diff --git a/opensubdiv/osd/glslPatchGregory.glsl b/opensubdiv/osd/glslPatchGregory.glsl
index 7c316da2..2459fbba 100644
--- a/opensubdiv/osd/glslPatchGregory.glsl
+++ b/opensubdiv/osd/glslPatchGregory.glsl
@@ -430,7 +430,6 @@ void main()
 #ifdef OSD_PATCH_TESS_EVAL_GREGORY_SHADER
 
 layout(quads) in;
-layout(cw) in;
 
 #if defined OSD_FRACTIONAL_ODD_SPACING
     layout(fractional_odd_spacing) in;
@@ -523,9 +522,7 @@ void main()
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            vec3 A = q[i + 4*j];
-
+            vec3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
             CUCP[i] += A * C[j];
@@ -569,7 +566,7 @@ void main()
     BiTangent = (OsdModelViewMatrix() * vec4(BiTangent, 0)).xyz;
     Tangent = (OsdModelViewMatrix() * vec4(Tangent, 0)).xyz;
 
-    normal = normalize(cross(BiTangent, Tangent));
+    normal = normalize(cross(Tangent, BiTangent));
 
     outpt.v.Nu = Nu;
     outpt.v.Nv = Nv;
@@ -583,9 +580,7 @@ void main()
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            vec3 A = q[i + 4*j];
-
+            vec3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
         }
@@ -605,16 +600,16 @@ void main()
     BiTangent = (OsdModelViewMatrix() * vec4(BiTangent, 0)).xyz;
     Tangent = (OsdModelViewMatrix() * vec4(Tangent, 0)).xyz;
 
-    vec3 normal = normalize(cross(BiTangent, Tangent));
+    vec3 normal = normalize(cross(Tangent, BiTangent));
 
 #endif
 
     outpt.v.position = OsdModelViewMatrix() * vec4(WorldPos, 1.0f);
     outpt.v.normal = normal;
-    outpt.v.tangent = BiTangent;
-    outpt.v.bitangent = Tangent;
+    outpt.v.tangent = Tangent;
+    outpt.v.bitangent = BiTangent;
 
-    vec2 UV = vec2(v, u);
+    vec2 UV = vec2(u, v);
 
     OSD_USER_VARYING_PER_EVAL_POINT(UV, 0, 1, 3, 2);
 
diff --git a/opensubdiv/osd/glslPatchGregoryBasis.glsl b/opensubdiv/osd/glslPatchGregoryBasis.glsl
index 92b81d5f..0c45a67c 100644
--- a/opensubdiv/osd/glslPatchGregoryBasis.glsl
+++ b/opensubdiv/osd/glslPatchGregoryBasis.glsl
@@ -101,7 +101,6 @@ void main()
 #ifdef OSD_PATCH_TESS_EVAL_GREGORY_BASIS_SHADER
 
 layout(quads) in;
-layout(cw) in;
 
 #if defined OSD_FRACTIONAL_ODD_SPACING
     layout(fractional_odd_spacing) in;
@@ -180,9 +179,7 @@ void main()
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            vec3 A = q[i + 4*j];
-
+            vec3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
             CUCP[i] += A * C[j];
@@ -207,7 +204,7 @@ void main()
     dVV *= 6 * level;
     dUV *= 9 * level;
 
-    vec3 n = cross(BiTangent, Tangent);
+    vec3 n = cross(Tangent, BiTangent);
     vec3 normal = normalize(n);
 
     float E = dot(Tangent, Tangent);
@@ -235,9 +232,7 @@ void main()
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            vec3 A = q[i + 4*j];
-
+            vec3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
         }
@@ -254,15 +249,15 @@ void main()
     BiTangent *= 3 * level;
     Tangent *= 3 * level;
 
-    vec3 normal = normalize(cross(BiTangent, Tangent));
+    vec3 normal = normalize(cross(Tangent, BiTangent));
 
 #endif
     outpt.v.position = vec4(WorldPos, 1.0f);
     outpt.v.normal = normal;
-    outpt.v.tangent = BiTangent;
-    outpt.v.bitangent = Tangent;
+    outpt.v.tangent = Tangent;
+    outpt.v.bitangent = BiTangent;
 
-    vec2 UV = vec2(v, u);
+    vec2 UV = vec2(u, v);
 
     OSD_USER_VARYING_PER_EVAL_POINT(UV, 0, 5, 15, 10);
 
diff --git a/opensubdiv/osd/hlslPatchGregory.hlsl b/opensubdiv/osd/hlslPatchGregory.hlsl
index 3cc6624f..ac322ef9 100644
--- a/opensubdiv/osd/hlslPatchGregory.hlsl
+++ b/opensubdiv/osd/hlslPatchGregory.hlsl
@@ -295,7 +295,7 @@ HS_CONSTANT_FUNC_OUT HSConstFunc(
 
