[Bf-blender-cvs] [de28a4d] master: Cycles: Add an AVX kernel for CPU rendering.
Thomas Dinges
noreply at git.blender.org
Thu Jan 16 17:05:16 CET 2014
Commit: de28a4d4b2c9397c5233a5ee1dbf1400f450a15c
Author: Thomas Dinges
Date: Thu Jan 16 17:04:11 2014 +0100
https://developer.blender.org/rBde28a4d4b2c9397c5233a5ee1dbf1400f450a15c
Cycles: Add an AVX kernel for CPU rendering.
* AVX is available on Intel Sandy Bridge and newer and AMD Bulldozer and newer.
* We don't use dedicated AVX intrinsics yet, but gcc auto vectorization gives a 3% performance improvement for Caminandes. Tested on an i5-3570, Linux x64.
* No change for Windows yet, MSVC 2008 does not support AVX.
Reviewed by: brecht
Differential Revision: https://developer.blender.org/D216
===================================================================
M intern/cycles/CMakeLists.txt
M intern/cycles/SConscript
M intern/cycles/device/device_cpu.cpp
M intern/cycles/kernel/CMakeLists.txt
M intern/cycles/kernel/kernel.h
A intern/cycles/kernel/kernel_avx.cpp
M intern/cycles/util/util_optimization.h
M intern/cycles/util/util_system.cpp
M intern/cycles/util/util_system.h
===================================================================
diff --git a/intern/cycles/CMakeLists.txt b/intern/cycles/CMakeLists.txt
index 6fa6260..cc1f597 100644
--- a/intern/cycles/CMakeLists.txt
+++ b/intern/cycles/CMakeLists.txt
@@ -14,10 +14,12 @@ if(WIN32 AND MSVC)
set(CYCLES_SSE2_KERNEL_FLAGS "/fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
set(CYCLES_SSE3_KERNEL_FLAGS "/fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
set(CYCLES_SSE41_KERNEL_FLAGS "/fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
+ set(CYCLES_AVX_KERNEL_FLAGS "/fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-") #/arch:AVX for VC2012 and above
else()
set(CYCLES_SSE2_KERNEL_FLAGS "/arch:SSE2 /fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
set(CYCLES_SSE3_KERNEL_FLAGS "/arch:SSE2 /fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
set(CYCLES_SSE41_KERNEL_FLAGS "/arch:SSE2 /fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
+ set(CYCLES_AVX_KERNEL_FLAGS "/arch:SSE2 /fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-") #/arch:AVX for VC2012 and above
endif()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /fp:fast -D_CRT_SECURE_NO_WARNINGS /Gs-")
@@ -28,11 +30,13 @@ elseif(CMAKE_COMPILER_IS_GNUCC)
set(CYCLES_SSE2_KERNEL_FLAGS "-ffast-math -msse -msse2 -mfpmath=sse")
set(CYCLES_SSE3_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3 -mfpmath=sse")
set(CYCLES_SSE41_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1 -mfpmath=sse")
+ set(CYCLES_AVX_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1 -mavx -mfpmath=sse")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ffast-math")
elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
set(CYCLES_SSE2_KERNEL_FLAGS "-ffast-math -msse -msse2")
set(CYCLES_SSE3_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3")
set(CYCLES_SSE41_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1")
+ set(CYCLES_SSE41_KERNEL_FLAGS "-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1 -mavx")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ffast-math")
endif()
diff --git a/intern/cycles/SConscript b/intern/cycles/SConscript
index b1dfeee..7ced461 100644
--- a/intern/cycles/SConscript
+++ b/intern/cycles/SConscript
@@ -38,6 +38,7 @@ sources.remove(path.join('util', 'util_view.cpp'))
sources.remove(path.join('kernel', 'kernel_sse2.cpp'))
sources.remove(path.join('kernel', 'kernel_sse3.cpp'))
sources.remove(path.join('kernel', 'kernel_sse41.cpp'))
+sources.remove(path.join('kernel', 'kernel_avx.cpp'))
incs = []
defs = []
@@ -78,23 +79,31 @@ if env['OURPLATFORM'] in ('win32-vc', 'win32-mingw', 'linuxcross', 'win64-vc', '
sse2_cxxflags = Split(env['CXXFLAGS'])
sse3_cxxflags = Split(env['CXXFLAGS'])
sse41_cxxflags = Split(env['CXXFLAGS'])
+avx_cxxflags = Split(env['CXXFLAGS'])
if env['OURPLATFORM'] == 'win32-vc':
# there is no /arch:SSE3, but intrinsics are available anyway
sse2_cxxflags.append('/arch:SSE /arch:SSE2 -D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
sse3_cxxflags.append('/arch:SSE /arch:SSE2 -D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
sse41_cxxflags.append('/arch:SSE /arch:SSE2 -D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
+ avx_cxxflags.append('/arch:SSE /arch:SSE2 -D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split()) #/arch:AVX for VC2012 and above
elif env['OURPLATFORM'] == 'win64-vc':
sse2_cxxflags.append('-D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
sse3_cxxflags.append('-D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
sse41_cxxflags.append('-D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split())
+ avx_cxxflags.append('-D_CRT_SECURE_NO_WARNINGS /fp:fast /Ox /Gs-'.split()) #/arch:AVX for VC2012 and above
else:
sse2_cxxflags.append('-ffast-math -msse -msse2 -mfpmath=sse'.split())
sse3_cxxflags.append('-ffast-math -msse -msse2 -msse3 -mssse3 -mfpmath=sse'.split())
sse41_cxxflags.append('-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1 -mfpmath=sse'.split())
+ avx_cxxflags.append('-ffast-math -msse -msse2 -msse3 -mssse3 -msse4.1 -mavx -mfpmath=sse'.split())
optim_defs = defs[:]
+cycles_avx = cycles.Clone()
+avx_sources = [path.join('kernel', 'kernel_avx.cpp')]
+cycles_avx.BlenderLib('bf_intern_cycles_avx', avx_sources, incs, optim_defs, libtype=['intern'], priority=[10], cxx_compileflags=avx_cxxflags)
+
cycles_sse41 = cycles.Clone()
sse41_sources = [path.join('kernel', 'kernel_sse41.cpp')]
cycles_sse41.BlenderLib('bf_intern_cycles_sse41', sse41_sources, incs, optim_defs, libtype=['intern'], priority=[10], cxx_compileflags=sse41_cxxflags)
diff --git a/intern/cycles/device/device_cpu.cpp b/intern/cycles/device/device_cpu.cpp
index b29d64e..76123fe 100644
--- a/intern/cycles/device/device_cpu.cpp
+++ b/intern/cycles/device/device_cpu.cpp
@@ -61,6 +61,7 @@ public:
system_cpu_support_sse2();
system_cpu_support_sse3();
system_cpu_support_sse41();
+ system_cpu_support_avx();
}
~CPUDevice()
@@ -166,6 +167,28 @@ public:
int start_sample = tile.start_sample;
int end_sample = tile.start_sample + tile.num_samples;
+#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
+ if(system_cpu_support_avx()) {
+ for(int sample = start_sample; sample < end_sample; sample++) {
+ if (task.get_cancel() || task_pool.canceled()) {
+ if(task.need_finish_queue == false)
+ break;
+ }
+
+ for(int y = tile.y; y < tile.y + tile.h; y++) {
+ for(int x = tile.x; x < tile.x + tile.w; x++) {
+ kernel_cpu_avx_path_trace(&kg, render_buffer, rng_state,
+ sample, x, y, tile.offset, tile.stride);
+ }
+ }
+
+ tile.sample = sample + 1;
+
+ task.update_progress(tile);
+ }
+ }
+ else
+#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41()) {
for(int sample = start_sample; sample < end_sample; sample++) {
@@ -270,6 +293,15 @@ public:
float sample_scale = 1.0f/(task.sample + 1);
if(task.rgba_half) {
+#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
+ if(system_cpu_support_avx()) {
+ for(int y = task.y; y < task.y + task.h; y++)
+ for(int x = task.x; x < task.x + task.w; x++)
+ kernel_cpu_avx_convert_to_half_float(&kernel_globals, (uchar4*)task.rgba_half, (float*)task.buffer,
