/*
* Copyright (C) 2020 Linux Studio Plugins Project
* (C) 2020 Vladimir Sadovnikov
*
* This file is part of lsp-dsp-lib
* Created on: 31 мар. 2020 г.
*
* lsp-dsp-lib 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 3 of the License, or
* any later version.
*
* lsp-dsp-lib 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 lsp-dsp-lib. If not, see .
*/
#ifndef PRIVATE_DSP_ARCH_X86_AVX_FFT_BUTTERFLY_H_
#define PRIVATE_DSP_ARCH_X86_AVX_FFT_BUTTERFLY_H_
#ifndef PRIVATE_DSP_ARCH_X86_AVX_IMPL
#error "This header should not be included directly"
#endif /* PRIVATE_DSP_ARCH_X86_AVX_IMPL */
namespace lsp
{
namespace avx
{
#define FFT_BUTTERFLY_BODY8(add_b, add_a, FMA_SEL) \
__IF_32(float *ptr1, *ptr2);\
\
ARCH_X86_ASM \
( \
/* Prepare angle */ \
__ASM_EMIT32("mov %[fft_a], %[ptr2]") \
__ASM_EMIT32("mov %[dst_re], %[ptr1]") \
__ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_a") "]), %%ymm6") /* ymm6 = x_re */ \
__ASM_EMIT("vmovaps 0x20(%[" __IF_32_64("ptr2", "fft_a") "]), %%ymm7") /* ymm7 = x_im */ \
__ASM_EMIT32("mov %[dst_im], %[ptr2]") \
/* Start loop */ \
__ASM_EMIT("1:") \
__ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1]), %%ymm0") /* ymm0 = a_re */ \
__ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2]), %%ymm2") /* ymm2 = b_re */ \
__ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1]), %%ymm1") /* ymm1 = a_im */ \
__ASM_EMIT("vmovups 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2]), %%ymm3") /* ymm3 = b_im */ \
/* Calculate complex multiplication */ \
__ASM_EMIT("vmulps %%ymm7, %%ymm2, %%ymm4") /* ymm4 = x_im * b_re */ \
__ASM_EMIT("vmulps %%ymm7, %%ymm3, %%ymm5") /* ymm5 = x_im * b_im */ \
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm2, %%ymm2", "")) /* ymm2 = x_re * b_re */ \
__ASM_EMIT(FMA_SEL("vmulps %%ymm6, %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_im */ \
__ASM_EMIT(FMA_SEL(add_b " %%ymm5, %%ymm2, %%ymm5", add_b " %%ymm6, %%ymm2, %%ymm5")) /* ymm5 = c_re = x_re * b_re +- x_im * b_im */ \
__ASM_EMIT(FMA_SEL(add_a " %%ymm4, %%ymm3, %%ymm4", add_a " %%ymm6, %%ymm3, %%ymm4")) /* ymm4 = c_im = x_re * b_im -+ x_im * b_re */ \
/* Perform butterfly */ \
__ASM_EMIT("vsubps %%ymm5, %%ymm0, %%ymm2") /* ymm2 = a_re - c_re */ \
__ASM_EMIT("vsubps %%ymm4, %%ymm1, %%ymm3") /* ymm3 = a_im - c_im */ \
__ASM_EMIT("vaddps %%ymm5, %%ymm0, %%ymm0") /* ymm0 = a_re + c_re */ \
__ASM_EMIT("vaddps %%ymm4, %%ymm1, %%ymm1") /* ymm1 = a_im + c_im */ \
/* Store values */ \
__ASM_EMIT("vmovups %%ymm0, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off1])") \
__ASM_EMIT("vmovups %%ymm2, 0x00(%[" __IF_32_64("ptr1", "dst_re") "], %[off2])") \
__ASM_EMIT("vmovups %%ymm1, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off1])") \
__ASM_EMIT("vmovups %%ymm3, 0x00(%[" __IF_32_64("ptr2", "dst_im") "], %[off2])") \
__ASM_EMIT("add $0x20, %[off1]") \
__ASM_EMIT("add $0x20, %[off2]") \
__ASM_EMIT32("subl $8, %[np]") \
__ASM_EMIT64("subq $8, %[np]") \
__ASM_EMIT("jz 2f") \
/* Rotate angle */ \
__ASM_EMIT32("mov %[fft_w], %[ptr2]") \
__ASM_EMIT("vmovaps 0x00(%[" __IF_32_64("ptr2", "fft_w") "]), %%ymm4") /* xmm4 = w_re */ \
__ASM_EMIT("vmovaps 0x20(%[" __IF_32_64("ptr2", "fft_w") "]), %%ymm5") /* xmm5 = w_im */ \
__ASM_EMIT32("mov %[dst_im], %[ptr2]") \
__ASM_EMIT("vmulps %%ymm5, %%ymm6, %%ymm2") /* ymm2 = w_im * x_re */ \
__ASM_EMIT("vmulps %%ymm5, %%ymm7, %%ymm3") /* ymm3 = w_im * x_im */ \
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm6, %%ymm6", "")) /* ymm6 = w_re * x_re */ \
__ASM_EMIT(FMA_SEL("vmulps %%ymm4, %%ymm7, %%ymm7", "")) /* ymm7 = w_re * x_im */ \
__ASM_EMIT(FMA_SEL("vsubps %%ymm3, %%ymm6, %%ymm6", "vfmsub132ps %%ymm4, %%ymm3, %%ymm6")) /* ymm6 = x_re' = w_re * x_re - w_im * x_im */ \
__ASM_EMIT(FMA_SEL("vaddps %%ymm2, %%ymm7, %%ymm7", "vfmadd132ps %%ymm4, %%ymm2, %%ymm7")) /* ymm7 = x_im' = w_re * x_im + w_im * x_re */ \
/* Repeat loop */ \
__ASM_EMIT("jmp 1b") \
__ASM_EMIT("2:") \
\
: __IF_32([ptr1] "=&r" (ptr1), [ptr2] "=&r" (ptr2), ) \
[off1] "+r" (off1), [off2] "+r" (off2), \
[np] X86_PGREG (np) \
: [dst_re] X86_GREG (dst_re), [dst_im] X86_GREG (dst_im), [fft_a] X86_GREG (fft_a), [fft_w] X86_GREG (fft_w) \
: "cc", "memory", \
"%xmm0", "%xmm1", "%xmm2", "%xmm3", \
"%xmm4", "%xmm5", "%xmm6", "%xmm7" \
);
#define FMA_OFF(a, b) a
#define FMA_ON(a, b) b
static inline void butterfly_direct8p(float *dst_re, float *dst_im, size_t rank, size_t blocks)
{
size_t pairs = 1 << rank;
size_t off1 = 0, shift = 4 << rank; //1 << (rank + 2);
const float *fft_a = &FFT_A[(rank - 2) << 4];
const float *fft_w = &FFT_DW[(rank - 2) << 4];
for (size_t b=0; b