[QNNPACK, Sparsity] Add dyanmic linear sparse kernel for arm64 (#50591)

Summary: Pull Request resolved: #50591 Adds sparse kernels for arm64. Reg blocking factor of 8x8. Test Plan: q8gemm-sparse-test Imported from OSS Reviewed By: AshkanAliabadi Differential Revision: D25925501 fbshipit-source-id: 8d62a8eb638f172ffaadfb1480ade0db35831189
pytorch · Feb 5, 2021 · a7ba051 · a7ba051
1 parent 70830b5
commit a7ba051
Show file tree

Hide file tree

Showing 9 changed files with 1,264 additions and 894 deletions.
diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
@@ -225,7 +225,9 @@ set(PYTORCH_QNNPACK_AARCH32_ASM_UKERNELS
 set(PYTORCH_QNNPACK_AARCH64_ASM_UKERNELS
   src/q8conv/8x8-aarch64-neon.S
   src/q8gemm/8x8-aarch64-neon.S
-  src/q8gemm/8x8-dq-aarch64-neon.S)
+  src/q8gemm/8x8-dq-aarch64-neon.S
+  src/q8gemm_sparse/8x4-packA-aarch64-neon.S
+  src/q8gemm_sparse/8x8c1x4-dq-packedA-aarch64-neon.S)
 
 set(PYTORCH_QNNPACK_X86_SSE2_UKERNELS
   src/q8avgpool/mp8x9p8q-sse2.c

diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/bench/q8gemm_sparse.cc b/aten/src/ATen/native/quantized/cpu/qnnpack/bench/q8gemm_sparse.cc
@@ -562,6 +562,51 @@ BENCHMARK_TEMPLATE_DEFINE_F(Q8GEMM_Op, 8x8__aarch64_neon, 8, 8, 8, 1)
 
 BENCHMARK_REGISTER_F(Q8GEMM_Op, 8x8__aarch64_neon)
     ->Apply(SparseGEMMBenchGemmArguments);
+
+BENCHMARK_TEMPLATE_DEFINE_F(Q8GEMMSparse_Op, 8x8c1x4_prepacked__aarch64_neon, 8, 8, 4)
+(benchmark::State& state) {
+  for (auto _ : state) {
+    auto m_blocks = (mc() + mr()  - 1) / mr();
+    auto k_blocks = (kc() + 4  - 1) / 4;
+    std::vector<uint8_t> a_packed(m_blocks * k_blocks * mr() * 4 + 8);
+    for (uint32_t m = 0; m < mc(); m += mr()) {
+      const uint32_t mrr = min(mc() - m, mr());
+      for (uint32_t n = 0, channel_offset = 0; n < nc();
+          n += nr(), channel_offset += nr()) {
+        const uint32_t nrr = min(nc() - n, nr());
+        pytorch_q8gemm_sparse_packA_ukernel_8x4__aarch64_neon(
+            mrr,
+            kc(),
+            a() + m * kc(),
+            kc() * sizeof(uint8_t),
+            a_packed.data() + (m >> 3) * (k_blocks << 2) * mr()
+            );
+      }
+    }
+    for (uint32_t m = 0; m < mc(); m += mr()) {
+      const uint32_t mrr = min(mc() - m, mr());
+      for (uint32_t n = 0, channel_offset = 0; n < nc();
+          n += nr(), channel_offset += nr()) {
+        const uint32_t nrr = min(nc() - n, nr());
+        pytorch_q8gemm_dq_sparse_1x4_ukernel_8x8_packedA__aarch64_neon(
+            mrr,
+            nrr,
+            a_packed.data() + (m >> 3) * (k_blocks << 2) * mr(),
+            bcsr_matrix_->values.data(),
+            bcsr_matrix_->row_values.data(),
+            bcsr_matrix_->col_indices.data(),
+            b() + n,
+            c() + m * nc() + n,
+            nc(),
+            channel_offset,
+            quantizationParams());
+      }
+    }
+  }
+}
+BENCHMARK_REGISTER_F(Q8GEMMSparse_Op, 8x8c1x4_prepacked__aarch64_neon)
+    ->Apply(SparseGEMMBenchGemmArguments);
+
 #endif
 
