/*
* Copyright (C) 2023 Linux Studio Plugins Project
* (C) 2023 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_AARCH64_ASIMD_PMATH_EXP_H_
#define PRIVATE_DSP_ARCH_AARCH64_ASIMD_PMATH_EXP_H_
#ifndef PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL
#error "This header should not be included directly"
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_IMPL */
namespace lsp
{
namespace asimd
{
IF_ARCH_AARCH64(
static const uint32_t EXP2_CONST[] __lsp_aligned16 =
{
LSP_DSP_VEC4(0x7f), // ME = 0x7f
LSP_DSP_VEC4(0x3f317218), // L2 = ln(2)
LSP_DSP_VEC4(0x40e00000), // C5 = 7!/6! = 7.0000000000000000
LSP_DSP_VEC4(0x42280000), // C4 = 7!/5! = 42.0000000000000000
LSP_DSP_VEC4(0x43520000), // C3 = 7!/4! = 210.0000000000000000
LSP_DSP_VEC4(0x44520000), // C2 = 7!/3! = 840.0000000000000000
LSP_DSP_VEC4(0x451d8000), // C1 = 7!/2! = 2520.0000000000000000
LSP_DSP_VEC4(0x459d8000), // C0 = 7!/1! = 5040.0000000000000000
LSP_DSP_VEC4(0x39500d01), // C6 = 1/7! = 0.0001984127011383
LSP_DSP_VEC4(0x3f800000), // C7 = 1.0
};
static const float EXP_LOG2E[] __lsp_aligned16 =
{
LSP_DSP_VEC4(M_LOG2E),
LSP_DSP_VEC4(M_LOG2E)
};
)
#define POW2_CORE_X8(ME, L2, C5, C4, C3, C2, C1, C0, C6, C7) \
/* v0 = x */ \
__ASM_EMIT("sshr v2.4s, v0.4s, #31") /* v2 = [ x < 0 ] */ \
__ASM_EMIT("sshr v3.4s, v1.4s, #31") \
__ASM_EMIT("fabs v0.4s, v0.4s") /* v0 = XP = fabs(x) */ \
__ASM_EMIT("fabs v1.4s, v1.4s") \
__ASM_EMIT("fcvtzs v4.4s, v0.4s") /* v4 = R = int(x) */ \
__ASM_EMIT("fcvtzs v5.4s, v1.4s") \
__ASM_EMIT("add v6.4s, v4.4s, " ME ".4s") /* v6 = R + (ME = 127) */ \
__ASM_EMIT("add v7.4s, v5.4s, " ME ".4s") \
__ASM_EMIT("scvtf v4.4s, v4.4s") /* v4 = float(R) */ \
__ASM_EMIT("scvtf v5.4s, v5.4s") \
__ASM_EMIT("shl v6.4s, v6.4s, #23") /* v6 = 1 << (R + 127) */ \
__ASM_EMIT("shl v7.4s, v7.4s, #23") \
__ASM_EMIT("fsub v0.4s, v0.4s, v4.4s") /* v0 = XP - float(R) */ \
__ASM_EMIT("fsub v1.4s, v1.4s, v5.4s") \
__ASM_EMIT("fmul v0.4s, v0.4s, " L2 ".4s") /* v0 = X = ln(2) * (XP - float(R)) */ \
__ASM_EMIT("fmul v1.4s, v1.4s, " L2 ".4s") \
/* v0 = X, v2 = [ x < 0 ], v6 = 1 << R */ \
__ASM_EMIT("fadd v4.4s, v0.4s, " C5 ".4s") /* v4 = C5+X */ \
__ASM_EMIT("fadd v5.4s, v1.4s, " C5 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C5+X) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C4 ".4s") /* v4 = C4+X*(C5+X) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C4 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C4+X*(C5+X)) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C3 ".4s") /* v4 = C3+X*(C4+X*(C5+X)) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C3 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C3+X*(C4+X*(C5+X))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C2 ".4s") /* v4 = C2+X*(C3+X*(C4+X*(C5+X))) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C2 