Turn off the use of uniform arrays in shaders as some devices don't like them, and the speedup seems to be very small if any.

GPU/GLES/FragmentShaderGenerator.cpp

@@ -148,6 +148,7 @@ void GenerateFragmentShader(char *buffer) {
 	char *p = buffer;
 
 #if defined(GLSL_ES_1_0)
+	WRITE(p, "#version 100\n");  // GLSL ES 1.0
 	WRITE(p, "precision lowp float;\n");
 #elif !defined(FORCE_OPENGL_2_0)
 	WRITE(p, "#version 110\n");

GPU/GLES/ShaderManager.cpp

@@ -105,8 +105,16 @@ LinkedShader::LinkedShader(Shader *vs, Shader *fs)
 	u_view = glGetUniformLocation(program, "u_view");
 	u_world = glGetUniformLocation(program, "u_world");
 	u_texmtx = glGetUniformLocation(program, "u_texmtx");
-	u_bone = glGetUniformLocation(program, "u_bone");
 	numBones = gstate.getNumBoneWeights();
+#ifdef USE_BONE_ARRAY
+	u_bone = glGetUniformLocation(program, "u_bone");
+#else
+	for (int i = 0; i < numBones; i++) {
+		char name[10];
+		sprintf(name, "u_bone%i", i);
+		u_bone[i] = glGetUniformLocation(program, name);
+	}
+#endif
 
 	// Lighting, texturing
 	u_ambient = glGetUniformLocation(program, "u_ambient");
@@ -323,6 +331,7 @@ void LinkedShader::updateUniforms() {
 	}
 
 	// TODO: Could even set all bones in one go if they're all dirty.
+#ifdef USE_BONE_ARRAY
 	if (u_bone != -1) {
 		float allBones[8 * 16];
 
@@ -346,6 +355,15 @@ void LinkedShader::updateUniforms() {
 			}
 		}
 	}
+#else
+	float bonetemp[16];
+	for (int i = 0; i < numBones; i++) {
+		if (dirtyUniforms & (DIRTY_BONEMATRIX0 << i)) {
+			ConvertMatrix4x3To4x4(gstate.boneMatrix + 12 * i, bonetemp);
+			glUniformMatrix4fv(u_bone[i], 1, GL_FALSE, bonetemp);
+		}
+	}
+#endif
 
 	// Lighting
 	if (u_ambient != -1 && (dirtyUniforms & DIRTY_AMBIENT)) {

GPU/GLES/ShaderManager.h

@@ -54,7 +54,11 @@ public:
 	int u_view;
 	int u_texmtx;
 	int u_world;
+#ifdef USE_BONE_ARRAY
 	int u_bone;  // array, size is numBones
+#else
+	int u_bone[8];
+#endif
 	int numBones;
 	
 	// Fragment processing inputs

GPU/GLES/VertexShaderGenerator.cpp

@@ -128,6 +128,7 @@ void GenerateVertexShader(int prim, char *buffer) {
 // #define USE_FOR_LOOP
 
 #if defined(USING_GLES2)
+	WRITE(p, "#version 100\n");  // GLSL ES 1.0
 	WRITE(p, "precision highp float;\n");
 
 #elif !defined(FORCE_OPENGL_2_0)
@@ -213,7 +214,13 @@ void GenerateVertexShader(int prim, char *buffer) {
 			WRITE(p, "uniform mediump mat4 u_texmtx;\n");
 		if ((vertType & GE_VTYPE_WEIGHT_MASK) != GE_VTYPE_WEIGHT_NONE) {
 			int numBones = 1 + ((vertType & GE_VTYPE_WEIGHTCOUNT_MASK) >> GE_VTYPE_WEIGHTCOUNT_SHIFT);
+#ifdef USE_BONE_ARRAY
 			WRITE(p, "uniform mediump mat4 u_bone[%i];\n", numBones);
+#else
+			for (int i = 0; i < numBones; i++) {
+				WRITE(p, "uniform mediump mat4 u_bone%i;\n", i);
+			}
+#endif
 		}
 		for (int i = 0; i < 4; i++) {
 			if (doLight[i] != LIGHT_OFF) {
@@ -303,7 +310,7 @@ void GenerateVertexShader(int prim, char *buffer) {
 				"a_w2.x", "a_w2.y", "a_w2.z", "a_w2.w",
 			};
 			
-#ifdef USE_FOR_LOOP
+#if defined(USE_FOR_LOOP) && defined(USE_BONE_ARRAY)
 
 			// To loop through the weights, we unfortunately need to put them in a float array.
 			// GLSL ES sucks - no way to directly initialize an array!
@@ -326,6 +333,8 @@ void GenerateVertexShader(int prim, char *buffer) {
 			}
 
 #else
+
+#ifdef USE_BONE_ARRAY
 			if (numWeights == 1)
 				WRITE(p, "  mat4 skinMatrix = a_w1 * u_bone[0]");
 			else
@@ -337,9 +346,24 @@ void GenerateVertexShader(int prim, char *buffer) {
 				if (numWeights == 5 && i == 4) weightAttr = "a_w2";
 				WRITE(p, " + %s * u_bone[%i]", weightAttr, i);
 			}
-			WRITE(p, ";\n");
+#else
+			if (numWeights == 1)
+				WRITE(p, "  mat4 skinMatrix = a_w1 * u_bone0");
+			else
+				WRITE(p, "  mat4 skinMatrix = a_w1.x * u_bone0");
+			for (int i = 1; i < numWeights; i++) {
+				const char *weightAttr = boneWeightAttr[i];
+				// workaround for "cant do .x of scalar" issue
+				if (numWeights == 1 && i == 0) weightAttr = "a_w1";
+				if (numWeights == 5 && i == 4) weightAttr = "a_w2";
+				WRITE(p, " + %s * u_bone%i", weightAttr, i);
+			}
 #endif
 
+#endif
+
+			WRITE(p, ";\n");
+
 			// Trying to simplify this results in bugs in LBP...
 			WRITE(p, "  vec3 skinnedpos = (skinMatrix * vec4(a_position, 1.0)).xyz * %f;\n", factor);
 			WRITE(p, "  vec3 worldpos = (u_world * vec4(skinnedpos, 1.0)).xyz;\n");

GPU/GLES/VertexShaderGenerator.h

@@ -19,6 +19,8 @@
 
 #include "Globals.h"
 
+// #define USE_BONE_ARRAY
+
 struct VertexShaderID
 {
 	VertexShaderID() {d[0] = 0xFFFFFFFF;}