From c3c6cbd3ee3fbcca486565fd46f8e92c956dde88 Mon Sep 17 00:00:00 2001
From: Yanzhan Yang <yangyanzhan@baidu.com>
Date: Fri, 10 May 2019 20:01:39 +0800
Subject: [PATCH] change pixelFormat only when useAggressiveOptimization is set
 (#1614)

---
 .../paddle-mobile/Src/Framework/Texture.swift   | 17 +++++++++++++----
 .../Src/Operators/Kernels/ConvAddKernel.swift   | 12 ++++++++++--
 2 files changed, 23 insertions(+), 6 deletions(-)

diff --git a/metal/paddle-mobile/paddle-mobile/Src/Framework/Texture.swift b/metal/paddle-mobile/paddle-mobile/Src/Framework/Texture.swift
index c99b545f39..63250fa085 100644
--- a/metal/paddle-mobile/paddle-mobile/Src/Framework/Texture.swift
+++ b/metal/paddle-mobile/paddle-mobile/Src/Framework/Texture.swift
@@ -71,6 +71,7 @@ extension InputTexture {
 public class Texture: Tensorial {
     public var dim: Dim
     public var tensorDim: Dim
+    public var useMPS = false
     
     /// tensor dim pad to four
     public var padToFourDim: Dim
@@ -135,14 +136,22 @@ public class Texture: Tensorial {
         }
         
         if computePrecision == .Float16 {
-            if tensorDim[1] == 1 {
-                tmpTextureDes.pixelFormat = .r16Float
+            if useMPS {
+                if tensorDim[1] == 1 {
+                    tmpTextureDes.pixelFormat = .r16Float
+                } else {
+                    tmpTextureDes.pixelFormat = .rgba16Float
+                }
             } else {
                 tmpTextureDes.pixelFormat = .rgba16Float
             }
         } else if computePrecision == .Float32 {
-            if tensorDim[1] == 1 {
-                tmpTextureDes.pixelFormat = .r32Float
+            if useMPS {
+                if tensorDim[1] == 1 {
+                    tmpTextureDes.pixelFormat = .r32Float
+                } else {
+                    tmpTextureDes.pixelFormat = .rgba32Float
+                }
             } else {
                 tmpTextureDes.pixelFormat = .rgba32Float
             }
diff --git a/metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvAddKernel.swift b/metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvAddKernel.swift
index e4fa5b1d67..0e893d35e6 100644
--- a/metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvAddKernel.swift
+++ b/metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvAddKernel.swift
@@ -112,8 +112,14 @@ class ConvAddKernel<P: PrecisionProtocol>: Kernel, Computable {
         var shouldUseMPS = false
         let functionName = type(of: self).kernelFunctionName(param: param, useAggressiveOptimization: initContext.useAggresiveOptimization)
         if #available(iOS 11.0, *), (initContext.useMPS || initContext.useAggresiveOptimization) {
-            if (param.input.tensorDim[1] == 1 || param.input.tensorDim[1] > 4) && (param.output.tensorDim[1] == 1 || param.output.tensorDim[1] > 4) {
-                shouldUseMPS = true
+            if initContext.useAggresiveOptimization {
+                if (param.input.tensorDim[1] == 1 || param.input.tensorDim[1] > 4) && (param.output.tensorDim[1] == 1 || param.output.tensorDim[1] > 4) {
+                    shouldUseMPS = true
+                }
+            } else {
+                if param.input.tensorDim[1] > 4 && param.output.tensorDim[1] > 4 {
+                    shouldUseMPS = true
+                }
             }
         }
         if type(of: self).isWinoGrad(functionName: functionName) {
@@ -166,6 +172,8 @@ class ConvAddKernel<P: PrecisionProtocol>: Kernel, Computable {
         
         let isDepthWise = param.filter.tensorDim[1] == 1 && param.filter.tensorDim[0] == param.input.tensorDim[1]
         if #available(iOS 11.0, *) {
+            param.input.useMPS = true
+            param.output.useMPS = true
             let desc: MPSCNNConvolutionDescriptor = isDepthWise ?
                 MPSCNNDepthWiseConvolutionDescriptor(kernelWidth: param.filter.tensorDim[3],
                                                      kernelHeight: param.filter.tensorDim[2],
-- 
GitLab