[CONTRIB] MPS DNN Dense (apache#615)

* mps

* update

Makefile

@@ -133,6 +133,7 @@ endif
 include make/contrib/cblas.mk
 include make/contrib/nnpack.mk
 include make/contrib/cudnn.mk
+include make/contrib/mps.mk
 
 ifdef ADD_CFLAGS
 	CFLAGS += $(ADD_CFLAGS)

make/config.mk

@@ -68,3 +68,6 @@ USE_NNPACK = 0
 
 # Whether use CuDNN
 USE_CUDNN = 0
+
+# Whether use MPS
+USE_MPS = 0

make/contrib/mps.mk

@@ -0,0 +1,14 @@
+MPS_CONTRIB_SRC = $(wildcard src/contrib/mps/*.mm, src/contrib/mps/*.cc)
+MPS_CONTRIB_OBJ = $(patsubst src/%.mm, build/%.o, $(MPS_CONTRIB_SRC))
+
+ifeq ($(USE_MPS), 1)
+FRAMEWORKS += -framework MetalPerformanceShaders
+CFLAGS += 
+ADD_LDFLAGS += 
+RUNTIME_DEP += $(MPS_CONTRIB_OBJ)
+endif
+
+build/contrib/mps/%.o: src/contrib/mps/%.mm src/contrib/mps/%.cc
+	@mkdir -p $(@D)
+	$(CXX) $(OBJCFLAGS) $(CFLAGS) -MM -MT build/contrib/mps/$*.o $< >build/contrib/mps/$*.d
+	$(CXX) $(OBJCFLAGS) -c $(CFLAGS) -c $< -o $@

python/tvm/contrib/mps.py

@@ -0,0 +1,35 @@
+"""External function interface to MPS libraroes."""
+from __future__ import absolute_import as _abs
+
+from .. import api as _api
+from .. import intrin as _intrin
+
+
+def matmul(lhs, rhs, transa=False, transb=False):
+    """Create an extern op that compute matrix mult of A and rhs with CrhsLAS
+
+    This function serves as an example on how to calle external libraries.
+
+    Parameters
+    ----------
+    lhs : Tensor
+        The left matrix operand
+    rhs : Tensor
+        The right matrix operand
+    transa : bool
+        Whether transpose lhs
+    transb : bool
+        Whether transpose rhs
+
+    Returns
+    -------
+    C : Tensor
+        The result tensor.
+    """
+    m = lhs.shape[0]
+    n = rhs.shape[1]
+    return _api.extern(
+        (n, m), [lhs, rhs],
+        lambda ins, outs: _intrin.call_packed(
+            "tvm.contrib.mps.matmul", ins[0], ins[1], outs[0], transa, transb),
+        name="C")

