[x265] [PATCH] count_nonzero primitive, downscaling quantCoeff from int32_t* to int16_t*
Steve Borho
steve at borho.org
Wed Aug 13 22:31:29 CEST 2014
On 08/12, praveen at multicorewareinc.com wrote:
> # HG changeset patch
> # User Praveen Tiwari
> # Date 1407834530 -19800
> # Node ID bb4d44663964237e4b66af6d92b2f13dbcf4f9b9
> # Parent 8a7f4bb1d1be32fe668d410450c2e320ccae6098
> count_nonzero primitive, downscaling quantCoeff from int32_t* to int16_t*
There's not much point in applying these patches until all of the quant
primitives are using short ints for coefficients. As-is this will just
be a slow-down.
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/common/dct.cpp
> --- a/source/common/dct.cpp Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/common/dct.cpp Tue Aug 12 14:38:50 2014 +0530
> @@ -815,7 +815,7 @@
> return numSig;
> }
>
> -int count_nonzero_c(const int32_t *quantCoeff, int numCoeff)
> +int count_nonzero_c(const int16_t *quantCoeff, int numCoeff)
> {
> X265_CHECK(((intptr_t)quantCoeff & 15) == 0, "quant buffer not aligned\n");
> X265_CHECK(numCoeff > 0 && (numCoeff & 15) == 0, "numCoeff invalid %d\n", numCoeff);
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/common/primitives.h
> --- a/source/common/primitives.h Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/common/primitives.h Tue Aug 12 14:38:50 2014 +0530
> @@ -163,7 +163,7 @@
> typedef uint32_t (*nquant_t)(int32_t *coef, int32_t *quantCoeff, int32_t *qCoef, int qBits, int add, int numCoeff);
> typedef void (*dequant_scaling_t)(const int32_t* src, const int32_t *dequantCoef, int32_t* dst, int num, int mcqp_miper, int shift);
> typedef void (*dequant_normal_t)(const int32_t* quantCoef, int32_t* coef, int num, int scale, int shift);
> -typedef int (*count_nonzero_t)(const int32_t *quantCoeff, int numCoeff);
> +typedef int (*count_nonzero_t)(const int16_t *quantCoeff, int numCoeff);
>
> typedef void (*weightp_pp_t)(pixel *src, pixel *dst, intptr_t srcStride, intptr_t dstStride, int width, int height, int w0, int round, int shift, int offset);
> typedef void (*weightp_sp_t)(int16_t *src, pixel *dst, intptr_t srcStride, intptr_t dstStride, int width, int height, int w0, int round, int shift, int offset);
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/common/quant.cpp
> --- a/source/common/quant.cpp Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/common/quant.cpp Tue Aug 12 14:38:50 2014 +0530
> @@ -2,6 +2,7 @@
> * Copyright (C) 2014 x265 project
> *
> * Authors: Steve Borho <steve at borho.org>
> + * Praveen Kumar Tiwari <praveen at multicorewareinc.com>
> *
> * This program is free software; you can redistribute it and/or modify
> * it under the terms of the GNU General Public License as published by
> @@ -463,7 +464,17 @@
> const uint32_t sizeIdx = log2TrSize - 2;
> int useDST = !sizeIdx && ttype == TEXT_LUMA && bIntra;
>
> - X265_CHECK((int)numSig == primitives.count_nonzero(coeff, 1 << log2TrSize * 2), "numSig differ\n");
> + /* This section of code is to safely convert int32_t coefficients to int16_t, once the caller function is
> + * optimize to take coefficients as int16_t*, it will be cleanse.*/
> + int numCoeff = (1 << (log2TrSize * 2));
