Add AVX512VNNI support for TVM

tests/python/contrib/test_gemm_acc32_vnni.py

@@ -0,0 +1,111 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=import-self, invalid-name, unused-argument, too-many-lines, len-as-condition
+
+import tvm
+import numpy as np
+from topi.x86.tensor_intrin import dot_16x1x16_int8_int8_int32_vnni
+from topi.x86.tensor_intrin import dot_16x1x16_int8_int8_int32
+from nose.tools import nottest
+
+
+@nottest
+def test_fc_int8_acc32():
+    n = 1024
+    k = 1024
+    m = 1024
+
+    X = tvm.placeholder((m, k), name='X', dtype="uint8")
+    W = tvm.placeholder((n, k), name='W', dtype="int8")
+
+    peak = 280
+    print("Peak {} Gops/s".format(peak))
+    memory_ops = n * k + m * k + 2 * n * n
+    gops_per_mm = 2 * n * m * k
+
+    # def verify(target="llvm -mcpu=skylake-avx512"):
+    # For LLVM < 8.0, it shows "'cascadelake' is not a recognized processor for this target
+    # (ignoring processor)" error with the following setting. After LLVM 8.0 is enabled in the
+    # test, we should use cascadelake setting.
+    def verify(target="llvm -mcpu=cascadelake"):
+        if not tvm.module.enabled(target):
+            print("skip because %s is not enabled..." % target)
+            return
+
+        ctx = tvm.context(target, 0)
+        pc = dot_16x1x16_int8_int8_int32_vnni()
+        # pc = dot_16x1x16_int8_int8_int32()
+        ak = tvm.reduce_axis((0, k), name='k')
+        packedW = tvm.placeholder(
+            (n // 16, 16 * (k // 4), 4), name='packedW', dtype="int8")
+
+        t_fc = tvm.compute((m, n), lambda i, j: tvm.sum(X[i, ak].astype(
+            "int32") * packedW[j / 16, (ak / 4) * 16 + j % 16, ak % 4].astype("int32"), axis=ak), name="F")
+        t_sch = tvm.create_schedule(t_fc.op)
+        a_x, a_y = t_fc.op.axis
+        a_k, = t_fc.op.reduce_axis
+
+        a_yo, a_yi = t_sch[t_fc].split(a_y, factor=16)
+        a_xo, a_xi = t_sch[t_fc].split(a_x, factor=32)
+        a_ko, a_ki = t_sch[t_fc].split(a_k, factor=4)
+        a_koo, a_koi = t_sch[t_fc].split(a_ko, factor=4)
+        t_sch[t_fc].reorder(a_yo, a_xo, a_xi, a_koo, a_koi, a_yi, a_ki)
+
+        # measure multiple threading
+        # t_sch[t_fc].parallel(a_xo)
+        t_sch[t_fc].unroll(a_koi)
+        t_sch[t_fc].tensorize(a_yi, pc)
+
+        # print(tvm.lower(t_sch, [X, packedW, t_fc], simple_mode=True))
+        t_func = tvm.build(t_sch, [X, packedW, t_fc], target, name="intrinsic")
+        t_evaluator = t_func.time_evaluator(t_func.entry_name, ctx, number=10)
+
+        # generate the plain data
+        a_ = np.random.uniform(1, 10, size=(m, k)).astype("uint8")
+        b_ = np.random.uniform(1, 10, size=(n, k)).astype("int8")
+
+        packW = np.random.uniform(1, 10, size=(
+            n // 16, 16 * (k // 4), 4)).astype("int8")
+        # This occurs in pre_compute stage
+        for r_idx in range(n // 16):
+            for s_idx in range(16 * (k // 4)):
+                for t_idx in range(4):
+                    packW[r_idx][s_idx][t_idx] = b_[r_idx * 16 + s_idx %
+                                                    16][(s_idx // 16) * 4 + t_idx]
+
+        x = tvm.nd.array(a_, ctx)
+        w = tvm.nd.array(packW, ctx)
+        y = tvm.nd.array(np.zeros((m, n), dtype="int32"), ctx)
+        result = t_evaluator(x, w, y)
+
+        gops_per_sec = gops_per_mm / result.mean / 1e9
+        # verify the correctness
+        tvm.testing.assert_allclose(y.asnumpy(), np.dot(a_, b_.T), rtol=0)
+        print('Tensorization: running time: {:.3f} ms, {:.2f} Gops/s, effiency: {:.2f}'.format(
+            result.mean * 1000, gops_per_sec, gops_per_sec / peak))
+        # print('TVM with x86 micro-kernel: %f' % result.mean)
+        t_func.export_library("tensorize_acc32.o")
+
+    verify()
+
+
+if __name__ == "__main__":
+    # The test requires Cascade Lake and newer Intel machines to generate the
+    # correct AVX512 VNNI instruction. So, disabling the test.
+
+    # test_fc_int8_acc32()
+    pass