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C2+X*(C3+X*(C4+X*(C5+X)))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C1 ".4s") /* v4 = C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C1 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C0 ".4s") /* v4 = C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C0 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v1.4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, " C6 ".4s") /* v4 = 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, " C6 ".4s") \
__ASM_EMIT("fadd v4.4s, v4.4s, " C7 ".4s") /* v4 = 1 + 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fadd v5.4s, v5.4s, " C7 ".4s") \
__ASM_EMIT("fmul v4.4s, v4.4s, v6.4s") /* v4 = E = (1 << R) * (1 + 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))))) */ \
__ASM_EMIT("fmul v5.4s, v5.4s, v7.4s") \
/* Calc reciprocals into v0 */ \
__ASM_EMIT("frecpe v8.4s, v4.4s") /* v8 = e2 */ \
__ASM_EMIT("frecpe v9.4s, v5.4s") \
__ASM_EMIT("frecps v10.4s, v8.4s, v4.4s") /* v10 = (2 - E*e2) */ \
__ASM_EMIT("frecps v11.4s, v9.4s, v5.4s") \
__ASM_EMIT("fmul v8.4s, v10.4s, v8.4s") /* v8 = e2' = e2 * (2 - E*e2) */ \
__ASM_EMIT("fmul v9.4s, v11.4s, v9.4s") \
__ASM_EMIT("frecps v10.4s, v8.4s, v4.4s") /* v10 = (2 - E*e2') */ \
__ASM_EMIT("frecps v11.4s, v9.4s, v5.4s") \
__ASM_EMIT("fmul v0.4s, v10.4s, v8.4s") /* v0 = 1/E = e2" = e2' * (2 - E*e2) */ \
__ASM_EMIT("fmul v1.4s, v11.4s, v9.4s") \
/* Perform conditional output */ \
__ASM_EMIT("bif v0.16b, v4.16b, v2.16b") /* v0 = ((1/E) & [ x < 0 ]) | (E & [x >= 0]) */ \
__ASM_EMIT("bif v1.16b, v5.16b, v3.16b")
#define POW2_CORE_X4(ME, L2, C5, C4, C3, C2, C1, C0, C6, C7) \
/* v0 = x */ \
__ASM_EMIT("sshr v2.4s, v0.4s, #31") /* v2 = [ x < 0 ] */ \
__ASM_EMIT("fabs v0.4s, v0.4s") /* v0 = XP = fabs(x) */ \
__ASM_EMIT("fcvtzs v4.4s, v0.4s") /* v4 = R = int(x) */ \
__ASM_EMIT("add v6.4s, v4.4s, " ME ".4s") /* v6 = R + (ME = 127) */ \
__ASM_EMIT("scvtf v4.4s, v4.4s") /* v4 = float(R) */ \
__ASM_EMIT("shl v6.4s, v6.4s, #23") /* v6 = 1 << (R + 127) */ \
__ASM_EMIT("fsub v0.4s, v0.4s, v4.4s") /* v0 = XP - float(R) */ \
__ASM_EMIT("fmul v0.4s, v0.4s, " L2 ".4s") /* v0 = X = ln(2) * (XP - float(R)) */ \
/* v0 = X, v2 = [ x < 0 ], v6 = 1 << R */ \
__ASM_EMIT("fadd v4.4s, v0.4s, " C5 ".4s") /* v4 = C5+X */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C5+X) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C4 ".4s") /* v4 = C4+X*(C5+X) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C4+X*(C5+X)) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C3 ".4s") /* v4 = C3+X*(C4+X*(C5+X)) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C3+X*(C4+X*(C5+X))) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C2 ".4s") /* v4 = C2+X*(C3+X*(C4+X*(C5+X))) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C2+X*(C3+X*(C4+X*(C5+X)))) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C1 ".4s") /* v4 = C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C0 ".4s") /* v4 = C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v0.4s") /* v4 = X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, " C6 ".4s") /* v4 = 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fadd v4.4s, v4.4s, " C7 ".4s") /* v4 = 1 + 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X)))))) */ \
__ASM_EMIT("fmul v4.4s, v4.4s, v6.4s") /* v4 = E = (1 << R) * (1 + 1/7! * X*(C0+X*(C1+X*(C2+X*(C3+X*(C4+X*(C5+X))))))) */ \
/* Calc reciprocals into v0 */ \
__ASM_EMIT("frecpe v8.4s, v4.4s") /* v8 = e2 */ \
__ASM_EMIT("frecps v10.4s, v8.4s, v4.4s") /* v10 = (2 - E*e2) */ \
__ASM_EMIT("fmul v8.4s, v10.4s, v8.4s") /* v8 = e2' = e2 * (2 - E*e2) */ \
__ASM_EMIT("frecps v10.4s, v8.4s, v4.4s") /* v10 = (2 - E*e2') */ \
__ASM_EMIT("fmul v0.4s, v10.4s, v8.4s") /* v0 = 1/E = e2" = e2' * (2 - E*e2) */ \
/* Perform conditional output */ \
__ASM_EMIT("bif v0.16b, v4.16b, v2.16b") /* v0 = ((1/E) & [ x < 0 ]) | (E & [x >= 0]) */
#define EXP_CORE_LOAD \
__ASM_EMIT("ldp q26, q27, [%[LOG2E]]") \
__ASM_EMIT("ldp q16, q17, [%[E2C], #0x00]") /* v16 = ME, v17 = L2 */ \
__ASM_EMIT("ldp q18, q19, [%[E2C], #0x20]") /* v18 = C5, v19 = C4 */ \
__ASM_EMIT("ldp q20, q21, [%[E2C], #0x40]") /* v20 = C3, v21 = C2 */ \
__ASM_EMIT("ldp q22, q23, [%[E2C], #0x60]") /* v22 = C1, v23 = C0 */ \
__ASM_EMIT("ldp q24, q25, [%[E2C], #0x80]") /* v22 = C6, v23 = C7 */
#define EXP_CORE_X8_NOLOAD \
/* in: v0 = x0, v1 = x1 */ \
__ASM_EMIT("fmul v0.4s, v0.4s, v26.4s") \
__ASM_EMIT("fmul v1.4s, v1.4s, v27.4s") \
POW2_CORE_X8("v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25") \
/* out: v0 = exp(x0), v1 = exp(x1) */
#define EXP_CORE_X4_NOLOAD \
/* in: v0 = x0 */ \
__ASM_EMIT("fmul v0.4s, v0.4s, v26.4s") \
POW2_CORE_X4("v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25") \
/* out: v0 = exp(x0) */
#define EXP_CORE_X8 \
/* in: v0 = x0, v1 = x1 */ \
EXP_CORE_LOAD \
EXP_CORE_X8_NOLOAD \
/* out: v0 = exp(x0), v1 = exp(x1) */
#define EXP_CORE_X4 \
/* in: v0 = x0 */ \
EXP_CORE_LOAD \
EXP_CORE_X4_NOLOAD \
/* out: v0 = exp(x0) */
void exp2(float *dst, const float *src, size_t count)
{
ARCH_AARCH64_ASM(
EXP_CORE_LOAD
// x8 blocks
__ASM_EMIT("subs %[count], %[count], #8")
__ASM_EMIT("b.lo 2f")
__ASM_EMIT("1:")
__ASM_EMIT("ldp q0, q1, [%[src]]")
EXP_CORE_X8_NOLOAD
__ASM_EMIT("subs %[count], %[count], #8")
__ASM_EMIT("stp q0, q1, [%[dst]]")
__ASM_EMIT("add %[src], %[src], #0x20")
__ASM_EMIT("add %[dst], %[dst], #0x20")
__ASM_EMIT("bhs 1b")
__ASM_EMIT("2:")
// x4 block
__ASM_EMIT("adds %[count], %[count], #4")
__ASM_EMIT("blt 4f")
__ASM_EMIT("ldr q0, [%[src]]")
EXP_CORE_X4_NOLOAD
