diff options
| author | Brecht Van Lommel <brecht@blender.org> | 2021-11-05 23:01:23 +0300 |
|---|---|---|
| committer | Brecht Van Lommel <brecht@blender.org> | 2021-11-06 00:04:36 +0300 |
| commit | 97ff37bf54474efbce39653a1387ad55091d4964 (patch) | |
| tree | 58ff9592807dbd98d126b179627e5c56f5309956 /intern/cycles/integrator/pass_accessor_cpu.h | |
| parent | d1a9425a2fde32b6786b333ab55661da507e818b (diff) | |
Cycles: perform CPU film reading in the kernel, to use AVX2 half conversion
Adds a bunch of CPU kernel function to process on row of pixels, and use those
instead of calling unoptimized implementations.
Fixes T92598
Diffstat (limited to 'intern/cycles/integrator/pass_accessor_cpu.h')
| -rw-r--r-- | intern/cycles/integrator/pass_accessor_cpu.h | 32 |
1 files changed, 14 insertions, 18 deletions
diff --git a/intern/cycles/integrator/pass_accessor_cpu.h b/intern/cycles/integrator/pass_accessor_cpu.h index 0313dc5bb0d..9ed38ab256e 100644 --- a/intern/cycles/integrator/pass_accessor_cpu.h +++ b/intern/cycles/integrator/pass_accessor_cpu.h @@ -16,6 +16,8 @@ #pragma once +#include "device/cpu/kernel.h" + #include "integrator/pass_accessor.h" CCL_NAMESPACE_BEGIN @@ -28,25 +30,19 @@ class PassAccessorCPU : public PassAccessor { using PassAccessor::PassAccessor; protected: - template<typename Processor> - inline void run_get_pass_kernel_processor(const RenderBuffers *render_buffers, - const BufferParams &buffer_params, - const Destination &destination, - const Processor &processor) const; - - template<typename Processor> - inline void run_get_pass_kernel_processor_float(const KernelFilmConvert *kfilm_convert, - const RenderBuffers *render_buffers, - const BufferParams &buffer_params, - const Destination &destination, - const Processor &processor) const; + inline void run_get_pass_kernel_processor_float( + const KernelFilmConvert *kfilm_convert, + const RenderBuffers *render_buffers, + const BufferParams &buffer_params, + const Destination &destination, + const CPUKernels::FilmConvertFunction func) const; - template<typename Processor> - inline void run_get_pass_kernel_processor_half_rgba(const KernelFilmConvert *kfilm_convert, - const RenderBuffers *render_buffers, - const BufferParams &buffer_params, - const Destination &destination, - const Processor &processor) const; + inline void run_get_pass_kernel_processor_half_rgba( + const KernelFilmConvert *kfilm_convert, + const RenderBuffers *render_buffers, + const BufferParams &buffer_params, + const Destination &destination, + const CPUKernels::FilmConvertHalfRGBAFunction func) const; #define DECLARE_PASS_ACCESSOR(pass) \ virtual void get_pass_##pass(const RenderBuffers *render_buffers, \ |