topi/python/topi/x86/tensor_intrin.py

@@ -21,7 +21,7 @@
 
 def dot_16x1x16_int8_int8_int32():
     """
-    Int8 dot product by every 4 elements using AVX2 Skylake instructions.
+    Int8 dot product by every 4 elements using AVX512 Skylake instructions.
     This function takes two arrays of int8 datatype -- data[4] and
     kernel[16][4] -- and computes a dot product of data[4] with every
     4 elements of kernels, resulting in output[16] of int32 datatype.
@@ -30,9 +30,9 @@ def dot_16x1x16_int8_int8_int32():
         void dot_16x1x16_int8_int8_int32(int8 data[4], int8 kernel[16][4],
                 int32 output[16]){
             for (int i = 0; i < 16; i++){
-                out[i] = 0;
+                output[i] = 0;
                 for (int k = 0; k < 4; k++){
-                    out[i] += data[k] * kernel[i][k]
+                    output[i] += data[k] * kernel[i][k]
                 }
             }
         }
@@ -102,38 +102,41 @@ def _instr(index):
 
 def dot_16x1x16_int8_int8_int16():
     """
-    Int8 dot product by every 2 elements using AVX2 Skylake instructions.
+    Int8 dot product by every 2 elements using AVX512 Skylake instructions.
     This function takes two arrays of int8 datatype -- data[2] and
     kernel[4][32][2] -- and computes a dot product of data[2] with every
     2 elements of kernels, resulting in output[4][32] of int16 datatype.
     The pseudo code is as follows.
     .. code-block:: c
         void dot_16x1x16_int8_int8_int16(int8 data[2], int8 kernel[32*4][2],
                 int16 output[32*4]){
-        for (int i = 0; i< 4; i++){
+            for (int i = 0; i< 4; i++){
                 for (int j = 0; j < 32; j++){
-                    out[i][i] = 0;
+                    output[i][i] = 0;
                     for (int k = 0; k < 2; k++){
-                        out[i][j][k] += data[k] * kernel[i][j][k]
+                        output[i][j][k] += data[k] * kernel[i][j][k]
                     }
                 }
+            }
         }
-        }
+
     Physically, the kernel array sits in four AVX512 vector registers and
     the data[2] is broadcasted to another AVX512 vector register. This
     function returns a TensorIntrin that can be used to tensorize
     a schedule.
+
     Returns
     -------
     intrin : TensorIntrin
         The Skylake int8 TensorIntrin that can be used in tensorizing schedule
     """
 
-    num_int8_elements = 2 # 2 int8 elements in int32
+    int16_lanes = 4*32 # 4*32 int32 lanes in 4 AVX512 vector registers
+    num_int8_elements = 2 # 2 int8 elements in int16
     data = tvm.placeholder((num_int8_elements,), dtype='uint8', name='data')
-    kernel = tvm.placeholder((128, num_int8_elements), dtype='int8', name='kernel')
+    kernel = tvm.placeholder((int16_lanes, num_int8_elements), dtype='int8', name='kernel')
     k = tvm.reduce_axis((0, num_int8_elements), name='k')
-    C = tvm.compute((128, ),
+    C = tvm.compute((int16_lanes, ),
                     lambda i: tvm.sum(data[k].astype('int16') *
                                       kernel[i, k].astype('int16'),
                                       axis=k),
@@ -177,3 +180,98 @@ def _instr(index):
 
     with tvm.build_config(offset_factor=1, partition_const_loop=True):
         return tvm.decl_tensor_intrin(C.op, _intrin_func, binds={data:a_buffer, kernel:b_buffer})
+
+
+def dot_16x1x16_int8_int8_int32_vnni():
+    """
+    Int8 dot product by every 4 elements using AVX512VNNI Cascade Lake instructions.
+    This function takes two arrays of int8 datatype -- data[4] and
+    kernel[16][4] -- and computes a dot product of data[4] with every
+    4 elements of kernels, resulting in output[16] of int32 datatype.
+    The pseudo code is as follows.
+    .. code-block:: c
+        void dot_16x1x16_int8_int8_int32_vnni(int8 data[4], int8 kernel[16][4],
+                int32 output[16]){
+            for (int i = 0; i < 16; i++){
+                output[i] = 0;
+                for (int k = 0; k < 4; k++){
+                    output[i] += data[k] * kernel[i][k]
+                }
+            }
+        }
+
+    Physically, the kernel array sits in an AVX512 vector register and
+    the data[4] is broadcasted to another AVX512 vector register. This
+    function returns a TensorIntrin that can be used to tensorize
+    a schedule.
+
+    Returns
+    -------
+    intrin : TensorIntrin
+        The Cascade Lake int8 TensorIntrin that can be used in tensorizing schedule
+    """
+
+    int32_lanes = 16 # 16 int32 lanes in AVX512
+    num_int8_elements = 4 # 4 int8 elements in int32
+    data = tvm.placeholder((num_int8_elements,), dtype='uint8', name='data')
+    kernel = tvm.placeholder((int32_lanes, num_int8_elements), dtype='int8', name='kernel')
+    k = tvm.reduce_axis((0, num_int8_elements), name='k')
+    C = tvm.compute((int32_lanes,),
+                    lambda i: tvm.sum(data[k].astype('int32') *
+                                      kernel[i, k].astype('int32'),
+                                      axis=k),
+                    name="C")
+
+    a_buffer = tvm.decl_buffer(data.shape, dtype='uint8', name="a_buffer",
+                               offset_factor=1,
+                               strides=[1])
+    b_buffer = tvm.decl_buffer(kernel.shape, dtype='int8', name="b_buffer",
+                               offset_factor=1,
+                               strides=[tvm.var('ldw'), 1])
+
+    def _intrin_func(ins, outs):
+        def _instr(index):
+            ib = tvm.ir_builder.create()
+            if index == 1:
+                ib.emit(outs[0].vstore(0, tvm.const(0, 'int32x16')))
+                return ib.get()
+
+            a_int8 = ins[0].vload([0], "uint8x4")
+            re_int32 = tvm.call_pure_intrin('int32', 'reinterpret', a_int8)
+            vec_ai32 = re_int32.astype('int32x16')
+            vec_b = ins[1].vload([0, 0], "int8x64")
+
+            vnni_inst_name = 'llvm.x86.avx512.vpdpbusd.512'
+            llvm_id = tvm.codegen.llvm_lookup_intrinsic_id(vnni_inst_name)
+
+            if llvm_id != 0: # VNNI is available for current LLVM version
+                vec_bi32 = tvm.call_pure_intrin('int32x16', 'reinterpret', vec_b)
+                vec_zero = tvm.const(0, "int32x16")
+                quad_reduction = tvm.call_llvm_intrin('int32x16',
+                                                      'llvm.x86.avx512.vpdpbusd.512',
+                                                      tvm.const(0, 'uint32'),
+                                                      vec_zero,
+                                                      vec_ai32, vec_bi32)
+            else: # Fall back to the normal AVX512
+                vec_a = tvm.call_pure_intrin('int8x64', 'reinterpret', vec_ai32)
+                vec_one = tvm.const(1, "int16x32")
+                pair_reduction = tvm.call_llvm_intrin('int16x32',
+                                                      'llvm.x86.avx512.pmaddubs.w.512',
+                                                      tvm.const(0, 'uint32'),
+                                                      vec_a, vec_b)
+                quad_reduction = tvm.call_llvm_intrin('int32x16',
+                                                      'llvm.x86.avx512.pmaddw.d.512',
+                                                      tvm.const(0, 'uint32'),
+                                                      pair_reduction, vec_one)
+
+            if index == 0:
+                ib.emit(outs[0].vstore(0, quad_reduction))
+            else:
+                ib.emit(outs[0].vstore(0, quad_reduction + outs[0].vload([0], 'int32x16')))
+            return ib.get()
+
+        # body, reset, update
+        return _instr(0), _instr(1), _instr(2)
+
+    with tvm.build_config(offset_factor=1, partition_const_loop=True):
+        return tvm.decl_tensor_intrin(C.op, _intrin_func, binds={data:a_buffer, kernel:b_buffer})