__ASM_EMIT("sub %[count], %[count], #4")
__ASM_EMIT("str q0, [%[dst]]")
__ASM_EMIT("add %[src], %[src], #0x10")
__ASM_EMIT("add %[dst], %[dst], #0x10")
__ASM_EMIT("4:")
// Tail: 1x-3x block
__ASM_EMIT("adds %[count], %[count], #4")
__ASM_EMIT("bls 12f")
__ASM_EMIT("tst %[count], #1")
__ASM_EMIT("b.eq 6f")
__ASM_EMIT("ld1 {v0.s}[0], [%[src]]")
__ASM_EMIT("add %[src], %[src], #0x04")
__ASM_EMIT("6:")
__ASM_EMIT("tst %[count], #2")
__ASM_EMIT("b.eq 8f")
__ASM_EMIT("ld1 {v0.d}[1], [%[src]]")
__ASM_EMIT("8:")
EXP_CORE_X4_NOLOAD
__ASM_EMIT("tst %[count], #1")
__ASM_EMIT("b.eq 10f")
__ASM_EMIT("st1 {v0.s}[0], [%[dst]]")
__ASM_EMIT("add %[dst], %[dst], #0x04")
__ASM_EMIT("10:")
__ASM_EMIT("tst %[count], #2")
__ASM_EMIT("b.eq 12f")
__ASM_EMIT("st1 {v0.d}[1], [%[dst]]")
__ASM_EMIT("12:")
: [dst] "+r" (dst), [src] "+r" (src), [count] "+r" (count)
: [E2C] "r" (&EXP2_CONST[0]),
[LOG2E] "r" (&EXP_LOG2E[0])
: "cc", "memory",
"v0", "v1", "v2", "v3",
"v4", "v5", "v6", "v7",
"v8", "v9", "v10", "v11",
"v16", "v17", "v18", "v19",
"v20", "v21", "v22", "v23",
"v24", "v25", "v26", "v27"
);
}
void exp1(float *dst, size_t count)
{
ARCH_AARCH64_ASM(
EXP_CORE_LOAD
// x8 blocks
__ASM_EMIT("subs %[count], %[count], #8")
__ASM_EMIT("b.lo 2f")
__ASM_EMIT("1:")
__ASM_EMIT("ldp q0, q1, [%[dst]]")
EXP_CORE_X8
__ASM_EMIT("subs %[count], %[count], #8")
__ASM_EMIT("stp q0, q1, [%[dst]]")
__ASM_EMIT("add %[dst], %[dst], #0x20")
__ASM_EMIT("bhs 1b")
__ASM_EMIT("2:")
// x4 block
__ASM_EMIT("adds %[count], %[count], #4")
__ASM_EMIT("blt 4f")
__ASM_EMIT("ldr q0, [%[dst]]")
EXP_CORE_X4
__ASM_EMIT("sub %[count], %[count], #4")
__ASM_EMIT("str q0, [%[dst]]")
__ASM_EMIT("add %[dst], %[dst], #0x10")
__ASM_EMIT("4:")
// Tail: 1x-3x block
__ASM_EMIT("adds %[count], %[count], #4")
__ASM_EMIT("bls 12f")
__ASM_EMIT("tst %[count], #1")
__ASM_EMIT("b.eq 6f")
__ASM_EMIT("ld1 {v0.s}[0], [%[dst]]")
__ASM_EMIT("add %[dst], %[dst], #0x04")
__ASM_EMIT("6:")
__ASM_EMIT("tst %[count], #2")
__ASM_EMIT("b.eq 8f")
__ASM_EMIT("ld1 {v0.d}[1], [%[dst]]")
__ASM_EMIT("8:")
EXP_CORE_X4
__ASM_EMIT("tst %[count], #1")
__ASM_EMIT("b.eq 10f")
__ASM_EMIT("sub %[dst], %[dst], #0x04")
__ASM_EMIT("st1 {v0.s}[0], [%[dst]]")
__ASM_EMIT("add %[dst], %[dst], #0x04")
__ASM_EMIT("10:")
__ASM_EMIT("tst %[count], #2")
__ASM_EMIT("b.eq 12f")
__ASM_EMIT("st1 {v0.d}[1], [%[dst]]")
__ASM_EMIT("12:")
: [dst] "+r" (dst), [count] "+r" (count)
: [E2C] "r" (&EXP2_CONST[0]),
[LOG2E] "r" (&EXP_LOG2E[0])
: "cc", "memory",
"v0", "v1", "v2", "v3",
"v4", "v5", "v6", "v7",
"v8", "v9", "v10", "v11",
"v16", "v17", "v18", "v19",
"v20", "v21", "v22", "v23",
"v24", "v25", "v26", "v27"
);
}
} /* namespace asimd */
} /* namespace lsp */
#endif /* PRIVATE_DSP_ARCH_AARCH64_ASIMD_PMATH_EXP_H_ */