commit 4a7b70f2262b95276ee291b5455ae50bf1874e4f
author Marat Dukhan <maratek@google.com> Mon Aug 02 18:18:10 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Aug 02 18:19:05 2021 -0700
tree e52d929c45bc58594088e3560a5b052023182f6c
parent 79993410e8170efddd914a47d91e5544633c6a3c

QS8/QU8 VMUL[C] microkernels in NEON implementation

PiperOrigin-RevId: 388349072

src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c

diff --git a/src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c b/src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c
new file mode 100644
index 0000000..25e8845
--- /dev/null
+++ b/src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c
@@ -0,0 +1,144 @@
+// Auto-generated file. Do not edit!
+//   Template: src/qs8-vmulc/neon.c.in
+//   Generator: tools/xngen
+//
+// Copyright 2021 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.
+
+#include <assert.h>
+
+#include <arm_neon.h>
+
+#include <xnnpack/vadd.h>
+
+
+void xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16(
+    size_t n,
+    const uint8_t* input_a,
+    const uint8_t* input_b,
+    uint8_t* output,
+    const union xnn_qu8_mul_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+  const uint8x8_t va_zero_point = vld1_dup_u8(&params->fp32_neon.a_zero_point);
+  const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neon.scale);
+  const float32x4_t voutput_min_less_zero_point = vld1q_dup_f32(&params->fp32_neon.output_min_less_zero_point);
+  const float32x4_t voutput_max_less_zero_point = vld1q_dup_f32(&params->fp32_neon.output_max_less_zero_point);
+  const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
+  const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_zero_point);
+
+  const uint8x8_t vb = vld1_dup_u8(input_b);
+  const uint8x8_t vb_zero_point = vld1_dup_u8(&params->fp32_neon.b_zero_point);
+  const int16x8_t vxb = vreinterpretq_s16_u16(vsubl_u8(vb, vb_zero_point));
+  for (; n >= 16 * sizeof(uint8_t); n -= 16 * sizeof(uint8_t)) {
+    const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
+    const uint8x8_t va89ABCDEF = vld1_u8(input_a); input_a += 8;
+
+    const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
+    const int16x8_t vxa89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(va89ABCDEF, va_zero_point));
+
+    int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
+    int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
+    int32x4_t vacc89AB = vmull_s16(vget_low_s16(vxa89ABCDEF), vget_low_s16(vxb));
+    int32x4_t vaccCDEF = vmull_s16(vget_high_s16(vxa89ABCDEF), vget_high_s16(vxb));
+
+    float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
+    float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
+    float32x4_t vfpacc89AB = vcvtq_f32_s32(vacc89AB);
+    float32x4_t vfpaccCDEF = vcvtq_f32_s32(vaccCDEF);
+
+    vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
+    vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
+    vfpacc89AB = vmulq_f32(vfpacc89AB, vscale);
+    vfpaccCDEF = vmulq_f32(vfpaccCDEF, vscale);
+
+    vfpacc0123 = vmaxq_f32(vfpacc0123, voutput_min_less_zero_point);
+    vfpacc4567 = vmaxq_f32(vfpacc4567, voutput_min_less_zero_point);
+    vfpacc89AB = vmaxq_f32(vfpacc89AB, voutput_min_less_zero_point);
+    vfpaccCDEF = vmaxq_f32(vfpaccCDEF, voutput_min_less_zero_point);
+
+    vfpacc0123 = vminq_f32(vfpacc0123, voutput_max_less_zero_point);
+    vfpacc4567 = vminq_f32(vfpacc4567, voutput_max_less_zero_point);
+    vfpacc89AB = vminq_f32(vfpacc89AB, voutput_max_less_zero_point);
+    vfpaccCDEF = vminq_f32(vfpaccCDEF, voutput_max_less_zero_point);
+
+    vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
