CPUDevice(DeviceInfo& info, Stats &stats, bool background)
: Device(info, stats, background)
{
#ifdef WITH_OSL
kernel_globals.osl = &osl_globals;
#endif
/* do now to avoid thread issues */
system_cpu_support_sse2();
system_cpu_support_sse3();
system_cpu_support_sse41();
system_cpu_support_avx();
system_cpu_support_avx2();
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX2
if(system_cpu_support_avx2()) {
VLOG(1) << "Will be using AVX2 kernels.";
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
if(system_cpu_support_avx()) {
VLOG(1) << "Will be using AVX kernels.";
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41()) {
VLOG(1) << "Will be using SSE4.1 kernels.";
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
if(system_cpu_support_sse3()) {
VLOG(1) << "Will be using SSE3kernels.";
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
if(system_cpu_support_sse2()) {
VLOG(1) << "Will be using SSE2 kernels.";
}
else
#endif
{
VLOG(1) << "Will be using regular kernels.";
}
}
string device_cpu_capabilities(void)
{
string capabilities = "";
capabilities += system_cpu_support_sse2() ? "SSE2 " : "";
capabilities += system_cpu_support_sse3() ? "SSE3 " : "";
capabilities += system_cpu_support_sse41() ? "SSE41 " : "";
capabilities += system_cpu_support_avx() ? "AVX " : "";
capabilities += system_cpu_support_avx2() ? "AVX2" : "";
if(capabilities[capabilities.size() - 1] == ' ')
capabilities.resize(capabilities.size() - 1);
return capabilities;
}
KernelFunctions(
F kernel_default, F kernel_sse2, F kernel_sse3, F kernel_sse41, F kernel_avx, F kernel_avx2)
{
const char *architecture_name = "default";
kernel = kernel_default;
/* Silence potential warnings about unused variables
* when compiling without some architectures. */
(void)kernel_sse2;
(void)kernel_sse3;
(void)kernel_sse41;
(void)kernel_avx;
(void)kernel_avx2;
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX2
if (DebugFlags().cpu.has_avx2() && system_cpu_support_avx2()) {
architecture_name = "AVX2";
kernel = kernel_avx2;
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
if (DebugFlags().cpu.has_avx() && system_cpu_support_avx()) {
architecture_name = "AVX";
kernel = kernel_avx;
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if (DebugFlags().cpu.has_sse41() && system_cpu_support_sse41()) {
architecture_name = "SSE4.1";
kernel = kernel_sse41;
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
if (DebugFlags().cpu.has_sse3() && system_cpu_support_sse3()) {
architecture_name = "SSE3";
kernel = kernel_sse3;
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
if (DebugFlags().cpu.has_sse2() && system_cpu_support_sse2()) {
architecture_name = "SSE2";
kernel = kernel_sse2;
}
#endif
if (strcmp(architecture_name, logged_architecture) != 0) {
VLOG(1) << "Will be using " << architecture_name << " kernels.";
logged_architecture = architecture_name;
}
}
void thread_shader(DeviceTask& task)
{
KernelGlobals kg = kernel_globals;
#ifdef WITH_OSL
OSLShader::thread_init(&kg, &kernel_globals, &osl_globals);
#endif
void(*shader_kernel)(KernelGlobals*, uint4*, float4*, int, int, int, int);
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX2
if(system_cpu_support_avx2())
shader_kernel = kernel_cpu_avx2_shader;
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_AVX
if(system_cpu_support_avx())
shader_kernel = kernel_cpu_avx_shader;
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE41
if(system_cpu_support_sse41())
shader_kernel = kernel_cpu_sse41_shader;
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
if(system_cpu_support_sse3())
shader_kernel = kernel_cpu_sse3_shader;
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
if(system_cpu_support_sse2())
shader_kernel = kernel_cpu_sse2_shader;
else
#endif
shader_kernel = kernel_cpu_shader;
for(int sample = 0; sample < task.num_samples; sample++) {
for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++)
shader_kernel(&kg, (uint4*)task.shader_input, (float4*)task.shader_output,
task.shader_eval_type, x, task.offset, sample);
if(task.get_cancel() || task_pool.canceled())
break;
task.update_progress(NULL);
}
#ifdef WITH_OSL
OSLShader::thread_free(&kg);
#endif
}
CPUDevice(DeviceInfo& info, Stats &stats, bool background)
: Device(info, stats, background)
{
#ifdef WITH_OSL
kernel_globals.osl = &osl_globals;
#endif
/* do now to avoid thread issues */
