// Copyright 2020 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 CHANNEL_TILE % 4 == 0 $assert CHANNEL_TILE >= 4 $assert PIXEL_TILE == 1 $ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" #include #include #include "xnnpack/ibilinear.h" $ISA = "wasmrelaxedsimd" if FMA else "wasmsimd" void xnn_f32_ibilinear_ukernel__${ISA}_c${CHANNEL_TILE}${"" if PIXEL_TILE == 1 else "x%d" % PIXEL_TILE}( size_t output_pixels, size_t channels, const float** restrict input, size_t input_offset, const float* restrict weights, float* restrict output, size_t output_increment) XNN_OOB_READS { assert(output_pixels != 0); assert(channels != 0); assert(channels % sizeof(float) == 0); do { const float* i0 = (const float*) ((uintptr_t) input[0] + input_offset); const float* i1 = (const float*) ((uintptr_t) input[1] + input_offset); const float* i2 = (const float*) ((uintptr_t) input[2] + input_offset); const float* i3 = (const float*) ((uintptr_t) input[3] + input_offset); input += 4; const v128_t valphah = wasm_v128_load32_splat(weights); const v128_t valphav = wasm_v128_load32_splat(weights + 1); weights += 2; size_t c = channels; $if CHANNEL_TILE > 4: for (; c >= ${CHANNEL_TILE} * sizeof(float); c -= ${CHANNEL_TILE} * sizeof(float)) { const v128_t vtl${ABC[0:4]} = wasm_v128_load(i0); const v128_t vtr${ABC[0:4]} = wasm_v128_load(i1); const v128_t vbl${ABC[0:4]} = wasm_v128_load(i2); const v128_t vbr${ABC[0:4]} = wasm_v128_load(i3); $for C in range(4, CHANNEL_TILE, 4): const v128_t vtl${ABC[C:C+4]} = wasm_v128_load(i0 + ${C}); const v128_t vtr${ABC[C:C+4]} = wasm_v128_load(i1 + ${C}); const v128_t vbl${ABC[C:C+4]} = wasm_v128_load(i2 + ${C}); const v128_t vbr${ABC[C:C+4]} = wasm_v128_load(i3 + ${C}); i0 += ${CHANNEL_TILE}; i1 += ${CHANNEL_TILE}; i2 += ${CHANNEL_TILE}; i3 += ${CHANNEL_TILE}; $for C in range(0, CHANNEL_TILE, 4): const v128_t vtd${ABC[C:C+4]} = wasm_f32x4_sub(vtr${ABC[C:C+4]}, vtl${ABC[C:C+4]}); const v128_t vbd${ABC[C:C+4]} = wasm_f32x4_sub(vbr${ABC[C:C+4]}, vbl${ABC[C:C+4]}); $for C in range(0, CHANNEL_TILE, 4): $if FMA: const v128_t vt${ABC[C:C+4]} = wasm_f32x4_relaxed_madd(vtd${ABC[C:C+4]}, valphah, vtl${ABC[C:C+4]}); const v128_t vb${ABC[C:C+4]} = wasm_f32x4_relaxed_madd(vbd${ABC[C:C+4]}, valphah, vbl${ABC[C:C+4]}); $else: const v128_t vt${ABC[C:C+4]} = wasm_f32x4_add(wasm_f32x4_mul(vtd${ABC[C:C+4]}, valphah), vtl${ABC[C:C+4]}); const v128_t vb${ABC[C:C+4]} = wasm_f32x4_add(wasm_f32x4_mul(vbd${ABC[C:C+4]}, valphah), vbl${ABC[C:C+4]}); $for C in range(0, CHANNEL_TILE, 4): const v128_t vd${ABC[C:C+4]} = wasm_f32x4_sub(vb${ABC[C:C+4]}, vt${ABC[C:C+4]}); $for C in range(0, CHANNEL_TILE, 4): $if FMA: const v128_t vo${ABC[C:C+4]} = wasm_f32x4_relaxed_madd(vd${ABC[C:C+4]}, valphav, vt${ABC[C:C+4]}); $else: const v128_t vo${ABC[C:C+4]} = wasm_f32x4_add(wasm_f32x4_mul(vd${ABC[C:C+4]}, valphav), vt${ABC[C:C+4]}); wasm_v128_store(output, vo${ABC[0:4]}); $for C in range(4, CHANNEL_TILE, 4): wasm_v128_store(output + ${C}, vo${ABC[C:C+4]}); output += ${CHANNEL_TILE}; } for (; c >= 4 * sizeof(float); c -= 4 * sizeof(float)) { const v128_t vtl = wasm_v128_load(i0); const v128_t vtr = wasm_v128_load(i1); const v128_t vbl = wasm_v128_load(i2); const v128_t vbr = wasm_v128_load(i3); i0 += 4; i1 += 4; i2 += 4; i3 += 4; const v128_t vtd = wasm_f32x4_sub(vtr, vtl); const v128_t vbd = wasm_f32x4_sub(vbr, vbl); $if FMA: const v128_t vt = wasm_f32x4_relaxed_madd(vtd, valphah, vtl); const v128_t vb = wasm_f32x4_relaxed_madd(vbd, valphah, vbl); $else: const v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vtd, valphah), vtl); const v128_t vb = wasm_f32x4_add(wasm_f32x4_mul(vbd, valphah), vbl); const v128_t vd = wasm_f32x4_sub(vb, vt); $if FMA: const v128_t vo = wasm_f32x4_relaxed_madd(vd, valphav, vt); $else: const v128_t vo = wasm_f32x4_add(wasm_f32x4_mul(vd, valphav), vt); wasm_v128_store(output, vo); output += 4; } if XNN_UNLIKELY(c != 0) { const v128_t vtl = wasm_v128_load(i0); const v128_t vtr = wasm_v128_load(i1); const v128_t vbl = wasm_v128_load(i2); const v128_t vbr = wasm_v128_load(i3); const v128_t vtd = wasm_f32x4_sub(vtr, vtl); const v128_t vbd = wasm_f32x4_sub(vbr, vbl); $if FMA: const v128_t vt = wasm_f32x4_relaxed_madd(vtd, valphah, vtl); const v128_t vb = wasm_f32x4_relaxed_madd(vbd, valphah, vbl); $else: const v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vtd, valphah), vtl); const v128_t vb = wasm_f32x4_add(wasm_f32x4_mul(vbd, valphah), vbl); const v128_t vd = wasm_f32x4_sub(vb, vt); $if FMA: v128_t vo = wasm_f32x4_relaxed_madd(vd, valphav, vt); $else: v128_t vo = wasm_f32x4_add(wasm_f32x4_mul(vd, valphav), vt); if (c & (2 * sizeof(float))) { wasm_v128_store64_lane(output, vo, 0); vo = wasm_v64x2_shuffle(vo, vo, 1, 1); output += 2; } if (c & (1 * sizeof(float))) { wasm_v128_store32_lane(output, vo, 0); output += 1; } } output = (float*) ((uintptr_t) output + output_increment); } while (--output_pixels != 0); }