 #ifndef PYTORCH_QNNPACK_BENCHMARK_NO_MAIN

diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/src/init.c b/aten/src/ATen/native/quantized/cpu/qnnpack/src/init.c
@@ -164,6 +164,15 @@ static void init(void) {
   pytorch_qnnp_params.u8lut32norm = pytorch_u8lut32norm_ukernel__scalar;
   pytorch_qnnp_params.x8lut = pytorch_x8lut_ukernel__scalar;
 #elif CPUINFO_ARCH_ARM64
+  pytorch_qnnp_params.q8gemm_sparse = (struct pytorch_q8gemm_sparse_parameters){
+      .gemm_dq = NULL,
+      .packedA_gemm_dq = pytorch_q8gemm_dq_sparse_1x4_ukernel_8x8_packedA__aarch64_neon,
+      .packA = pytorch_q8gemm_sparse_packA_ukernel_8x4__aarch64_neon,
+      .mr = 8,
+      .nr = 8,
+      .kr = 4,
+      .log2_mr = 3,
+  };
   pytorch_qnnp_params.q8conv = (struct pytorch_q8conv_parameters){
       .gemm = pytorch_q8gemm_ukernel_8x8__aarch64_neon,
       .conv = pytorch_q8conv_ukernel_8x8__aarch64_neon,