 [domain("quad")]
 [partitioning(HS_PARTITION)]
-[outputtopology("triangle_ccw")]
+[outputtopology("triangle_cw")]
 [outputcontrolpoints(4)]
 [patchconstantfunc("HSConstFunc")]
 GregDomainVertex hs_main_patches(
@@ -537,9 +537,7 @@ void ds_main_patches(
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            float3 A = q[i + 4*j];
-
+            float3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
             CUCP[i] += A * C[j];
@@ -583,7 +581,7 @@ void ds_main_patches(
     BiTangent = mul(OsdModelViewMatrix(), float4(BiTangent, 0)).xyz;
     Tangent = mul(OsdModelViewMatrix(), float4(Tangent, 0)).xyz;
 
-    normal = normalize(cross(BiTangent, Tangent));
+    normal = normalize(cross(Tangent, BiTangent));
 
     output.Nu = Nu;
     output.Nv = Nv;
@@ -597,9 +595,7 @@ void ds_main_patches(
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            float3 A = q[i + 4*j];
-
+            float3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
         }
@@ -619,18 +615,18 @@ void ds_main_patches(
     BiTangent = mul(OsdModelViewMatrix(), float4(BiTangent, 0)).xyz;
     Tangent = mul(OsdModelViewMatrix(), float4(Tangent, 0)).xyz;
 
-    float3 normal = normalize(cross(BiTangent, Tangent));
+    float3 normal = normalize(cross(Tangent, BiTangent));
 
 #endif
 
     output.position = mul(OsdModelViewMatrix(), float4(WorldPos, 1.0f));
     output.normal = normal;
-    output.tangent = BiTangent;
-    output.bitangent = Tangent;
+    output.tangent = Tangent;
+    output.bitangent = BiTangent;
 
     output.edgeDistance = 0;
 
-    float2 UV = float2(v, u);
+    float2 UV = float2(u, v);
     output.patchCoord = OsdInterpolatePatchCoord(UV, patch[0].patchCoord);
 
     OSD_DISPLACEMENT_CALLBACK;
diff --git a/opensubdiv/osd/hlslPatchGregoryBasis.hlsl b/opensubdiv/osd/hlslPatchGregoryBasis.hlsl
index 10c1b808..081dc146 100644
--- a/opensubdiv/osd/hlslPatchGregoryBasis.hlsl
+++ b/opensubdiv/osd/hlslPatchGregoryBasis.hlsl
@@ -78,7 +78,7 @@ HSConstFunc(
 
 [domain("quad")]
 [partitioning(HS_PARTITION)]
-[outputtopology("triangle_ccw")]
+[outputtopology("triangle_cw")]
 [outputcontrolpoints(20)]
 [patchconstantfunc("HSConstFunc")]
 HullVertex hs_main_patches(
@@ -159,9 +159,7 @@ void ds_main_patches(
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            float3 A = q[i + 4*j];
-
+            float3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
             CUCP[i] += A * C[j];
@@ -186,7 +184,7 @@ void ds_main_patches(
     dVV *= 6 * level;
     dUV *= 9 * level;
 
-    float3 n = cross(BiTangent, Tangent);
+    float3 n = cross(Tangent, BiTangent);
     float3 normal = normalize(n);
 
     float E = dot(Tangent, Tangent);
@@ -214,9 +212,7 @@ void ds_main_patches(
 
     for (int i=0; i<4; ++i) {
         for (uint j=0; j<4; ++j) {
-            // reverse face front
-            float3 A = q[i + 4*j];
-
+            float3 A = q[4*i + j];
             BUCP[i] += A * B[j];
             DUCP[i] += A * D[j];
         }
@@ -233,18 +229,18 @@ void ds_main_patches(
     BiTangent *= 3 * level;
     Tangent *= 3 * level;
 
-    float3 normal = normalize(cross(BiTangent, Tangent));
+    float3 normal = normalize(cross(Tangent, BiTangent));
 
 #endif
 
     output.position = float4(WorldPos, 1.0f);
     output.normal = normal;
-    output.tangent = BiTangent;
-    output.bitangent = Tangent;
+    output.tangent = Tangent;
+    output.bitangent = BiTangent;
 
     output.edgeDistance = 0;
 
-    float2 UV = float2(v, u);
+    float2 UV = float2(u, v);
     output.patchCoord = OsdInterpolatePatchCoord(UV, patch[0].patchCoord);
 
     OSD_DISPLACEMENT_CALLBACK;