_output - vst1.s16 {q10, q11}, [r2] ; qcoeff = 0 - vst1.s16 {q10, q11}, [r3] ; dqcoeff = 0 - mov r0, #0 + str r2, [r1, #vp8_blockd_eob] + vst1.s16 {q0, q1}, [r6@128] ; qcoeff = 0 + vst1.s16 {q0, q1}, [r7@128] ; dqcoeff = 0 + ldmfd sp!, {r4-r7} bx lr ENDP +; default inverse zigzag table is defined in vp8/common/entropy.c +_inv_zig_zag_ + DCD inv_zig_zag + + ALIGN 16 ; enable use of @128 bit aligned loads +inv_zig_zag + DCW 0x0001, 0x0002, 0x0006, 0x0007 + DCW 0x0003, 0x0005, 0x0008, 0x000d + DCW 0x0004, 0x0009, 0x000c, 0x000e + DCW 0x000a, 0x000b, 0x000f, 0x0010 + END + diff --git a/vp8/encoder/arm/quantize_arm.c b/vp8/encoder/arm/quantize_arm.c deleted file mode 100644 index 0e3334ac7..000000000 --- a/vp8/encoder/arm/quantize_arm.c +++ /dev/null @@ -1,79 +0,0 @@ -/* - * Copyright (c) 2010 The WebM project authors. All Rights Reserved. - * - * Use of this source code is governed by a BSD-style license - * that can be found in the LICENSE file in the root of the source - * tree. An additional intellectual property rights grant can be found - * in the file PATENTS. All contributing project authors may - * be found in the AUTHORS file in the root of the source tree. - */ - - -#include -#include "vpx_mem/vpx_mem.h" - -#include "vp8/encoder/quantize.h" -#include "vp8/common/entropy.h" - -DECLARE_ALIGNED(16, const short, vp8_rvsplus1_default_zig_zag1d[16]) = -{ - 1, 2, 6, 7, - 3, 5, 8, 13, - 4, 9, 12, 14, - 10, 11, 15, 16, -}; - - -extern int vp8_fast_quantize_b_neon_func(short *coeff_ptr, short *zbin_ptr, short *qcoeff_ptr, short *dqcoeff_ptr, short *dequant_ptr, const short *scan_mask, short *round_ptr, short *quant_ptr); - -void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d) -{ - d->eob = vp8_fast_quantize_b_neon_func(b->coeff, b->zbin, d->qcoeff, d->dqcoeff, d->dequant, vp8_rvsplus1_default_zig_zag1d, b->round, b->quant_fast); -} - -/* -//neon code is written according to the following rewritten c code -void vp8_fast_quantize_b_neon(BLOCK *b,BLOCKD *d) -{ - int i, rc, eob; - int zbin; - int x, x1, y, z, sz; - 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]; - - eob = 0; - - for(i=0;i<16;i++) - { - z = coeff_ptr[i]; - zbin = zbin_ptr[i] ; - x = abs(z); // x = abs(z) - - if(x>=zbin) - { - sz = (z>>31); // sign of z - y = ((x+round_ptr[i])*quant_ptr[i])>>16; // quantize (x) - x1 = (y^sz) - sz; // get the sign back - - qcoeff_ptr[i] = x1; // write to destination - dqcoeff_ptr[i] = x1 * dequant_ptr[i]; // dequantized value - - if(y) - { - if(eobeob = eob; -} -*/ diff --git a/vp8/encoder/arm/quantize_arm.h b/vp8/encoder/arm/quantize_arm.h index 0c6adf4c2..af4187ac1 100644 --- a/vp8/encoder/arm/quantize_arm.h +++ b/vp8/encoder/arm/quantize_arm.h @@ -23,14 +23,13 @@ extern prototype_quantize_block(vp8_fast_quantize_b_armv6); #if HAVE_ARMV7 + extern prototype_quantize_block(vp8_fast_quantize_b_neon); -/* The neon quantizer has not been updated to match the new exact - * quantizer introduced in commit e04e2935 - */ -//#undef vp8_quantize_fastquantb -//#define vp8_quantize_fastquantb vp8_fast_quantize_b_neon +#undef vp8_quantize_fastquantb +#define vp8_quantize_fastquantb vp8_fast_quantize_b_neon -#endif +#endif /* HAVE_ARMV7 */ #endif + diff --git a/vp8/encoder/quantize.h b/vp8/encoder/quantize.h index b74718bfa..e4c32a570 100644 --- a/vp8/encoder/quantize.h +++ b/vp8/encoder/quantize.h @@ -17,6 +17,7 @@ #define prototype_quantize_block(sym) \ void (sym)(BLOCK *b,BLOCKD *d) + #if ARCH_X86 || ARCH_X86_64 #include "x86/quantize_x86.h" #endif @@ -41,6 +42,7 @@ typedef struct prototype_quantize_block(*fastquantb); } vp8_quantize_rtcd_vtable_t; + #if CONFIG_RUNTIME_CPU_DETECT #define QUANTIZE_INVOKE(ctx,fn) (ctx)->fn #else diff --git a/vp8/vp8cx_arm.mk b/vp8/vp8cx_arm.mk index 165dada2b..03d42d215 100644 --- a/vp8/vp8cx_arm.mk +++ b/vp8/vp8cx_arm.mk @@ -15,7 +15,6 @@ # encoder VP8_CX_SRCS-$(ARCH_ARM) += encoder/arm/arm_csystemdependent.c -VP8_CX_SRCS-$(HAVE_ARMV7) += encoder/arm/quantize_arm.c VP8_CX_SRCS-$(HAVE_ARMV7) += encoder/arm/picklpf_arm.c VP8_CX_SRCS-$(HAVE_ARMV6) += encoder/arm/dct_arm.c VP8_CX_SRCS-$(HAVE_ARMV6) += encoder/arm/variance_arm.c -- 2.40.0