+    vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
+    vacc89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc89AB, vmagic_bias));
+    vaccCDEF = vreinterpretq_s32_f32(vaddq_f32(vfpaccCDEF, vmagic_bias));
+
+    vacc0123 = vsubq_s32(vacc0123, vmagic_bias_less_zero_point);
+    vacc4567 = vsubq_s32(vacc4567, vmagic_bias_less_zero_point);
+    vacc89AB = vsubq_s32(vacc89AB, vmagic_bias_less_zero_point);
+    vaccCDEF = vsubq_s32(vaccCDEF, vmagic_bias_less_zero_point);
+
+#if XNN_ARCH_ARM64
+      const int16x8_t vacc01234567 = vuzp1q_s16(vreinterpretq_s16_s32(vacc0123), vreinterpretq_s16_s32(vacc4567));
+      const int16x8_t vacc89ABCDEF = vuzp1q_s16(vreinterpretq_s16_s32(vacc89AB), vreinterpretq_s16_s32(vaccCDEF));
+
+      uint8x16_t vout0123456789ABCDEF = vuzp1q_s8(vreinterpretq_u8_s16(vacc01234567), vreinterpretq_u8_s16(vacc89ABCDEF));
+#else
+      const int16x8_t vacc01234567 = vcombine_s16(vmovn_s32(vacc0123), vmovn_s32(vacc4567));
+      const int16x8_t vacc89ABCDEF = vcombine_s16(vmovn_s32(vacc89AB), vmovn_s32(vaccCDEF));
+
+      uint8x16_t vout0123456789ABCDEF = vreinterpretq_u8_s8(vcombine_s8(vmovn_s16(vacc01234567), vmovn_s16(vacc89ABCDEF)));
+#endif
+
+    vst1q_u8(output, vout0123456789ABCDEF); output += 16;
+  }
+  if XNN_UNLIKELY(n != 0) {
+    do {
+      const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;
+
+      const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));
+
+      int32x4_t vacc0123 = vmull_s16(vget_low_s16(vxa01234567), vget_low_s16(vxb));
+      int32x4_t vacc4567 = vmull_s16(vget_high_s16(vxa01234567), vget_high_s16(vxb));
+
+      float32x4_t vfpacc0123 = vcvtq_f32_s32(vacc0123);
+      float32x4_t vfpacc4567 = vcvtq_f32_s32(vacc4567);
+
+      vfpacc0123 = vmulq_f32(vfpacc0123, vscale);
+      vfpacc4567 = vmulq_f32(vfpacc4567, vscale);
+
+      vfpacc0123 = vmaxq_f32(vfpacc0123, voutput_min_less_zero_point);
+      vfpacc4567 = vmaxq_f32(vfpacc4567, voutput_min_less_zero_point);
+
+      vfpacc0123 = vminq_f32(vfpacc0123, voutput_max_less_zero_point);
+      vfpacc4567 = vminq_f32(vfpacc4567, voutput_max_less_zero_point);
+
+      vacc0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0123, vmagic_bias));
+      vacc4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4567, vmagic_bias));
+
+      vacc0123 = vsubq_s32(vacc0123, vmagic_bias_less_zero_point);
+      vacc4567 = vsubq_s32(vacc4567, vmagic_bias_less_zero_point);
+
+#if XNN_ARCH_ARM64
+      const int16x8_t vacc01234567 = vuzp1q_s16(vreinterpretq_s16_s32(vacc0123), vreinterpretq_s16_s32(vacc4567));
+      uint8x8_t vout01234567 = vreinterpret_u8_s8(vmovn_s16(vacc01234567));
+#else
+      const int16x8_t vacc01234567 = vcombine_s16(vmovn_s32(vacc0123), vmovn_s32(vacc4567));
+      uint8x8_t vout01234567 = vreinterpret_u8_s8(vmovn_s16(vacc01234567));
+#endif
+
+      if XNN_LIKELY(n >= (8 * sizeof(uint8_t))) {
+        vst1_u8(output, vout01234567); output += 8;
+        n -= 8 * sizeof(uint8_t);
+      } else {
+        if (n & (4 * sizeof(uint8_t))) {
+          vst1_lane_u32(__builtin_assume_aligned(output, 1), vreinterpret_u32_u8(vout01234567), 0); output += 4;
+          vout01234567 = vext_u8(vout01234567, vout01234567, 4);
+        }
+        if (n & (2 * sizeof(uint8_t))) {
+          vst1_lane_u16(__builtin_assume_aligned(output, 1), vreinterpret_u16_u8(vout01234567), 0); output += 2;
+          vout01234567 = vext_u8(vout01234567, vout01234567, 2);
+        }
+        if (n & (1 * sizeof(uint8_t))) {
+          vst1_lane_u8(output, vout01234567, 0);
+        }
+        n = 0;
+      }
+    } while (n != 0);
+  }
+}