Home
last modified time | relevance | path

Searched defs:vo4p0 (Results 1 – 25 of 27) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3p1-minmax-scalar-5x1.c111 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
197 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
D3x3p1-minmax-scalar-6x1.c120 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
220 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
D3x3p1-minmax-wasmsimd-arm-splat-5x4.c104 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
226 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
D3x3p1-minmax-wasmsimd-x86-splat-5x4.c104 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
226 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
D3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c137 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
252 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
D3x3p1-minmax-ssse3-5x4.c130 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
246 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
D3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c137 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
252 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
D3x3p1-minmax-neon-5x4.c101 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
224 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
D3x3p1-minmax-neonfma-5x4.c101 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
224 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
D3x3p1-minmax-wasmsimd-arm-splat-6x4.c112 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
252 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
D3x3p1-minmax-neon-6x4.c109 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
250 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
D3x3p1-minmax-neonfma-6x4.c109 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
250 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
D3x3p1-minmax-wasmsimd-x86-splat-6x4.c112 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
252 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
D3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c148 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
281 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
D3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c148 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
281 v128_t vo4p0 = wasm_f32x4_add(vbias, wasm_f32x4_mul(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
D3x3p1-minmax-ssse3-6x4.c141 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
275 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
D5x5p2-minmax-wasmsimd-x86-splat-5x4.c117 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
361 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
618 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
D5x5p2-minmax-wasmsimd-arm-splat-5x4.c117 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
361 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
618 v128_t vo4p0 = wasm_v32x4_shuffle(vw0123, vw0123, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
D3x3p1-minmax-sse-5x4.c159 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
319 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
D5x5p2-minmax-neonfma-5x4.c114 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
359 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
617 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
D5x5p2-minmax-neon-5x4.c114 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
359 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
617 float32x4_t vo4p0 = vdupq_lane_f32(vget_low_f32(vw0123), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
D5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c143 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
387 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
644 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
D5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c143 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
387 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
644 v128_t vo4p0 = vbias; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
D3x3p1-minmax-sse-6x4.c174 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
358 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk01)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
D5x5p2-minmax-sse-5x4.c142 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
395 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
661 __m128 vo4p0 = _mm_add_ps(vbias, _mm_mul_ps(vi4x4567, vk02)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local

12