| // Copyright 2019 Google LLC |
| // |
| // This source code is licensed under the BSD-style license found in the |
| // LICENSE file in the root directory of this source tree. |
| |
| $assert BATCH_TILE % 4 == 0 |
| $assert BATCH_TILE >= 4 |
| $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" |
| $assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB"] |
| #include <assert.h> |
| |
| #include <psimd.h> |
| |
| #include <xnnpack/common.h> |
| #include <xnnpack/vbinary.h> |
| |
| |
| $PSIMD_OP_F32 = { |
| $ "ADD": lambda x, y: "psimd_add_f32(%s, %s)" % (x, y), |
| $ "DIV": lambda x, y: "psimd_div_f32(%s, %s)" % (x, y), |
| $ "MAX": lambda x, y: "psimd_max_f32(%s, %s)" % (x, y), |
| $ "MIN": lambda x, y: "psimd_min_f32(%s, %s)" % (x, y), |
| $ "MUL": lambda x, y: "psimd_mul_f32(%s, %s)" % (x, y), |
| $ "SUB": lambda x, y: "psimd_sub_f32(%s, %s)" % (x, y), |
| $}[OP] |
| void xnn_f32_v${OP.lower()}_ukernel__psimd_x${BATCH_TILE}( |
| size_t n, |
| const float* a, |
| const float* b, |
| float* y, |
| const union xnn_f32_minmax_params params[restrict static 1]) |
| { |
| assert(n != 0); |
| assert(n % sizeof(float) == 0); |
| |
| const psimd_f32 vy_min = psimd_load_splat_f32(¶ms->scalar.min); |
| const psimd_f32 vy_max = psimd_load_splat_f32(¶ms->scalar.max); |
| |
| for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { |
| const psimd_f32 va${ABC[0:4]} = psimd_load_f32(a); |
| $for N in range(4, BATCH_TILE, 4): |
| const psimd_f32 va${ABC[N:N+4]} = psimd_load_f32(a + ${N}); |
| a += ${BATCH_TILE}; |
| |
| const psimd_f32 vb${ABC[0:4]} = psimd_load_f32(b); |
| $for N in range(4, BATCH_TILE, 4): |
| const psimd_f32 vb${ABC[N:N+4]} = psimd_load_f32(b + ${N}); |
| b += ${BATCH_TILE}; |
| |
| $for N in range(0, BATCH_TILE, 4): |
| psimd_f32 vy${ABC[N:N+4]} = ${PSIMD_OP_F32("va" + ABC[N:N+4], "vb" + ABC[N:N+4])}; |
| |
| $for N in range(0, BATCH_TILE, 4): |
| vy${ABC[N:N+4]} = psimd_max_f32(vy${ABC[N:N+4]}, vy_min); |
| |
| $for N in range(0, BATCH_TILE, 4): |
| vy${ABC[N:N+4]} = psimd_min_f32(vy${ABC[N:N+4]}, vy_max); |
| |
| psimd_store_f32(y, vy${ABC[0:4]}); |
| $for N in range(4, BATCH_TILE, 4): |
| psimd_store_f32(y + ${N}, vy${ABC[N:N+4]}); |
| y += ${BATCH_TILE}; |
| } |
| $if BATCH_TILE >= 4: |
| for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) { |
| const psimd_f32 va0123 = psimd_load_f32(a); |
| a += 4; |
| |
| const psimd_f32 vb0123 = psimd_load_f32(b); |
| b += 4; |
| |
| psimd_f32 vy0123 = ${PSIMD_OP_F32("va0123", "vb0123")}; |
| vy0123 = psimd_max_f32(vy0123, vy_min); |
| vy0123 = psimd_min_f32(vy0123, vy_max); |
| psimd_store_f32(y, vy0123); |
| y += 4; |
| } |
| if XNN_UNLIKELY(n != 0) { |
| const psimd_f32 va0123 = psimd_load_f32(a); |
| const psimd_f32 vb0123 = psimd_load_f32(b); |
| |
| psimd_f32 vy0123 = ${PSIMD_OP_F32("va0123", "vb0123")}; |
| vy0123 = psimd_max_f32(vy0123, vy_min); |
| vy0123 = psimd_min_f32(vy0123, vy_max); |
| if (n & (2 * sizeof(float))) { |
| psimd_store2_f32(y, vy0123); |
| vy0123 = psimd_concat_hi_f32(vy0123, vy0123); |
| y += 2; |
| } |
| if (n & (1 * sizeof(float))) { |
| psimd_store1_f32(y, vy0123); |
| } |
| } |
| } |