diff --git a/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8gemm_sparse/8x4-packA-aarch64-neon.S b/aten/src/ATen/native/quantized/cpu/qnnpack/src/q8gemm_sparse/8x4-packA-aarch64-neon.S
@@ -0,0 +1,243 @@
+/*
+ * Copyright (c) Facebook, Inc. and its affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <qnnpack/assembly.h>
+
+# Packed A format.
+# 8kx4m blocks for alls blocks given 4 rows (4m) are placed in contiguous memory.
+# Original A
+# --------- K -----------          -- (K + 4 - 1) / 4 --
+# |                     |          |                   |
+# |                     |        (M + 8 - 1)/8         |
+# |                     | Packed   |                   |
+# M                     |  =>      |-------------------|
+# |                     |        Thus Packed A has (K + 4 - 1)/4 * (M + 8 -1)/8 blocks
+# |                     |
+# |---------------------|
+# 
+# Each 8 x 4 blocks is transposed and stored.
+# Each of the (K + 4 - 1)/4 blocks for a given group of 8 m blocks
+# are stored adjacent in memory
+# Thus, each block:
+# |----8m-----|----8m-----|
+# 4k          |           | ..... (K + 4 - 1)/4 blocks
+# |-----------|-----------|
+# This locality helps in loading 8kx8m blocks of activations
+# Note when M is not multiple of 8, the rest can contain arbitrary
+# data in packed A as we will not be writing those out.
+# This wil be taken care by just copying the appropriate valid data
+
+# void pytorch_q8gemm_sparse_packA_ukernel_8x4__aarch32_neon(
+#     size_t mr,
+#     size_t K,
+#     const uint8_t* a,
+#     size_t a_stride,
+#     uint8_t* packed_a,
+BEGIN_FUNCTION pytorch_q8gemm_sparse_packA_ukernel_8x4__aarch64_neon
+
+    # x2 = a0 = a pointer
+    # x4 = packed_a pointer
+
+    CMP x0, 2
+    # x5 = a1
+    ADD x5, x2, x3
+    CSEL x5, x2, x5, LO
+
+    # x6 = a2
+    ADD x6, x5, x3
+    CSEL x6, x5, x6, LS
+
+    CMP x0, 4
+    # x7 = a3
+    ADD x7, x6, x3
+    CSEL x7, x6, x7, LO
+
+    # x8 = a4
+    ADD x8, x7, x3
+    CSEL x8, x7, x8, LS
+
+    CMP x0, 6
+    # x9 = a5
+    ADD x9, x8, x3
+    CSEL x9, x8, x9, LO
+
+    # x10 = a6
+    ADD x10, x9, x3
+    CSEL x10, x9, x10, LS
+
+    CMP x0, 8
+    # x11 = a7
+    ADD x11, x10, x3
+    CSEL x11, x10, x11, NE
+
+    # num_k_blocks = (k + (4 - 1)) / 4
+    ADD x1, x1, 3
+    LSR x1, x1, 2
+
+    SUBS x1, x1, 2
+    B.LO 1f
+
+    .p2align 5
+k_loop:
+    LD1 {v0.d}[0], [x2], 8
+    LD1 {v0.d}[1], [x8], 8
+    LD1 {v1.d}[0], [x5], 8
+    LD1 {v1.d}[1], [x9], 8
+    LD1 {v2.d}[0], [x6], 8
+    LD1 {v2.d}[1], [x10], 8
+    LD1 {v3.d}[0], [x7], 8
+    LD1 {v3.d}[1], [x11], 8
+
+    #  Now we have 8x8 block of values that we will tranpose
+    #  A matrix
+    #  ------------------------
+    #  |                      |
+    #  |a0-----a3a4-----a7....|
+    #  |b0 B00 b3b4 B01 b7....|
+    #  |c0     c3c4     c7....|
+    #  |d0-----d3d4-----d7....|
+    #  |e0-----e3e4-----e7....|
+    #  |f0 B10 f3f4 B11 f7....|
+    #  |g0     g3g4     g7....|
+    #  |h0-----h3h4-----h7....|
+    #  |                      |
+    #  |                      |
+    #  ------------------------
+    #  {v0.2d[1], v0.2d[0]} = B00[0]+ B01[0] + B10[0] + B11[0]
+    #  {v1.2d[1], v1.2d[0]} = B00[1]+ B01[1] + B10[1] + B11[1]
+    #  {v2.2d[1], v2.2d[0]} = B00[2]+ B01[2] + B10[2] + B11[2]
+    #  {v3.2d[1], v3.2d[0]} = B00[3]+ B01[3] + B10[3] + B11[3]
+    #  v0 = e7 e6 e5 e4 e3 e2 e1 e0; a7 a6 a5 a4 a3 a2 a1 a0
+    #  v1 = f7 f6 f5 f4 f3 f2 f1 f0; b7 b6 b5 b4 b3 b2 b1 b0
+    #  v2 = g7 g6 g5 g4 g3 g2 g1 g0; c7 c6 c5 c4 c3 c2 c1 c0
+    #  v3 = h7 h6 h5 h4 h3 h2 h1 h0; d7 d6 d5 d4 d3 d2 d1 d0
+    #  Sequence:
+    #  TRN1 v4.16b, v0.16b, v1.16b
+    #  TRN2 v5.16b, v0.16b, v1.16b
+    #  TRN1 v6.16b, v2.16b, v3.16b