> + assert(numCoeff <= 1024);
> + ALIGN_VAR_16(int16_t, qCoeff[32 * 32]);
> + for (int i = 0; i < numCoeff; i++)
> + {
> + qCoeff[i] = (coeff[i] & 0xFFFF);
> + }
> +
> + X265_CHECK((int)numSig == primitives.count_nonzero(qCoeff, 1 << log2TrSize * 2), "numSig differ\n");
>
> // DC only
> if (numSig == 1 && coeff[0] != 0 && !useDST)
> @@ -501,7 +512,16 @@
> int numCoeff = 1 << log2TrSize * 2;
> uint32_t numSig = primitives.nquant(m_resiDctCoeff, qCoef, dstCoeff, qbits, add, numCoeff);
These two loops are only here for an X265_CHECK statement that is
usually compiled out. All of this code should have been wrapped within
#if CHECKED_BUILD || _DEBUG
> - X265_CHECK((int)numSig == primitives.count_nonzero(dstCoeff, numCoeff), "numSig differ\n");
> + /* This section of code is to safely convert int32_t coefficients to int16_t, once the caller function is
> + * optimize to take coefficients as int16_t*, it will be cleanse.*/
> + assert(numCoeff <= 1024);
> + ALIGN_VAR_16(int16_t, qCoeff[32 * 32]);
> + for (int i = 0; i < numCoeff; i++)
> + {
> + qCoeff[i] = (dstCoeff[i] & 0xFFFF);
> + }
> +
> + X265_CHECK((int)numSig == primitives.count_nonzero(qCoeff, numCoeff), "numSig differ\n");
> if (!numSig)
> return 0;
>
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/common/x86/pixel-util.h
> --- a/source/common/x86/pixel-util.h Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/common/x86/pixel-util.h Tue Aug 12 14:38:50 2014 +0530
> @@ -2,6 +2,7 @@
> * Copyright (C) 2013 x265 project
> *
> * Authors: Steve Borho <steve at borho.org>
> + * Praveen Kumar Tiwari <praveen at multicorewareinc.com>
> *
> * This program is free software; you can redistribute it and/or modify
> * it under the terms of the GNU General Public License as published by
> @@ -47,7 +48,7 @@
> uint32_t x265_quant_sse4(int32_t *coef, int32_t *quantCoeff, int32_t *deltaU, int32_t *qCoef, int qBits, int add, int numCoeff);
> uint32_t x265_nquant_sse4(int32_t *coef, int32_t *quantCoeff, int32_t *qCoef, int qBits, int add, int numCoeff);
> void x265_dequant_normal_sse4(const int32_t* quantCoef, int32_t* coef, int num, int scale, int shift);
> -int x265_count_nonzero_ssse3(const int32_t *quantCoeff, int numCoeff);
> +int x265_count_nonzero_ssse3(const int16_t *quantCoeff, int numCoeff);
>
> void x265_weight_pp_sse4(pixel *src, pixel *dst, intptr_t srcStride, intptr_t dstStride, int width, int height, int w0, int round, int shift, int offset);
> void x265_weight_sp_sse4(int16_t *src, pixel *dst, intptr_t srcStride, intptr_t dstStride, int width, int height, int w0, int round, int shift, int offset);
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/common/x86/pixel-util8.asm
> --- a/source/common/x86/pixel-util8.asm Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/common/x86/pixel-util8.asm Tue Aug 12 14:38:50 2014 +0530
> @@ -3,6 +3,7 @@
> ;*
> ;* Authors: Min Chen <chenm003 at 163.com> <min.chen at multicorewareinc.com>
> ;* Nabajit Deka <nabajit at multicorewareinc.com>
> +;* Praveen Kumar Tiwari <praveen at multicorewareinc.com>
> ;*
> ;* This program is free software; you can redistribute it and/or modify
