/external/XNNPACK/src/f32-dwconv/gen/ |
D | up8x3-minmax-neonfma-acc2.c | 71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local 103 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local 127 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma_acc2() local
|
D | up8x3-minmax-wasmsimd-arm-acc2.c | 74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local 115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local 143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_arm_acc2() local
|
D | up8x3-minmax-neon-acc2.c | 71 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local 103 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local 127 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__neon_acc2() local
|
D | up8x3-minmax-sse-acc2.c | 74 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local 114 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local 142 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__sse_acc2() local
|
D | up8x3-minmax-wasmsimd-x86-acc2.c | 74 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local 115 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local 143 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86_acc2() local
|
D | up8x4-minmax-wasmsimd-x86-acc2.c | 79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local 163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
|
D | up8x4-minmax-sse-acc2.c | 79 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 128 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local 162 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
|
D | up8x4-minmax-wasmsimd-arm-acc2.c | 79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local 163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
|
D | up8x4-minmax-neon-acc2.c | 76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local 143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
|
D | up8x4-minmax-neonfma-acc2.c | 76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local 143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
|
D | up16x3-minmax-neon-acc2.c | 83 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neon_acc2() local 131 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neon_acc2() local 155 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neon_acc2() local
|
D | up16x3-minmax-neonfma-acc2.c | 83 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma_acc2() local 131 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma_acc2() local 155 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x3__neonfma_acc2() local
|
D | up4x3-minmax-sse-acc2.c | 68 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__sse_acc2() local 96 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__sse_acc2() local
|
D | up4x3-minmax-wasmsimd-arm-acc2.c | 68 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__wasmsimd_arm_acc2() local 97 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__wasmsimd_arm_acc2() local
|
D | up4x3-minmax-wasmsimd-x86-acc2.c | 68 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__wasmsimd_x86_acc2() local 97 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__wasmsimd_x86_acc2() local
|
D | up4x3-minmax-neonfma-acc2.c | 65 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__neonfma_acc2() local 89 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__neonfma_acc2() local
|
D | up4x3-minmax-neon-acc2.c | 65 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__neon_acc2() local 89 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x3__neon_acc2() local
|
D | up4x4-minmax-wasmsimd-arm-acc2.c | 73 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local 108 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm_acc2() local
|
D | up4x4-minmax-sse-acc2.c | 73 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local 107 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__sse_acc2() local
|
D | up4x4-minmax-wasmsimd-x86-acc2.c | 73 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local 108 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_x86_acc2() local
|
D | up4x4-minmax-neon-acc2.c | 70 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local 98 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__neon_acc2() local
|
D | up4x4-minmax-neonfma-acc2.c | 70 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local 98 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x4__neonfma_acc2() local
|
D | up16x4-minmax-neon-acc2.c | 88 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neon_acc2() local 149 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neon_acc2() local 177 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neon_acc2() local
|
D | up16x4-minmax-neonfma-acc2.c | 88 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neonfma_acc2() local 149 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neonfma_acc2() local 177 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up16x4__neonfma_acc2() local
|
D | up8x9-minmax-wasmsimd-arm-acc2.c | 104 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 199 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local 263 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
|