+    #  TRN2 v7.16b, v2.16b, v3.16b
+    #  Now we have
+    #  v4 = f6 e6 f4 e4 f2 e2 f0 e0; b6 a6 b4 a4 b2 a2 b0 a0
+    #  v5 = f7 e7 f5 e5 f3 e3 f1 e1; b7 a7 b5 a5 b3 a3 b1 a1
+    #  v6 = h6 g6 h4 g4 h2 g2 h0 g0; d6 c6 d4 c4 d2 c2 d0 c0
+    #  v7 = h7 g7 h5 g5 h3 g3 h1 g1; d7 c7 d5 c5 d3 c3 d1 c1
+    #  TRN1 v0.8h, v4.8h, v6.8h
+    #  TRN2 v2.8h, v4.8h, v6.8h
+    #  TRN1 v1.8h, v5.8h, v7.8h
+    #  TRN2 v3.8h, v5.8h, v7.8h
+    #  v0 = h4 g4 f4 e4 h0 g0 f0 e0; d4 c4 b4 a4 d0 c0 b0 a0
+    #  v1 = h5 g5 f5 e5 h1 g1 f1 e1; d5 c5 b5 a5 d1 c1 b1 a1
+    #  v2 = h6 g6 f6 e6 h2 g2 f2 e2; d6 c6 b6 a6 d2 c2 b2 a2
+    #  v3 = h7 g7 f7 e7 h3 g3 f3 e3; d7 c7 b7 a7 d3 c3 b3 a3
+    #  UZP1 v4.4s, v0.4s, v1.4s
+    #  UZP2 v6.4s, v0.4s, v1.4s
+    #  UZP1 v5.4s, v2.4s, v3.4s
+    #  UZP2 v7.4s, v2.4s, v3.4s
+    #  v4 = h1 g1 f1 e1 d1 c1 b1 a1; h0 g0 f0 e0 d0 c0 b0 a0
+    #  v5 = h3 g3 f3 e3 d3 c3 b3 a3; h2 g2 f2 e2 d2 c2 b2 a2
+    #  v6 = h5 g5 f5 e5 d5 c5 b5 a5; h4 g4 f4 e4 d4 c4 b4 a4
+    #  v7 = h7 g7 f7 e7 d7 c7 b7 a7; h6 g6 f6 e6 d6 c6 b6 a6
+    #  Thus 2 8x4 blocks are transposed.
+
+    TRN1 v4.16b, v0.16b, v1.16b
+    TRN2 v5.16b, v0.16b, v1.16b
+    TRN1 v6.16b, v2.16b, v3.16b
+    TRN2 v7.16b, v2.16b, v3.16b
+
+    TRN1 v0.8h, v4.8h, v6.8h
+    TRN2 v2.8h, v4.8h, v6.8h
+    TRN1 v1.8h, v5.8h, v7.8h
+    TRN2 v3.8h, v5.8h, v7.8h
+
+    UZP1 v4.4s, v0.4s, v1.4s
+    UZP2 v6.4s, v0.4s, v1.4s
+    UZP1 v5.4s, v2.4s, v3.4s
+    UZP2 v7.4s, v2.4s, v3.4s
+
+    ST1 {v4.16b}, [x4], 16
+    ST1 {v5.16b}, [x4], 16
+    ST1 {v6.16b}, [x4], 16
+    ST1 {v7.16b}, [x4], 16
+
+    SUBS x1, x1, 2
+
+    B.HS k_loop
+1:
+    CMP x1, -2
+    B.EQ 2f
+
+    LD1 {v0.s}[0], [x2]
+    LD1 {v0.s}[1], [x8]
+    LD1 {v1.s}[0], [x5]
+    LD1 {v1.s}[1], [x9]
+    LD1 {v2.s}[0], [x6]
+    LD1 {v2.s}[1], [x10]
+    LD1 {v3.s}[0], [x7]
+    LD1 {v3.s}[1], [x11]
+
+    #  Now we have 8x4 block of values that we will tranpose
+    #  A matrix
+    #  ----------------------------
+    #  |                          |
+    #  |                 a0-----a3|
+    #  |                 b0 B00 b3|
+    #  |   last block    c0     c3|
+    #  |                 d0-----d3|
+    #  |                 e0-----e3|
+    #  |                 f0 B01 f3|
+    #  |                 g0     g3|
+    #  |                 h0-----h3|
+    #  |                          |
+    #  |                          |
+    #  ---------------------------
+    #  v0 = -; e3 e2 e1 e0 a3 a2 a1 a0
+    #  v1 = -; f3 f2 f1 f0 b3 b2 b1 b0
+    #  v2 = -; g3 g2 g1 g0 c3 c2 c1 c0
+    #  v3 = -; h3 h2 h1 h0 d3 d2 d1 d0
+    #  Sequence:
+    #  TRN1 v4.16b, v0.16b, v1.16b
+    #  TRN2 v5.16b, v0.16b, v1.16b
+    #  TRN1 v6.16b, v2.16b, v3.16b
+    #  TRN2 v7.16b, v2.16b, v3.16b
+    #  Now we have
+    #  v4 = -;f2 e2 f0 e0 b2 a2 b0 a0
+    #  v5 = -;f3 e3 f1 e1 b3 a3 b1 a1
+    #  v6 = -;h2 g2 h0 g0 d2 c2 d0 c0
+    #  v7 = -;h3 g3 h1 g1 d3 c3 d1 c1
+    #  TRN1 v0.8h, v4.8h, v6.8h
+    #  TRN2 v2.8h, v4.8h, v6.8h
+    #  TRN1 v1.8h, v5.8h, v7.8h
+    #  TRN2 v3.8h, v5.8h, v7.8h
+    #  v0 = -;h0 g0 f0 e0 d0 c0 b0 a0
+    #  v1 = -;h1 g1 f1 e1 d1 c1 b1 a1
+    #  v2 = -;h2 g2 f2 e2 d2 c2 b2 a2
+    #  v3 = -;h3 g3 f3 e3 d3 c3 b3 a3
+    #  Thus 1 8x4 blocks are transposed.
+
+    TRN1 v4.16b, v0.16b, v1.16b
+    TRN2 v5.16b, v0.16b, v1.16b
+    TRN1 v6.16b, v2.16b, v3.16b
+    TRN2 v7.16b, v2.16b, v3.16b
+    TRN1 v0.8h, v4.8h, v6.8h
+    TRN2 v2.8h, v4.8h, v6.8h
+    TRN1 v1.8h, v5.8h, v7.8h
+    TRN2 v3.8h, v5.8h, v7.8h
+
+    ST1 {v0.8b}, [x4], 8
+    ST1 {v1.8b}, [x4], 8
+    ST1 {v2.8b}, [x4], 8
+    ST1 {v3.8b}, [x4]
+    .p2align 4
+2:
+    RET
+
+END_FUNCTION pytorch_q8gemm_sparse_packA_ukernel_8x4__aarch64_neon
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif