| // Auto-generated file. Do not edit! |
| // Template: src/qs8-gemm/c2-neon-mull-dup.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/gemm.h> |
| #include <xnnpack/math.h> |
| |
| void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup( |
| size_t mr, |
| size_t nc, |
| size_t kc, |
| const int8_t* restrict a, |
| size_t a_stride, |
| const void* restrict w, |
| int8_t* restrict c, |
| size_t cm_stride, |
| size_t cn_stride, |
| const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN |
| { |
| assert(mr != 0); |
| assert(mr <= 1); |
| assert(nc != 0); |
| assert(kc != 0); |
| assert(kc % sizeof(int8_t) == 0); |
| assert(a != NULL); |
| assert(w != NULL); |
| assert(c != NULL); |
| |
| kc = round_up_po2(kc, 2 * sizeof(int8_t)); |
| const int8_t* a0 = a; |
| int8_t* c0 = c; |
| |
| do { |
| int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t)); |
| |
| size_t k = kc; |
| |
| while (k >= 16 * sizeof(int8_t)) { |
| const int8x8_t va0x0 = vld1_s8(a0); a0 += 8; |
| const int8x8_t va0x1 = vld1_s8(a0); a0 += 8; |
| const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| |
| const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)); |
| const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)); |
| |
| int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0); |
| const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0); |
| const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)); |
| const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)); |
| |
| int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0); |
| const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); |
| const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)); |
| const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)); |
| |
| int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0); |
| const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0); |
| const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)); |
| const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)); |
| |
| int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0); |
| const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); |
| int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0); |
| const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); |
| |
| k -= 16 * sizeof(int8_t); |
| } |
| |
| if (k >= 8 * sizeof(int8_t)) { |
| const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| |
| const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| |
| const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)); |
| |
| const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)); |
| |
| const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)); |
| |
| const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)); |
| |
| const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3); |
| const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3); |
| |
| k -= 8 * sizeof(int8_t); |
| } |
| |
| if XNN_UNLIKELY(k != 0) { |
| const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| |
| const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| |
| const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)); |
| const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0); |
| const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0); |
| |
| if (k > 2 * sizeof(int8_t)) { |
| const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| |
| const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)); |
| const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1); |
| const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); |
| |
| if (k > 4 * sizeof(int8_t)) { |
| const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t)); |
| |
| const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)); |
| const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2); |
| vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2); |
| const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2); |
| vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2); |
| } |
| } |
| } |
| |
| float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123); |
| float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567); |
| |
| const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4); |
| vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123); |
| const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4); |
| vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567); |
| |
| const float32x4_t voutput_min_less_zero_point = vld1q_dup_f32(¶ms->neon_fp32.output_min_less_zero_point); |
| vfpacc0x0123 = vmaxq_f32(vfpacc0x0123, voutput_min_less_zero_point); |
| vfpacc0x4567 = vmaxq_f32(vfpacc0x4567, voutput_min_less_zero_point); |
| |
| const float32x4_t voutput_max_less_zero_point = vld1q_dup_f32(¶ms->neon_fp32.output_max_less_zero_point); |
| vfpacc0x0123 = vminq_f32(vfpacc0x0123, voutput_max_less_zero_point); |
| vfpacc0x4567 = vminq_f32(vfpacc0x4567, voutput_max_less_zero_point); |
| |
| const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neon_fp32.magic_bias); |
| vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias)); |
| vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias)); |
| |
| const int32x4_t vmagic_bias_less_zero_point = vld1q_dup_s32(¶ms->neon_fp32.magic_bias_less_zero_point); |
| vacc0x0123 = vsubq_s32(vacc0x0123, vmagic_bias_less_zero_point); |
| vacc0x4567 = vsubq_s32(vacc0x4567, vmagic_bias_less_zero_point); |
| |
| #if XNN_ARCH_ARM64 |
| const int16x8_t vacc0x01234567 = vuzp1q_s16(vreinterpretq_s16_s32(vacc0x0123), vreinterpretq_s16_s32(vacc0x4567)); |
| |
| int8x8_t vout0x01234567 = vmovn_s16(vacc0x01234567); |
| #else |
| const int16x8_t vacc0x01234567 = vcombine_s16(vmovn_s32(vacc0x0123), vmovn_s32(vacc0x4567)); |
| |
| int8x8_t vout0x01234567 = vmovn_s16(vacc0x01234567); |
| #endif |
| |
| if (nc >= 8) { |
| vst1_s8(c0 + 0, vout0x01234567); |
| |
| c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| |
| a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| |
| nc -= 8; |
| } else { |
| // Final case where not all of the 8 columns fit in the destination. |
| if (nc & 4) { |
| vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4; |
| vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4); |
| } |
| if (nc & 2) { |
| vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2; |
| vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2); |
| } |
| if (nc & 1) { |
| vst1_lane_s8(c0, vout0x01234567, 0); |
| } |
| |
| nc = 0; |
| } |
| } while (nc != 0); |
| } |