From 86de65fbf0743b3159f9c7b3af97203f8ec9743e Mon Sep 17 00:00:00 2001 From: Rodger Combs Date: Tue, 30 Jul 2019 02:51:42 -0500 Subject: [PATCH] build: add support for building CUDA files with clang This avoids using the CUDA SDK at all; instead, we provide a minimal reimplementation of the basic functionality that lavfi actually uses. It generates very similar code to what NVCC produces. The header contains no implementation code derived from the SDK. The function and type declarations are derived from the SDK only to the extent required to build a compatible implementation. This is generally accepted to qualify as fair use. Because this option does not require the proprietary SDK, it does not require the "--enable-nonfree" flag in configure. Signed-off-by: Timo Rothenpieler --- Changelog | 1 + compat/cuda/cuda_runtime.h | 131 +++++++++++++++++++++++++++++++++++++ configure | 67 ++++++++++++------- ffbuild/common.mak | 3 +- 4 files changed, 178 insertions(+), 24 deletions(-) create mode 100644 compat/cuda/cuda_runtime.h diff --git a/Changelog b/Changelog index beb2d2615d..389ca6c4db 100644 --- a/Changelog +++ b/Changelog @@ -36,6 +36,7 @@ version 4.2: - derain filter - deesser filter - mov muxer writes tracks with unspecified language instead of English by default +- add support for using clang to compile CUDA kernels version 4.1: diff --git a/compat/cuda/cuda_runtime.h b/compat/cuda/cuda_runtime.h new file mode 100644 index 0000000000..dbe50f8711 --- /dev/null +++ b/compat/cuda/cuda_runtime.h @@ -0,0 +1,131 @@ +/* + * Minimum CUDA compatibility definitions header + * + * Copyright (c) 2019 Rodger Combs + * + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#ifndef AV_COMPAT_CUDA_CUDA_RUNTIME_H +#define AV_COMPAT_CUDA_CUDA_RUNTIME_H + +// Common macros +#define __global__ __attribute__((global)) +#define __device__ __attribute__((device)) +#define __device_builtin__ __attribute__((device_builtin)) +#define __align__(N) __attribute__((aligned(N))) +#define __inline__ __inline__ __attribute__((always_inline)) + +#define max(a, b) ((a) > (b) ? (a) : (b)) +#define min(a, b) ((a) < (b) ? (a) : (b)) +#define abs(x) ((x) < 0 ? -(x) : (x)) + +#define atomicAdd(a, b) (__atomic_fetch_add(a, b, __ATOMIC_SEQ_CST)) + +// Basic typedefs +typedef __device_builtin__ unsigned long long cudaTextureObject_t; + +typedef struct __device_builtin__ __align__(2) uchar2 +{ + unsigned char x, y; +} uchar2; + +typedef struct __device_builtin__ __align__(4) ushort2 +{ + unsigned short x, y; +} ushort2; + +typedef struct __device_builtin__ uint3 +{ + unsigned int x, y, z; +} uint3; + +typedef struct uint3 dim3; + +typedef struct __device_builtin__ __align__(8) int2 +{ + int x, y; +} int2; + +typedef struct __device_builtin__ __align__(4) uchar4 +{ + unsigned char x, y, z, w; +} uchar4; + +typedef struct __device_builtin__ __align__(8) ushort4 +{ + unsigned char x, y, z, w; +} ushort4; + +typedef struct __device_builtin__ __align__(16) int4 +{ + int x, y, z, w; +} int4; + +// Accessors for special registers +#define GETCOMP(reg, comp) \ + asm("mov.u32 %0, %%" #reg "." #comp ";" : "=r"(tmp)); \ + ret.comp = tmp; + +#define GET(name, reg) static inline __device__ uint3 name() {\ + uint3 ret; \ + unsigned tmp; \ + GETCOMP(reg, x) \ + GETCOMP(reg, y) \ + GETCOMP(reg, z) \ + return ret; \ +} + +GET(getBlockIdx, ctaid) +GET(getBlockDim, ntid) +GET(getThreadIdx, tid) + +// Instead of externs for these registers, we turn access to them into calls into trivial ASM +#define blockIdx (getBlockIdx()) +#define blockDim (getBlockDim()) +#define threadIdx (getThreadIdx()) + +// Basic initializers (simple macros rather than inline functions) +#define make_uchar2(a, b) ((uchar2){.x = a, .y = b}) +#define make_ushort2(a, b) ((ushort2){.x = a, .y = b}) +#define make_uchar4(a, b, c, d) ((uchar4){.x = a, .y = b, .z = c, .w = d}) +#define make_ushort4(a, b, c, d) ((ushort4){.x = a, .y = b, .z = c, .w = d}) + +// Conversions from the tex instruction's 4-register output to various types +#define TEX2D(type, ret) static inline __device__ void conv(type* out, unsigned a, unsigned b, unsigned c, unsigned d) {*out = (ret);} + +TEX2D(unsigned char, a & 0xFF) +TEX2D(unsigned short, a & 0xFFFF) +TEX2D(uchar2, make_uchar2(a & 0xFF, b & 0xFF)) +TEX2D(ushort2, make_ushort2(a & 0xFFFF, b & 0xFFFF)) +TEX2D(uchar4, make_uchar4(a & 0xFF, b & 0xFF, c & 0xFF, d & 0xFF)) +TEX2D(ushort4, make_ushort4(a & 0xFFFF, b & 0xFFFF, c & 0xFFFF, d & 0xFFFF)) + +// Template calling tex instruction and converting the output to the selected type +template +static inline __device__ T tex2D(cudaTextureObject_t texObject, float x, float y) +{ + T ret; + unsigned ret1, ret2, ret3, ret4; + asm("tex.2d.v4.u32.f32 {%0, %1, %2, %3}, [%4, {%5, %6}];" : + "=r"(ret1), "=r"(ret2), "=r"(ret3), "=r"(ret4) : + "l"(texObject), "f"(x), "f"(y)); + conv(&ret, ret1, ret2, ret3, ret4); + return ret; +} + +#endif diff --git a/configure b/configure index 5a4f507246..bddc05b850 100755 --- a/configure +++ b/configure @@ -322,6 +322,7 @@ External library support: --disable-amf disable AMF video encoding code [autodetect] --disable-audiotoolbox disable Apple AudioToolbox code [autodetect] --enable-cuda-nvcc enable Nvidia CUDA compiler [no] + --disable-cuda-llvm disable CUDA compilation using clang [autodetect] --disable-cuvid disable Nvidia CUVID support [autodetect] --disable-d3d11va disable Microsoft Direct3D 11 video acceleration code [autodetect] --disable-dxva2 disable Microsoft DirectX 9 video acceleration code [autodetect] @@ -370,7 +371,7 @@ Toolchain options: --cxx=CXX use C compiler CXX [$cxx_default] --objcc=OCC use ObjC compiler OCC [$cc_default] --dep-cc=DEPCC use dependency generator DEPCC [$cc_default] - --nvcc=NVCC use Nvidia CUDA compiler NVCC [$nvcc_default] + --nvcc=NVCC use Nvidia CUDA compiler NVCC or clang [$nvcc_default] --ld=LD use linker LD [$ld_default] --pkg-config=PKGCONFIG use pkg-config tool PKGCONFIG [$pkg_config_default] --pkg-config-flags=FLAGS pass additional flags to pkgconf [] @@ -1038,12 +1039,16 @@ test_nvcc(){ tmpcu_=$TMPCU tmpo_=$TMPO [ -x "$(command -v cygpath)" ] && tmpcu_=$(cygpath -m $tmpcu_) && tmpo_=$(cygpath -m $tmpo_) - test_cmd $nvcc -ptx $NVCCFLAGS "$@" $NVCC_C $(nvcc_o $tmpo_) $tmpcu_ + test_cmd $nvcc $nvccflags "$@" $NVCC_C $(nvcc_o $tmpo_) $tmpcu_ } check_nvcc() { log check_nvcc "$@" - test_nvcc <= 0x0a090500" || die "ERROR: Decklink API version must be >= 10.9.5."; } } @@ -6701,16 +6734,6 @@ if [ -z "$optflags" ]; then fi fi -if [ -z "$nvccflags" ]; then - nvccflags=$nvccflags_default -fi - -if enabled x86_64 || enabled ppc64 || enabled aarch64; then - nvccflags="$nvccflags -m64" -else - nvccflags="$nvccflags -m32" -fi - check_optflags(){ check_cflags "$@" enabled lto && check_ldflags "$@" diff --git a/ffbuild/common.mak b/ffbuild/common.mak index d2b33320c0..7355508ea0 100644 --- a/ffbuild/common.mak +++ b/ffbuild/common.mak @@ -38,7 +38,6 @@ OBJCCFLAGS = $(CPPFLAGS) $(CFLAGS) $(OBJCFLAGS) ASFLAGS := $(CPPFLAGS) $(ASFLAGS) CXXFLAGS := $(CPPFLAGS) $(CFLAGS) $(CXXFLAGS) X86ASMFLAGS += $(IFLAGS:%=%/) -I$($@ -%.ptx: %.cu +%.ptx: %.cu $(SRC_PATH)/compat/cuda/cuda_runtime.h $(COMPILE_NVCC) %.ptx.c: %.ptx