/* * 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_FASTCONV_PREPARE_H_ #define PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_PREPARE_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 FASTCONV_DIRECT_PREPARE_BODY(FMA_SEL) \ size_t off; \ \ ARCH_X86_ASM( \ __ASM_EMIT("lea (,%[np], 8), %[off]") \ __ASM_EMIT("vmovups 0x00(%[ak]), %%ymm6") /* ymm6 = x_re */ \ __ASM_EMIT("vmovups 0x20(%[ak]), %%ymm7") /* ymm7 = x_im */ \ __ASM_EMIT("vmovups 0x00(%[wk]), %%ymm4") /* ymm4 = w_re */ \ __ASM_EMIT("vmovups 0x20(%[wk]), %%ymm5") /* ymm5 = w_im */ \ __ASM_EMIT("vxorps %%ymm1, %%ymm1, %%ymm1") /* ymm1 = a_im = 0 */ \ /* x8 blocks */ \ __ASM_EMIT32("subl $8, %[np]") \ __ASM_EMIT64("sub $8, %[np]") \ __ASM_EMIT64("jb 2f") \ __ASM_EMIT("1:") \ __ASM_EMIT("vmovups 0x00(%[src]), %%ymm0") /* ymm0 = a_re = re */ \ __ASM_EMIT("vmulps %%ymm0, %%ymm7, %%ymm3") /* ymm3 = x_im * re */ \ __ASM_EMIT("vmulps %%ymm0, %%ymm6, %%ymm2") /* ymm2 = b_re = x_re * re */ \ __ASM_EMIT("vsubps %%ymm3, %%ymm1, %%ymm3") /* ymm3 = b_im = -x_im * re */ \ __ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") \ __ASM_EMIT("vmovups %%ymm1, 0x20(%[dst])") \ __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst], %[off])") \ __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst], %[off])") \ __ASM_EMIT("add $0x20, %[src]") \ __ASM_EMIT("add $0x40, %[dst]") \ __ASM_EMIT32("subl $8, %[np]") \ __ASM_EMIT64("sub $8, %[np]") \ __ASM_EMIT("jb 2f") \ /* Rotate angle */ \ __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 */ \ __ASM_EMIT("jmp 1b") \ __ASM_EMIT("2:") \ : [dst] "+r" (dst), [src] "+r" (src), \ [off] "=&r" (off), [np] __ASM_ARG_RW(np) \ : [ak] "r" (ak), [wk] "r" (wk) \ : "%xmm0", "%xmm1", "%xmm2", "%xmm3", \ "%xmm4", "%xmm5", "%xmm6", "%xmm7" \ ) #define FASTCONV_REVERSE_PREPARE_BODY(FMA_SEL) \ ARCH_X86_ASM( \ /* 2x blocks of 4x-butterfly loop */ \ __ASM_EMIT("sub $2, %[nb]") \ __ASM_EMIT("jb 2f") \ __ASM_EMIT("1:") \ __ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \ __ASM_EMIT("vmovups 0x10(%[dst]), %%xmm4") /* xmm4 = r4 r5 r6 r7 */ \ __ASM_EMIT("vmovups 0x20(%[dst]), %%xmm2") /* xmm2 = i0 i1 i2 i3 */ \ __ASM_EMIT("vmovups 0x30(%[dst]), %%xmm6") /* xmm6 = i4 i5 i6 i7 */ \ __ASM_EMIT("vinsertf128 $1, 0x40(%[dst]), %%ymm0, %%ymm0") /* ymm0 = r0 r1 r2 r3 r8 r9 r10 r11 */ \ __ASM_EMIT("vinsertf128 $1, 0x50(%[dst]), %%ymm4, %%ymm4") /* ymm4 = r4 r5 r6 r7 r12 r13 r14 r15 */ \ __ASM_EMIT("vinsertf128 $1, 0x60(%[dst]), %%ymm2, %%ymm2") /* ymm2 = i0 i1 i2 i3 i8 i9 i10 i11 */ \ __ASM_EMIT("vinsertf128 $1, 0x70(%[dst]), %%ymm6, %%ymm6") /* ymm6 = i4 i5 i6 i7 i12 i13 i14 i15 */ \ /* 1st-order 4x butterfly */ \ __ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ \ __ASM_EMIT("vhsubps %%ymm6, %%ymm2, %%ymm3") /* ymm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ \ __ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ \ __ASM_EMIT("vhaddps %%ymm6, %%ymm2, %%ymm2") /* ymm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ \ /* 2nd-order 4x butterfly */ \ __ASM_EMIT("vblendps $0xaa, %%ymm3, %%ymm1, %%ymm4") /* ymm4 = r1' i3' r5' i7' */ \ __ASM_EMIT("vblendps $0xaa, %%ymm1, %%ymm3, %%ymm5") /* ymm5 = i1' r3' i5' r7' */ \ __ASM_EMIT("vhsubps %%ymm4, %%ymm0, %%ymm1") /* ymm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ \ __ASM_EMIT("vhsubps %%ymm5, %%ymm2, %%ymm3") /* ymm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ \ __ASM_EMIT("vhaddps %%ymm4, %%ymm0, %%ymm0") /* ymm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ \ __ASM_EMIT("vhaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ \ __ASM_EMIT("vblendps $0xcc, %%ymm1, %%ymm0, %%ymm4") /* ymm4 = r0" i4" r1" r5" */ \ __ASM_EMIT("vblendps $0xcc, %%ymm0, %%ymm1, %%ymm5") /* ymm5 = r2" r6" r3" r7" */ \ __ASM_EMIT("vshufps $0x88, %%ymm3, %%ymm2, %%ymm6") /* ymm6 = i0" i1" i2" i3" */ \ __ASM_EMIT("vshufps $0xdd, %%ymm3, %%ymm2, %%ymm7") /* ymm7 = i4" i5" i6" i7" */ \ __ASM_EMIT("vshufps $0x88, %%ymm5, %%ymm4, %%ymm2") /* ymm2 = r0" r1" r2" r3" */ \ __ASM_EMIT("vshufps $0xdd, %%ymm5, %%ymm4, %%ymm3") /* ymm3 = r4" r5" r6" r7" */ \ /* 3rd-order 8x butterfly */ \ __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm3, %%ymm4") /* ymm4 = x_im * b_re */ \ __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%ymm7, %%ymm5") /* ymm5 = x_im * b_im */ \ __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%ymm3, %%ymm3", "")) /* ymm3 = x_re * b_re */ \ __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%ymm7, %%ymm7", "")) /* ymm7 = x_re * b_im */ \ __ASM_EMIT(FMA_SEL("vsubps %%ymm5, %%ymm3, %%ymm5", "vfmsub231ps 0x00 + %[FFT_A], %%ymm3, %%ymm5")) /* ymm5 = c_re = x_re * b_re - x_im * b_im */ \ __ASM_EMIT(FMA_SEL("vaddps %%ymm4, %%ymm7, %%ymm4", "vfmadd231ps 0x00 + %[FFT_A], %%ymm7, %%ymm4")) /* ymm4 = c_im = x_re * b_im + x_im * b_re */ \ __ASM_EMIT("vsubps %%ymm5, %%ymm2, %%ymm0") /* ymm0 = a_re - c_re */ \ __ASM_EMIT("vsubps %%ymm4, %%ymm6, %%ymm1") /* ymm1 = a_im - c_im */ \ __ASM_EMIT("vaddps %%ymm5, %%ymm2, %%ymm2") /* ymm2 = a_re + c_re */ \ __ASM_EMIT("vaddps %%ymm4, %%ymm6, %%ymm3") /* ymm3 = a_im + c_im */ \ /* Store */ \ __ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \ __ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \ __ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \ __ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \ __ASM_EMIT("vextractf128 $1, %%ymm2, 0x40(%[dst])") \ __ASM_EMIT("vextractf128 $1, %%ymm0, 0x50(%[dst])") \ __ASM_EMIT("vextractf128 $1, %%ymm3, 0x60(%[dst])") \ __ASM_EMIT("vextractf128 $1, %%ymm1, 0x70(%[dst])") \ __ASM_EMIT("add $0x80, %[dst]") \ __ASM_EMIT("sub $2, %[nb]") \ __ASM_EMIT("jae 1b") \ /* 1x block of 4-butterfly */ \ __ASM_EMIT("2:") \ __ASM_EMIT("add $1, %[nb]") \ __ASM_EMIT("jl 4f") \ __ASM_EMIT("vmovups 0x00(%[dst]), %%xmm0") /* xmm0 = r0 r1 r2 r3 */ \ __ASM_EMIT("vmovups 0x10(%[dst]), %%xmm4") /* xmm4 = r4 r5 r6 r7 */ \ __ASM_EMIT("vmovups 0x20(%[dst]), %%xmm2") /* xmm2 = i0 i1 i2 i3 */ \ __ASM_EMIT("vmovups 0x30(%[dst]), %%xmm6") /* xmm6 = i4 i5 i6 i7 */ \ /* 1st-order 4x butterfly */ \ __ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0-r1 r2-r3 r4-r5 r6-r7 = r1' r3' r5' r7' */ \ __ASM_EMIT("vhsubps %%xmm6, %%xmm2, %%xmm3") /* xmm3 = i0-i1 i2-i3 i4-i5 i6-i7 = i1' i3' i5' i7' */ \ __ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0+r1 r2+r3 r4+r5 r6+r7 = r0' r2' r4' r6' */ \ __ASM_EMIT("vhaddps %%xmm6, %%xmm2, %%xmm2") /* xmm2 = i0+i1 i2+i3 i4+i5 i6+i7 = i0' i2' i4' i6' */ \ /* 2nd-order 4x butterfly */ \ __ASM_EMIT("vblendps $0xaa, %%xmm3, %%xmm1, %%xmm4") /* xmm4 = r1' i3' r5' i7' */ \ __ASM_EMIT("vblendps $0xaa, %%xmm1, %%xmm3, %%xmm5") /* xmm5 = i1' r3' i5' r7' */ \ __ASM_EMIT("vhsubps %%xmm4, %%xmm0, %%xmm1") /* xmm1 = r0'-r2' r4'-r6' r1'-i3' r5'-i7' = r2" r6" r1" r5" */ \ __ASM_EMIT("vhsubps %%xmm5, %%xmm2, %%xmm3") /* xmm3 = i0'-i2' i4'-i6' i1'-r3' i5'-r7' = i2" i6" i3" i7" */ \ __ASM_EMIT("vhaddps %%xmm4, %%xmm0, %%xmm0") /* xmm0 = r0'+r2' r4'+r6' r1'+i3' r5'+i7' = r0" r4" r3" r7" */ \ __ASM_EMIT("vhaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = i0'+i2' i4'+i6' i1'+r3' i5'+r7' = i0" i4" i1" i5" */ \ __ASM_EMIT("vblendps $0xcc, %%xmm1, %%xmm0, %%xmm4") /* xmm4 = r0" i4" r1" r5" */ \ __ASM_EMIT("vblendps $0xcc, %%xmm0, %%xmm1, %%xmm5") /* xmm5 = r2" r6" r3" r7" */ \ __ASM_EMIT("vshufps $0x88, %%xmm3, %%xmm2, %%xmm6") /* xmm6 = i0" i1" i2" i3" */ \ __ASM_EMIT("vshufps $0xdd, %%xmm3, %%xmm2, %%xmm7") /* xmm7 = i4" i5" i6" i7" */ \ __ASM_EMIT("vshufps $0x88, %%xmm5, %%xmm4, %%xmm2") /* xmm2 = r0" r1" r2" r3" */ \ __ASM_EMIT("vshufps $0xdd, %%xmm5, %%xmm4, %%xmm3") /* xmm3 = r4" r5" r6" r7" */ \ /* 3rd-order 8x butterfly */ \ __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm3, %%xmm4") /* xmm4 = x_im * b_re */ \ __ASM_EMIT("vmulps 0x20 + %[FFT_A], %%xmm7, %%xmm5") /* xmm5 = x_im * b_im */ \ __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%xmm3, %%xmm3", "")) /* xmm3 = x_re * b_re */ \ __ASM_EMIT(FMA_SEL("vmulps 0x00 + %[FFT_A], %%xmm7, %%xmm7", "")) /* xmm7 = x_re * b_im */ \ __ASM_EMIT(FMA_SEL("vsubps %%xmm5, %%xmm3, %%xmm5", "vfmsub231ps 0x00 + %[FFT_A], %%xmm3, %%xmm5")) /* xmm5 = c_re = x_re * b_re - x_im * b_im */ \ __ASM_EMIT(FMA_SEL("vaddps %%xmm4, %%xmm7, %%xmm4", "vfmadd231ps 0x00 + %[FFT_A], %%xmm7, %%xmm4")) /* xmm4 = c_im = x_re * b_im + x_im * b_re */ \ __ASM_EMIT("vsubps %%xmm5, %%xmm2, %%xmm0") /* xmm0 = a_re - c_re */ \ __ASM_EMIT("vsubps %%xmm4, %%xmm6, %%xmm1") /* xmm1 = a_im - c_im */ \ __ASM_EMIT("vaddps %%xmm5, %%xmm2, %%xmm2") /* xmm2 = a_re + c_re */ \ __ASM_EMIT("vaddps %%xmm4, %%xmm6, %%xmm3") /* xmm3 = a_im + c_im */ \ /* Store */ \ __ASM_EMIT("vmovups %%xmm2, 0x00(%[dst])") \ __ASM_EMIT("vmovups %%xmm0, 0x10(%[dst])") \ __ASM_EMIT("vmovups %%xmm3, 0x20(%[dst])") \ __ASM_EMIT("vmovups %%xmm1, 0x30(%[dst])") \ __ASM_EMIT("4:") \ : [dst] "+r" (dst), [nb] "+r" (nb) \ : [FFT_A] "o" (FFT_A) \ : "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 fastconv_direct_prepare(float *dst, const float *src, const float *ak, const float *wk, size_t np) { FASTCONV_DIRECT_PREPARE_BODY(FMA_OFF); } static inline void fastconv_reverse_prepare(float *dst, size_t nb) { FASTCONV_REVERSE_PREPARE_BODY(FMA_OFF); } static inline void fastconv_direct_prepare_fma3(float *dst, const float *src, const float *ak, const float *wk, size_t np) { FASTCONV_DIRECT_PREPARE_BODY(FMA_OFF); } static inline void fastconv_reverse_prepare_fma3(float *dst, size_t nb) { FASTCONV_REVERSE_PREPARE_BODY(FMA_OFF); } static inline void fastconv_direct_unpack(float *dst, const float *src) { ARCH_X86_ASM( __ASM_EMIT("vmovups (%[src]), %%xmm0") __ASM_EMIT("vxorps %%ymm1, %%ymm1, %%ymm1") __ASM_EMIT("vmovups %%ymm0, 0x00(%[dst])") __ASM_EMIT("vmovups %%ymm1, 0x20(%[dst])") : : [dst] "r" (dst), [src] "r" (src) : "%xmm0", "%xmm1" ); } static inline void fastconv_reverse_unpack(float *dst, const float *src, size_t rank) { size_t blocks = 1 << rank; float norm = 1.0f / float(blocks); // Perform 4-element butterflies ARCH_X86_ASM ( __ASM_EMIT("vbroadcastss %[norm], %%ymm0") // 16x blocks __ASM_EMIT("sub $16, %[blocks]") __ASM_EMIT("jb 2f") __ASM_EMIT("vmovaps %%ymm0, %%ymm1") __ASM_EMIT("1:") __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */ __ASM_EMIT("vmulps 0x40(%[src]), %%ymm1, %%ymm3") /* ymm3 = r8 r9 r10 r11 r12 r13 r14 r15 */ __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])") __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst])") __ASM_EMIT("add $0x80, %[dst]") __ASM_EMIT("add $0x40, %[src]") __ASM_EMIT("sub $16, %[blocks]") __ASM_EMIT("jae 1b") __ASM_EMIT("2:") // 8x block __ASM_EMIT("add $8, %[blocks]") __ASM_EMIT("jl 4f") __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */ __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])") __ASM_EMIT("4:") : [dst] "+r"(dst), [src] "+r" (src), [blocks] "+r" (blocks) : [norm] "o" (norm) : "cc", "memory", "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7" ); } static inline void fastconv_reverse_unpack_adding(float *dst, const float *src, size_t rank) { size_t blocks = 1 << rank; float norm = 1.0f / float(blocks); // Perform 4-element butterflies ARCH_X86_ASM ( __ASM_EMIT("vbroadcastss %[norm], %%ymm0") // 16x blocks __ASM_EMIT("sub $16, %[blocks]") __ASM_EMIT("jb 2f") __ASM_EMIT("vmovaps %%ymm0, %%ymm1") __ASM_EMIT("1:") __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */ __ASM_EMIT("vmulps 0x40(%[src]), %%ymm1, %%ymm3") /* ymm3 = r8 r9 r10 r11 r12 r13 r14 r15 */ __ASM_EMIT("vaddps 0x00(%[dst]), %%ymm2, %%ymm2") __ASM_EMIT("vaddps 0x20(%[dst]), %%ymm3, %%ymm3") __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])") __ASM_EMIT("vmovups %%ymm3, 0x20(%[dst])") __ASM_EMIT("add $0x80, %[dst]") __ASM_EMIT("add $0x40, %[src]") __ASM_EMIT("sub $16, %[blocks]") __ASM_EMIT("jae 1b") __ASM_EMIT("2:") // 8x block __ASM_EMIT("add $8, %[blocks]") __ASM_EMIT("jl 4f") __ASM_EMIT("vmulps 0x00(%[src]), %%ymm0, %%ymm2") /* ymm2 = r0 r1 r2 r3 r4 r5 r6 r7 */ __ASM_EMIT("vaddps 0x00(%[dst]), %%ymm2, %%ymm2") __ASM_EMIT("vmovups %%ymm2, 0x00(%[dst])") __ASM_EMIT("4:") : [dst] "+r"(dst), [src] "+r" (src), [blocks] "+r" (blocks) : [norm] "o" (norm) : "cc", "memory", "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7" ); } #undef FASTCONV_DIRECT_PREPARE_BODY #undef FASTCONV_REVERSE_PREPARE_BODY #undef FMA_OFF #undef FMA_ON } } #endif /* PRIVATE_DSP_ARCH_X86_AVX_FASTCONV_PREPARE_H_ */