src/contrib/mps/gemm.mm

@@ -0,0 +1,93 @@
+#include "../../runtime/metal/metal_common.h"
+#include <MetalPerformanceShaders/MetalPerformanceShaders.h>
+#include <dmlc/logging.h>
+#include <tvm/runtime/device_api.h>
+#include <tvm/runtime/registry.h>
+#include <tvm/runtime/util.h>
+
+namespace tvm {
+namespace contrib {
+
+using namespace runtime;
+
+TVM_REGISTER_GLOBAL("tvm.contrib.mps.matmul")
+    .set_body([](TVMArgs args, TVMRetValue *ret) {
+      DLTensor *A = args[0];
+      DLTensor *B = args[1];
+      DLTensor *C = args[2];
+      bool transa = args[3];
+      bool transb = args[4];
+      // call gemm for simple compact code.
+      CHECK_EQ(A->ndim, 2);
+      CHECK_EQ(B->ndim, 2);
+      CHECK_EQ(C->ndim, 2);
+      CHECK(C->strides == nullptr);
+      CHECK(B->strides == nullptr);
+      CHECK(A->strides == nullptr);
+      CHECK(TypeMatch(A->dtype, kDLFloat, 32));
+      CHECK(TypeMatch(B->dtype, kDLFloat, 32));
+      CHECK(TypeMatch(C->dtype, kDLFloat, 32));
+      // Get Metal device API
+      MetalThreadEntry* entry_ptr = MetalThreadEntry::ThreadLocal();
+      CHECK_EQ(A->ctx, B->ctx);
+      CHECK_EQ(A->ctx, C->ctx);
+      id<MTLDevice> dev = entry_ptr->metal_api->GetDevice(A->ctx);
+      id<MTLCommandQueue> queue = entry_ptr->metal_api->GetCommandQueue(A->ctx);
+      id<MTLCommandBuffer> cb = [queue commandBuffer];
+      NSUInteger M = A->shape[0 + transa?1:0];
+      NSUInteger N = B->shape[1 - transb?1:0];
+      NSUInteger K = B->shape[0 + transb?1:0];
+      CHECK_EQ(A->shape[1-transa?1:0], K);
+      // mps a
+      MPSDataType dtype = MPSType::DLTypeToMPSType(A->dtype);
+      MPSMatrixDescriptor *descA = [MPSMatrixDescriptor
+          matrixDescriptorWithDimensions:M
+                                 columns:K
+                                rowBytes:M * sizeof(dtype)
+                                dataType:dtype];
+      id<MTLBuffer> bufA = (__bridge id<MTLBuffer>)(A->data);
+      MPSMatrix *matrixA =
+          [[MPSMatrix alloc] initWithBuffer:bufA descriptor:descA];
+      // mps b
+      MPSMatrixDescriptor *descB = [MPSMatrixDescriptor
+          matrixDescriptorWithDimensions:K
+                                 columns:N
+                                rowBytes:K * sizeof(dtype)
+                                dataType:dtype];
+      id<MTLBuffer> bufB = (__bridge id<MTLBuffer>)(B->data);
+      MPSMatrix *matrixB =
+          [[MPSMatrix alloc] initWithBuffer:bufB descriptor:descB];
+      // mps c
+      MPSMatrixDescriptor *descC = [MPSMatrixDescriptor
+          matrixDescriptorWithDimensions:M
+                                 columns:N
+                                rowBytes:M * sizeof(dtype)
+                                dataType:dtype];
+      id<MTLBuffer> bufC = (__bridge id<MTLBuffer>)(C->data);
+      MPSMatrix *matrixC =
+          [[MPSMatrix alloc] initWithBuffer:bufC descriptor:descC];
+      // kernel
+
+      MPSMatrixMultiplication *mul_obj = [[MPSMatrixMultiplication alloc] init];
+      MPSMatrixMultiplication *sgemm = [mul_obj initWithDevice:dev
+                                                 transposeLeft:transa
+                                                transposeRight:transb
+                                                    resultRows:M
+                                                 resultColumns:N
+                                               interiorColumns:K
+                                                         alpha:1.0f
+                                                          beta:0.0f];
+      CHECK(sgemm != nil);
+      [sgemm encodeToCommandBuffer:cb
+                        leftMatrix:matrixA
+                       rightMatrix:matrixB
+                      resultMatrix:matrixC];
+      [cb commit];
+      [mul_obj dealloc];
+      [matrixA dealloc];
+      [matrixB dealloc];
+      [matrixC dealloc];
+    });
+
+} // namespace contrib
+} // namespace tvm

src/contrib/mps/mps_utils.cc

@@ -0,0 +1,58 @@
+/*!
+ *  Copyright (c) 2017 by Contributors
+ * \file Use external mps utils function
+ */
+#include "mps_utils.h"
+#include <MetalPerformanceShaders/MetalPerformanceShaders.h>
+#include <dmlc/thread_local.h>
+#include <tvm/runtime/registry.h>
+
+
+namespace tvm {
+namespace contrib {
+
+// MPS Data Type
+MPSDataType MPSType::DLTypeToMPSType(const DLDataType &dtype) {
+  switch (dtype.code) {
+      case kDLInt:
+        if (dtype.bits == 8 && dtype.lanes == 1) return MPSDataTypeInt8;
+        else if (dtype.bits == 16 && dtype.lanes == 1) return MPSDataTypeInt16;
+        else
+          LOG(FATAL) << "Unsupported type";
+        break;
+      case kDLUInt:
+        if (dtype.bits == 8 && dtype.lanes == 1) return MPSDataTypeUInt8;
+        else if (dtype.bits == 16 && dtype.lanes == 1) return MPSDataTypeUInt16;
+        else if (dtype.bits == 32 && dtype.lanes == 1) return MPSDataTypeUInt32;
+        LOG(FATAL) << "Unsupported type";
+        break;
+      case kDLFloat:
+        if (dtype.bits == 16 && dtype.lanes == 1) return MPSDataTypeFloat16;
+        else if (dtype.bits == 32 && dtype.lanes == 1) return MPSDataTypeFloat32;
+        else
+          LOG(FATAL) << "Unsupported type";
+        break;
+      default:
+        LOG(FATAL) << "Unsupported type";
+    }
+}
+
+// MetalThreadEntry
+
+MetalThreadEntry::MetalThreadEntry() {
+  auto func = runtime::Registry::Get("device_api.metal");
+  void *ret = (*func)();
+  metal_api = static_cast<runtime::metal::MetalWorkspace *>(ret);
+}
+
+MetalThreadEntry::~MetalThreadEntry() {
+}
+
+typedef dmlc::ThreadLocalStore<MetalThreadEntry> MetalThreadStore;
+
+MetalThreadEntry* MetalThreadEntry::ThreadLocal() {
+  return MetalThreadStore::Get();
+}
+
+}  // namespace contrib
+}  // namespace tvm

src/contrib/mps/mps_utils.h

@@ -0,0 +1,33 @@
+/*!
+ *  Copyright (c) 2017 by Contributors
+ * \file Use external mps utils function
+ */
+
+#ifndef TVM_CONTRIB_MPS_MPS_UTILS_H_
+#define TVM_CONTRIB_MPS_MPS_UTILS_H_
+
+#include <dmlc/logging.h>
+#include <tvm/runtime/device_api.h>
+#include "../../runtime/metal/metal_common.h"
+
+
+namespace tvm {
+namespace contrib {
+
+/*! breif Convert DLTensor type to MPS type */
+struct MPSType {
+  static MPSDataType DLTypeToMPSType(const DLDataType &dtype);
+};  // struct MPSType
+
+
+struct MetalThreadEntry {
+  MetalThreadEntry();
+  ~MetalThreadEntry();
+  runtime::MetalWorkspace *metal_api{nullptr};
+  static MetalThreadEntry* ThreadLocal();
+};  // MetalThreadEntry
+
+}  // namespace contrib
+}  // namespace tvm
+
+#endif  // TVM_CONTRIB_MPS_MPS_UTILS_H_

tests/python/contrib/test_mps.py

@@ -0,0 +1,40 @@
+import tvm
+import numpy as np
+from tvm.contrib import mps
+
+def test_matmul_add():
+    n = 1024
+    l = 128
+    m = 235
+    bias = tvm.var('bias', dtype=tvm.float32)
+    A = tvm.placeholder((n, l), name='A')
+    B = tvm.placeholder((l, m), name='B')
+    C1 = mps.matmul(A, B)
+    C2 = mps.matmul(B, A, True, True)
+    D1 = tvm.compute(C1.shape, lambda i, j: C1[i,j] + bias, name="D1")
+    D2 = tvm.compute(C2.shape, lambda i, j: C2[i,j] + bias, name="D2")
+    s1 = tvm.create_schedule(D1.op)
+    s2 = tvm.create_schedule(D2.op)
+
+    def verify(A, B, D, s, bias, target="llvm"):
+        if not tvm.module.enabled(target):
+            print("skip because %s is not enabled..." % target)
+            return
+        if not tvm.get_global_func("tvm.contrib.mps.matmul", True):
+            print("skip because extern function is not avalable")
+            return
+        ctx = tvm.cpu(0)
+        f = tvm.build(s, [A, B, D, bias], target)
+        a = tvm.nd.array(np.random.uniform(size=(n, l)).astype(A.dtype), ctx)
+        b = tvm.nd.array(np.random.uniform(size=(l, m)).astype(B.dtype), ctx)
+        d = tvm.nd.array(np.zeros((n, m), dtype=D.dtype), ctx)
+        bb = 10.0
+        f(a, b, d, bb)
+        np.testing.assert_allclose(
+            d.asnumpy(), np.dot(a.asnumpy(), b.asnumpy()) + bb, rtol=1e-5)
+    verify(A, B, D1, s1, bias)
+    verify(A, B, D2, s2, bias)
+
+
+if __name__ == "__main__":
+    test_matmul_add()