/external/XNNPACK/src/f32-hswish/gen/ |
D | hswish-wasm-x4.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__wasm_x4() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__wasm_x4() 40 float vacc0 = vx0 + vthree; in xnn_f32_hswish_ukernel__wasm_x4() 42 float vacc1 = vx1 + vthree; in xnn_f32_hswish_ukernel__wasm_x4() 44 float vacc2 = vx2 + vthree; in xnn_f32_hswish_ukernel__wasm_x4() 46 float vacc3 = vx3 + vthree; in xnn_f32_hswish_ukernel__wasm_x4() 73 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__wasm_x4()
|
D | hswish-scalar-x4.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__scalar_x4() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__scalar_x4() 40 float vacc0 = vx0 + vthree; in xnn_f32_hswish_ukernel__scalar_x4() 42 float vacc1 = vx1 + vthree; in xnn_f32_hswish_ukernel__scalar_x4() 44 float vacc2 = vx2 + vthree; in xnn_f32_hswish_ukernel__scalar_x4() 46 float vacc3 = vx3 + vthree; in xnn_f32_hswish_ukernel__scalar_x4() 73 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__scalar_x4()
|
D | hswish-wasm-x2.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__wasm_x2() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__wasm_x2() 38 float vacc0 = vx0 + vthree; in xnn_f32_hswish_ukernel__wasm_x2() 40 float vacc1 = vx1 + vthree; in xnn_f32_hswish_ukernel__wasm_x2() 58 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__wasm_x2()
|
D | hswish-scalar-x2.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__scalar_x2() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__scalar_x2() 38 float vacc0 = vx0 + vthree; in xnn_f32_hswish_ukernel__scalar_x2() 40 float vacc1 = vx1 + vthree; in xnn_f32_hswish_ukernel__scalar_x2() 58 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__scalar_x2()
|
D | hswish-wasmsimd-x16.c | 28 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three); in xnn_f32_hswish_ukernel__wasmsimd_x16() local 39 v128_t vacc0123 = wasm_f32x4_add(vx0123, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16() 41 v128_t vacc4567 = wasm_f32x4_add(vx4567, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16() 43 v128_t vacc89AB = wasm_f32x4_add(vx89AB, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16() 45 v128_t vaccCDEF = wasm_f32x4_add(vxCDEF, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16() 73 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16() 85 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x16()
|
D | hswish-neon-x16.c | 28 const float32x4_t vthree = vld1q_dup_f32(¶ms->scalar.three); in xnn_f32_hswish_ukernel__neon_x16() local 38 float32x4_t vacc0123 = vaddq_f32(vx0123, vthree); in xnn_f32_hswish_ukernel__neon_x16() 40 float32x4_t vacc4567 = vaddq_f32(vx4567, vthree); in xnn_f32_hswish_ukernel__neon_x16() 42 float32x4_t vacc89AB = vaddq_f32(vx89AB, vthree); in xnn_f32_hswish_ukernel__neon_x16() 44 float32x4_t vaccCDEF = vaddq_f32(vxCDEF, vthree); in xnn_f32_hswish_ukernel__neon_x16() 69 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x16() 78 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x16()
|
D | hswish-wasmsimd-x8.c | 28 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three); in xnn_f32_hswish_ukernel__wasmsimd_x8() local 37 v128_t vacc0123 = wasm_f32x4_add(vx0123, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x8() 39 v128_t vacc4567 = wasm_f32x4_add(vx4567, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x8() 59 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x8() 71 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x8()
|
D | hswish-neon-x8.c | 28 const float32x4_t vthree = vld1q_dup_f32(¶ms->scalar.three); in xnn_f32_hswish_ukernel__neon_x8() local 36 float32x4_t vacc0123 = vaddq_f32(vx0123, vthree); in xnn_f32_hswish_ukernel__neon_x8() 38 float32x4_t vacc4567 = vaddq_f32(vx4567, vthree); in xnn_f32_hswish_ukernel__neon_x8() 55 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x8() 64 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x8()
|
D | hswish-scalar-x1.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__scalar_x1() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__scalar_x1() 35 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__scalar_x1()
|
D | hswish-wasm-x1.c | 27 const float vthree = params->scalar.three; in xnn_f32_hswish_ukernel__wasm_x1() local 30 assert(vthree == 3.0f); in xnn_f32_hswish_ukernel__wasm_x1() 35 float vacc = vx + vthree; in xnn_f32_hswish_ukernel__wasm_x1()
|
D | hswish-wasmsimd-x4.c | 28 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three); in xnn_f32_hswish_ukernel__wasmsimd_x4() local 36 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x4() 48 v128_t vacc = wasm_f32x4_add(vx, vthree); in xnn_f32_hswish_ukernel__wasmsimd_x4()
|
D | hswish-neon-x4.c | 28 const float32x4_t vthree = vld1q_dup_f32(¶ms->scalar.three); in xnn_f32_hswish_ukernel__neon_x4() local 34 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x4() 43 float32x4_t vacc = vaddq_f32(vx, vthree); in xnn_f32_hswish_ukernel__neon_x4()
|
/external/XNNPACK/src/f32-hswish/ |
D | scalar.c.in | 27 const float vthree = params->scalar.three; 30 assert(vthree == 3.0f); 40 float vacc${ABC[N]} = vx${ABC[N]} + vthree; 60 float vacc = vx + vthree; 70 float vacc = vx + vthree; 80 float vacc = vx + vthree;
|
D | wasmsimd.c.in | 27 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three); variable 39 v128_t vacc${ABC[N:N+4]} = wasm_f32x4_add(vx${ABC[N:N+4]}, vthree); 60 v128_t vacc = wasm_f32x4_add(vx, vthree); 72 v128_t vacc = wasm_f32x4_add(vx, vthree);
|
D | neon.c.in | 27 const float32x4_t vthree = vld1q_dup_f32(¶ms->scalar.three); variable 37 float32x4_t vacc${ABC[N:N+4]} = vaddq_f32(vx${ABC[N:N+4]}, vthree); 54 float32x4_t vacc = vaddq_f32(vx, vthree); 63 float32x4_t vacc = vaddq_f32(vx, vthree);
|
/external/XNNPACK/src/f16-hswish/gen/ |
D | hswish-neonfp16arith-x16.c | 31 const float16x8_t vthree = vld1q_dup_f16(¶ms->three); in xnn_f16_hswish_ukernel__neonfp16arith_x16() local 39 float16x8_t vacc01234567 = vaddq_f16(vx01234567, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x16() 41 float16x8_t vacc89ABCDEF = vaddq_f16(vx89ABCDEF, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x16() 58 float16x8_t vacc = vaddq_f16(vx, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x16() 67 float16x8_t vacc = vaddq_f16(vx, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x16()
|
D | hswish-neonfp16arith-x8.c | 31 const float16x8_t vthree = vld1q_dup_f16(¶ms->three); in xnn_f16_hswish_ukernel__neonfp16arith_x8() local 37 float16x8_t vacc = vaddq_f16(vx, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x8() 46 float16x8_t vacc = vaddq_f16(vx, vthree); in xnn_f16_hswish_ukernel__neonfp16arith_x8()
|
/external/XNNPACK/src/f16-hswish/ |
D | neonfp16arith.c.in | 30 const float16x8_t vthree = vld1q_dup_f16(¶ms->three); variable 40 float16x8_t vacc${ABC[N:N+8]} = vaddq_f16(vx${ABC[N:N+8]}, vthree); 57 float16x8_t vacc = vaddq_f16(vx, vthree); 66 float16x8_t vacc = vaddq_f16(vx, vthree);
|