+ sample_scale, x, y, task.offset, task.stride);
+ }
+ else
+#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41()) {
for(int y = task.y; y < task.y + task.h; y++)
@@ -305,6 +337,15 @@ public:
}
}
else {
+#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
+ if(system_cpu_support_avx()) {
+ for(int y = task.y; y < task.y + task.h; y++)
+ for(int x = task.x; x < task.x + task.w; x++)
+ kernel_cpu_avx_convert_to_byte(&kernel_globals, (uchar4*)task.rgba_byte, (float*)task.buffer,
+ sample_scale, x, y, task.offset, task.stride);
+ }
+ else
+#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41()) {
for(int y = task.y; y < task.y + task.h; y++)
@@ -349,6 +390,17 @@ public:
OSLShader::thread_init(&kg, &kernel_globals, &osl_globals);
#endif
+#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
+ if(system_cpu_support_avx()) {
+ for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++) {
+ kernel_cpu_avx_shader(&kg, (uint4*)task.shader_input, (float4*)task.shader_output, task.shader_eval_type, x);
+
+ if(task_pool.canceled())
+ break;
+ }
+ }
+ else
+#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41()) {
for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++) {
diff --git a/intern/cycles/kernel/CMakeLists.txt b/intern/cycles/kernel/CMakeLists.txt
index 08427b1..6f5f744 100644
--- a/intern/cycles/kernel/CMakeLists.txt
+++ b/intern/cycles/kernel/CMakeLists.txt
@@ -15,6 +15,7 @@ set(SRC
kernel_sse2.cpp
kernel_sse3.cpp
kernel_sse41.cpp
+ kernel_avx.cpp
kernel.cl
kernel.cu
)
@@ -197,6 +198,7 @@ include_directories(SYSTEM ${INC_SYS})
set_source_files_properties(kernel_sse2.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE2_KERNEL_FLAGS}")
set_source_files_properties(kernel_sse3.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE3_KERNEL_FLAGS}")
set_source_files_properties(kernel_sse41.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_SSE41_KERNEL_FLAGS}")
+set_source_files_properties(kernel_avx.cpp PROPERTIES COMPILE_FLAGS "${CYCLES_AVX_KERNEL_FLAGS}")
add_library(cycles_kernel ${SRC} ${SRC_HEADERS} ${SRC_CLOSURE_HEADERS} ${SRC_SVM_HEADERS})
diff --git a/intern/cycles/kernel/kernel.h b/intern/cycles/kernel/kernel.h
index 01bea10..039dc79 100644
--- a/intern/cycles/kernel/kernel.h
+++ b/intern/cycles/kernel/kernel.h
@@ -76,6 +76,17 @@ void kernel_cpu_sse41_shader(KernelGlobals *kg, uint4 *input, float4 *output,
int type, int i);
#endif
+#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
+void kernel_cpu_avx_path_trace(KernelGlobals *kg, float *buffer, unsigned int *rng_state,
+ int sample, int x, int y, int offset, int stride);
+void kernel_cpu_avx_convert_to_byte(KernelGlobals *kg, uchar4 *rgba, float *buffer,
+ float sample_scale, int x, int y, int offset, int stride);
+void kernel_cpu_avx_convert_to_half_float(KernelGlobals *kg, uchar4 *rgba, float *buffer,
+ float sample_scale, int x, int y, int offset, int stride);
+void kernel_cpu_avx_shader(KernelGlobals *kg, uint4 *input, float4 *output,
+ int type, int i);
+#endif
+
CCL_NAMESPACE_END
#endif /* __KERNEL_H__ */
diff --git a/intern/cycles/kernel/kernel_avx.cpp b/intern/cycles/kernel/kernel_avx.cpp
new file mode 100644
index 0000000..d2a7142
--- /dev/null
+++ b/intern/cycles/kernel/kernel_avx.cpp
@@ -0,0 +1,82 @@
+/*
+ * Copyright 2011-2013 Blender Foundation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License
+ */
+
+/* Optimized CPU kernel entry points. This file is compiled with AVX
+ * optimization flags and nearly all functions inline
@@ Diff output truncated at 10240 characters. @@
More information about the Bf-blender-cvs
mailing list