system_cpu_support_sse2();
system_cpu_support_sse3();
system_cpu_support_sse41();
system_cpu_support_avx();
}
void thread_shader(DeviceTask& task)
{
KernelGlobals kg = kernel_globals;
#ifdef WITH_OSL
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.get_cancel() || 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++) {
kernel_cpu_sse41_shader(&kg, (uint4*)task.shader_input, (float4*)task.shader_output, task.shader_eval_type, x);
if(task.get_cancel() || task_pool.canceled())
break;
}
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE3
if(system_cpu_support_sse3()) {
for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++) {
kernel_cpu_sse3_shader(&kg, (uint4*)task.shader_input, (float4*)task.shader_output, task.shader_eval_type, x);
if(task.get_cancel() || task_pool.canceled())
break;
}
}
else
#endif
#ifdef WITH_CYCLES_OPTIMIZED_KERNEL_SSE2
if(system_cpu_support_sse2()) {
for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++) {
kernel_cpu_sse2_shader(&kg, (uint4*)task.shader_input, (float4*)task.shader_output, task.shader_eval_type, x);
if(task.get_cancel() || task_pool.canceled())
break;
}
}
else
#endif
{
for(int x = task.shader_x; x < task.shader_x + task.shader_w; x++) {
kernel_cpu_shader(&kg, (uint4*)task.shader_input, (float4*)task.shader_output, task.shader_eval_type, x);
if(task.get_cancel() || task_pool.canceled())
break;
}
}
#ifdef WITH_OSL
OSLShader::thread_free(&kg);
#endif
}
void thread_film_convert(DeviceTask& task)
{
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++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse41_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_SSE3
if(system_cpu_support_sse3()) {
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse3_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_SSE2
if(system_cpu_support_sse2()) {
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse2_convert_to_half_float(&kernel_globals, (uchar4*)task.rgba_half, (float*)task.buffer,
sample_scale, x, y, task.offset, task.stride);
}
else
#endif
{
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_convert_to_half_float(&kernel_globals, (uchar4*)task.rgba_half, (float*)task.buffer,
sample_scale, x, y, task.offset, task.stride);
}
}
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++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse41_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_SSE3
if(system_cpu_support_sse3()) {
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse3_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_SSE2
if(system_cpu_support_sse2()) {
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_sse2_convert_to_byte(&kernel_globals, (uchar4*)task.rgba_byte, (float*)task.buffer,
sample_scale, x, y, task.offset, task.stride);
}
else
#endif
{
for(int y = task.y; y < task.y + task.h; y++)
for(int x = task.x; x < task.x + task.w; x++)
kernel_cpu_convert_to_byte(&kernel_globals, (uchar4*)task.rgba_byte, (float*)task.buffer,
sample_scale, x, y, task.offset, task.stride);
}
}
}
void thread_path_trace(DeviceTask& task)
{
if(task_pool.canceled()) {
if(task.need_finish_queue == false)
return;
}
KernelGlobals kg = kernel_globals;
#ifdef WITH_OSL
OSLShader::thread_init(&kg, &kernel_globals, &osl_globals);
#endif
RenderTile tile;
while(task.acquire_tile(this, tile)) {
float *render_buffer = (float*)tile.buffer;
uint *rng_state = (uint*)tile.rng_state;
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++) {
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_sse41_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_SSE3
if(system_cpu_support_sse3()) {
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_sse3_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_SSE2
if(system_cpu_support_sse2()) {
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_sse2_path_trace(&kg, render_buffer, rng_state,
sample, x, y, tile.offset, tile.stride);
}
}
tile.sample = sample + 1;
task.update_progress(tile);
}
}
else
#endif
{
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_path_trace(&kg, render_buffer, rng_state,
sample, x, y, tile.offset, tile.stride);
}
}
tile.sample = sample + 1;
task.update_progress(tile);
}
}
task.release_tile(tile);
if(task_pool.canceled()) {
if(task.need_finish_queue == false)
break;
}
}
#ifdef WITH_OSL
OSLShader::thread_free(&kg);
#endif
}