> ;* it under the terms of the GNU General Public License as published by
> @@ -1091,10 +1092,10 @@
>
>
> ;-----------------------------------------------------------------------------
> -; int count_nonzero(const int32_t *quantCoeff, int numCoeff);
> +; int count_nonzero(const int16_t *quantCoeff, int numCoeff);
> ;-----------------------------------------------------------------------------
> INIT_XMM ssse3
> -cglobal count_nonzero, 2,2,5
> +cglobal count_nonzero, 2,2,4
> pxor m0, m0
> shr r1d, 4
> movd m1, r1d
> @@ -1103,12 +1104,8 @@
> .loop:
> mova m2, [r0 + 0]
> mova m3, [r0 + 16]
> - packssdw m2, m3
> - mova m3, [r0 + 32]
> - mova m4, [r0 + 48]
> - add r0, 64
> - packssdw m3, m4
> packsswb m2, m3
> + add r0, 32
> pcmpeqb m2, m0
> paddb m1, m2
> dec r1d
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/encoder/entropy.cpp
> --- a/source/encoder/entropy.cpp Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/encoder/entropy.cpp Tue Aug 12 14:38:50 2014 +0530
> @@ -2,6 +2,7 @@
> * Copyright (C) 2013 x265 project
> *
> * Authors: Steve Borho <steve at borho.org>
> +* Praveen Kumar Tiwari <praveen at multicorewareinc.com>
> *
> * This program is free software; you can redistribute it and/or modify
> * it under the terms of the GNU General Public License as published by
> @@ -1488,8 +1489,18 @@
> {
> uint32_t trSize = 1 << log2TrSize;
>
> + /* This section of code is to safely convert int32_t coefficients to int16_t, once the caller function is
> + * optimize to take coefficients as int16_t*, it will be cleanse.*/
> + int numCoeff = (1 << (log2TrSize << 1));
> + assert(numCoeff <= 1024);
> + ALIGN_VAR_16(int16_t, qCoeff[32 * 32]);
> + for (int i = 0; i < numCoeff; i++)
> + {
> + qCoeff[i] = (coeff[i] & 0xFFFF);
> + }
> +
> // compute number of significant coefficients
> - uint32_t numSig = primitives.count_nonzero(coeff, (1 << (log2TrSize << 1)));
> + uint32_t numSig = primitives.count_nonzero(qCoeff, (1 << (log2TrSize << 1)));
>
> X265_CHECK(numSig > 0, "cbf check fail\n");
>
> diff -r 8a7f4bb1d1be -r bb4d44663964 source/test/mbdstharness.cpp
> --- a/source/test/mbdstharness.cpp Tue Aug 12 01:11:39 2014 -0500
> +++ b/source/test/mbdstharness.cpp Tue Aug 12 14:38:50 2014 +0530
> @@ -366,7 +366,7 @@
>
> bool MBDstHarness::check_count_nonzero_primitive(count_nonzero_t ref, count_nonzero_t opt)
> {
> - ALIGN_VAR_32(int32_t, qcoeff[32 * 32]);
> + ALIGN_VAR_32(int16_t, qcoeff[32 * 32]);
>
> for (int i = 0; i < 4; i++)
> {
> @@ -376,7 +376,7 @@
>
> for (int n = 0; n <= num; n++)
> {
> - memset(qcoeff, 0, num * sizeof(int32_t));
> + memset(qcoeff, 0, num * sizeof(int16_t));
>
> for (int j = 0; j < n; j++)
> {
> @@ -386,7 +386,7 @@
> k = (k + 11) & mask;
> }
>
> - qcoeff[k] = rand() - RAND_MAX / 2;
> + qcoeff[k] = (int16_t)rand() - RAND_MAX / 2;
> }
>
> int refval = ref(qcoeff, num);
> @@ -516,7 +516,7 @@
> for (int i = 4; i <= 32; i <<= 1)
> {
> printf("count_nonzero[%dx%d]", i, i);
> - REPORT_SPEEDUP(opt.count_nonzero, ref.count_nonzero, mbufidct, i * i)
> + REPORT_SPEEDUP(opt.count_nonzero, ref.count_nonzero, mbuf1, i * i)
> }
> }
> }
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
--
Steve Borho
More information about the x265-devel
mailing list