[x265] [PATCH 307 of 307] x86:AVX512 Set run time flag to enable/disable avx512
Mateusz
mateuszb at poczta.onet.pl
Fri Apr 13 06:35:53 CEST 2018
W dniu 07.04.2018 o 04:35, mythreyi at multicorewareinc.com pisze:
> # HG changeset patch
> # User Jayashree <jayashree.c at multicorewareinc.com>
> # Date 1522928767 -19800
> # Thu Apr 05 17:16:07 2018 +0530
> # Node ID f6ad2fa637fd3c8f9e2811982b89aa28228e9f6b
> # Parent 876b6e006f2080072c0684dbf75e7cfde974ba79
> x86:AVX512 Set run time flag to enable/disable avx512
>
> diff -r 876b6e006f20 -r f6ad2fa637fd source/common/cpu.cpp
> --- a/source/common/cpu.cpp Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/common/cpu.cpp Thu Apr 05 17:16:07 2018 +0530
> @@ -122,7 +122,7 @@
> #pragma warning(disable: 4309) // truncation of constant value
> #endif
>
> -uint32_t cpu_detect(void)
> +uint32_t cpu_detect(bool benableavx512 )
> {
> uint32_t cpu = 0;
>
> @@ -184,11 +184,13 @@
> {
> if (ebx & 0x00000020)
> cpu |= X265_CPU_AVX2;
> -
> - if ((xcr0 & 0xE0) == 0xE0) /* OPMASK/ZMM state */
> + if (benableavx512)
> {
> - if ((ebx & 0xD0030000) == 0xD0030000)
> - cpu |= X265_CPU_AVX512;
> + if ((xcr0 & 0xE0) == 0xE0) /* OPMASK/ZMM state */
> + {
> + if ((ebx & 0xD0030000) == 0xD0030000)
> + cpu |= X265_CPU_AVX512;
> + }
> }
> }
> }
> @@ -327,7 +329,7 @@
> int PFX(cpu_fast_neon_mrc_test)(void);
> }
>
> -uint32_t cpu_detect(void)
> +uint32_t cpu_detect(bool benableavx512)
> {
> int flags = 0;
>
> @@ -370,7 +372,7 @@
>
> #elif X265_ARCH_POWER8
>
> -uint32_t cpu_detect(void)
> +uint32_t cpu_detect(bool benableavx512)
> {
> #if HAVE_ALTIVEC
> return X265_CPU_ALTIVEC;
> @@ -381,7 +383,7 @@
>
> #else // if X265_ARCH_POWER8
>
> -uint32_t cpu_detect(void)
> +uint32_t cpu_detect(bool benableavx512)
> {
> return 0;
> }
> diff -r 876b6e006f20 -r f6ad2fa637fd source/common/cpu.h
> --- a/source/common/cpu.h Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/common/cpu.h Thu Apr 05 17:16:07 2018 +0530
> @@ -50,7 +50,7 @@
> #endif
>
> namespace X265_NS {
> -uint32_t cpu_detect(void);
> +uint32_t cpu_detect(bool);
>
> struct cpu_name_t
> {
> diff -r 876b6e006f20 -r f6ad2fa637fd source/common/param.cpp
> --- a/source/common/param.cpp Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/common/param.cpp Thu Apr 05 17:16:07 2018 +0530
> @@ -99,13 +99,13 @@
> {
> x265_free(p);
> }
> -
> +bool benableavx512 = false;
> void x265_param_default(x265_param* param)
> {
> memset(param, 0, sizeof(x265_param));
>
> /* Applying default values to all elements in the param structure */
> - param->cpuid = X265_NS::cpu_detect();
> + param->cpuid = X265_NS::cpu_detect(benableavx512);
> param->bEnableWavefront = 1;
> param->frameNumThreads = 0;
>
> @@ -609,6 +609,17 @@
> if (0) ;
> OPT("asm")
> {
> + sscanf(value, "%s", p->asmname);
p->asmname is a pointer to unallocated memory. It is not used in x265.
VS 2015 and VS 2017 builds hangs at command-line:
x265 --asm sse4 -V
I think we should remove 'asmname' from x265_param
or allocate memory before we copy anything to this location.
> + if (strcmp(value, "avx512")==0)
> + {
> + p->bEnableavx512 = 1;
> + benableavx512 = true;
> + }
> + else
> + {
> + p->bEnableavx512 = 0;
> + benableavx512 = false;
> + }
> if (bValueWasNull)
> p->cpuid = atobool(value);
> else
> @@ -1072,7 +1083,7 @@
> if (isdigit(value[0]))
> cpu = x265_atoi(value, bError);
> else
> - cpu = !strcmp(value, "auto") || x265_atobool(value, bError) ? X265_NS::cpu_detect() : 0;
> + cpu = !strcmp(value, "auto") || x265_atobool(value, bError) ? X265_NS::cpu_detect(benableavx512) : 0;
>
> if (bError)
> {
> diff -r 876b6e006f20 -r f6ad2fa637fd source/test/pixelharness.cpp
> --- a/source/test/pixelharness.cpp Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/test/pixelharness.cpp Thu Apr 05 17:16:07 2018 +0530
> @@ -332,8 +332,9 @@
> memset(ref_dest, 0, 64 * 64 * sizeof(pixel));
> memset(opt_dest, 0, 64 * 64 * sizeof(pixel));
> int j = 0;
> + bool enableavx512 = true;
> int width = 16 * (rand() % 4 + 1);
> - int cpuid = X265_NS::cpu_detect();
> + int cpuid = X265_NS::cpu_detect(enableavx512);
> if (cpuid & X265_CPU_AVX512)
> width = 32 * (rand() % 2 + 1);
> int height = 8;
> diff -r 876b6e006f20 -r f6ad2fa637fd source/test/testbench.cpp
> --- a/source/test/testbench.cpp Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/test/testbench.cpp Thu Apr 05 17:16:07 2018 +0530
> @@ -96,7 +96,8 @@
>
> int main(int argc, char *argv[])
> {
> - int cpuid = X265_NS::cpu_detect();
> + bool enableavx512 = true;
> + int cpuid = X265_NS::cpu_detect(enableavx512);
> const char *testname = 0;
>
> if (!(argc & 1))
> diff -r 876b6e006f20 -r f6ad2fa637fd source/x265.h
> --- a/source/x265.h Mon Feb 05 10:39:00 2018 -0800
> +++ b/source/x265.h Thu Apr 05 17:16:07 2018 +0530
> @@ -585,7 +585,14 @@
> * somehow flawed on your target hardware. The asm function tables are
> * process global, the first encoder configures them for all encoders */
> int cpuid;
> -
> + /*==Assembly features ==*/
> + /* x265_param_parse() will detect if the avx512 is enabled (in cli )and set
> + * bEnableavx512 to 1 to use avx512 SIMD. By default this flag will not be set ,
> + * hence the encoding will happen without avx512 assembly primitives even if the cpu has
> + * avx512 capabilities.
> + * Ensure to use --asm avx512 if you need to encode with avx512 assembly primitives*/
> + int bEnableavx512;
> + char* asmname;
> /*== Parallelism Features ==*/
>
> /* Number of concurrently encoded frames between 1 and X265_MAX_FRAME_THREADS
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
More information about the x265-devel
mailing list