flexible coopmat mnk and unified elempack for vulkan convolution winograd gemm (#6196)

src/layer/vulkan/shader/convolution_1x1s1d1_cm.comp

@@ -331,8 +331,6 @@ void main()
                 {
                     if (elempack == 1)
                     {
-                        const uint cstepd4 = psc(cstep) / 4;
-
                         const uint Md4_K_USGM_USGK = Md4 * K * UNROLL_SG_M * UNROLL_SG_K;
                         const uint Md4_K_USGM_USGK_d_subgroupsize = (Md4_K_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                         [[unroll]] for (uint q = 0; q < Md4_K_USGM_USGK_d_subgroupsize; q++)
@@ -347,8 +345,6 @@ void main()
                     }
                     else // if (elempack == 4)
                     {
-                        const uint inchd4 = inch / 4;
-
                         const uint Kd4_M_USGM_USGK = Kd4 * M * UNROLL_SG_M * UNROLL_SG_K;
                         const uint Kd4_M_USGM_USGK_d_subgroupsize = (Kd4_M_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                         [[unroll]] for (uint q = 0; q < Kd4_M_USGM_USGK_d_subgroupsize; q++)
@@ -365,11 +361,7 @@ void main()
 
                 // load weight
                 {
-                    const uint Nd4_K_USGN = Nd4 * K * UNROLL_SG_N;
                     const uint Nd4_K_USGN_USGK = Nd4 * K * UNROLL_SG_N * UNROLL_SG_K;
-
-                    const uint w_offset = ((wgni * kk) * UNROLL_WG_N) * Nd4_K_USGN + ((k / UNROLL_SG_K) * UNROLL_WG_N + sgni) * Nd4_K_USGN_USGK;
-
                     const uint Nd4_K_USGN_USGK_d_subgroupsize = (Nd4_K_USGN_USGK + (ncnn_subgroupSize * UNROLL_WG_M - 1)) / (ncnn_subgroupSize * UNROLL_WG_M);
                     [[unroll]] for (uint q = 0; q < Nd4_K_USGN_USGK_d_subgroupsize; q++)
                     {
@@ -537,29 +529,29 @@ void main()
                 {
                     if (elempack == 1)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         coopMatLoad(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Md4 * K), Md4, gl_CooperativeMatrixLayoutColumnMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         coopMatLoadNV(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Md4 * K), Md4, true);
-    #endif
+#endif
                     }
                     else // if (elempack == 4)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         coopMatLoad(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Kd4 * M), Kd4, gl_CooperativeMatrixLayoutRowMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         coopMatLoadNV(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Kd4 * M), Kd4, false);
-    #endif
+#endif
                     }
                 }
 
                 [[unroll]] for (uint zn = 0; zn < UNROLL_SG_N; zn++)
                 {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                     coopMatLoad(B[zn], tmp_k[sgni], (zk * UNROLL_SG_N + zn) * (Nd4 * K), Nd4, gl_CooperativeMatrixLayoutRowMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                     coopMatLoadNV(B[zn], tmp_k[sgni], (zk * UNROLL_SG_N + zn) * (Nd4 * K), Nd4, false);
-    #endif
+#endif
                 }
 
                 // sum += k * v
@@ -567,11 +559,11 @@ void main()
                 {
                     [[unroll]] for (uint zm = 0; zm < UNROLL_SG_M; zm++)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         sum[zn][zm] = coopMatMulAdd(A[zm], B[zn], sum[zn][zm]);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         sum[zn][zm] = coopMatMulAddNV(A[zm], B[zn], sum[zn][zm]);
-    #endif
+#endif
                     }
                 }
             }
@@ -585,8 +577,6 @@ void main()
             {
                 if (elempack == 1)
                 {
-                    const uint cstepd4 = psc(cstep) / 4;
-
                     const uint Md4_K_USGM_USGK = Md4 * K * UNROLL_SG_M * UNROLL_SG_K;
                     const uint Md4_K_USGM_USGK_d_subgroupsize = (Md4_K_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                     [[unroll]] for (uint q = 0; q < Md4_K_USGM_USGK_d_subgroupsize; q++)
@@ -601,8 +591,6 @@ void main()
                 }
                 else // if (elempack == 4)
                 {
-                    const uint inchd4 = inch / 4;
-
                     const uint Kd4_M_USGM_USGK = Kd4 * M * UNROLL_SG_M * UNROLL_SG_K;
                     const uint Kd4_M_USGM_USGK_d_subgroupsize = (Kd4_M_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                     [[unroll]] for (uint q = 0; q < Kd4_M_USGM_USGK_d_subgroupsize; q++)
@@ -619,11 +607,7 @@ void main()
 
             // load weight
             {
-                const uint Nd4_K_USGN = Nd4 * K * UNROLL_SG_N;
                 const uint Nd4_K_USGN_USGK = Nd4 * K * UNROLL_SG_N * UNROLL_SG_K;
-
-                const uint w_offset = ((wgni * kk) * UNROLL_WG_N) * Nd4_K_USGN + ((k / UNROLL_SG_K) * UNROLL_WG_N + sgni) * Nd4_K_USGN_USGK;
-
                 const uint Nd4_K_USGN_USGK_d_subgroupsize = (Nd4_K_USGN_USGK + (ncnn_subgroupSize * UNROLL_WG_M - 1)) / (ncnn_subgroupSize * UNROLL_WG_M);
                 [[unroll]] for (uint q = 0; q < Nd4_K_USGN_USGK_d_subgroupsize; q++)
                 {

src/layer/vulkan/shader/convolution_3x3s1d1_winograd23_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfp bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 4x4
@@ -115,22 +115,22 @@ void main()
     v33 = m33 - m31;
 
     // store 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v10);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v11);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v12);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v13);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v20);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v21);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v22);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v23);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v30);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v31);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v32);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v33);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v10);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v11);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v12);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v13);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v20);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v21);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v22);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v23);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v30);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v31);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v32);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v33);
 }

src/layer/vulkan/shader/convolution_3x3s1d1_winograd23_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfp top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfp bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,28 +44,28 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afp v00 = buffer_ld1(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afp v01 = buffer_ld1(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afp v02 = buffer_ld1(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afp v03 = buffer_ld1(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afp v10 = buffer_ld1(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afp v11 = buffer_ld1(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afp v12 = buffer_ld1(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afp v13 = buffer_ld1(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afp v20 = buffer_ld1(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afp v21 = buffer_ld1(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afp v22 = buffer_ld1(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afp v23 = buffer_ld1(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afp v30 = buffer_ld1(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afp v31 = buffer_ld1(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afp v32 = buffer_ld1(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afp v33 = buffer_ld1(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afp v00 = buffer_ld1(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afp v01 = buffer_ld1(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afp v02 = buffer_ld1(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afp v03 = buffer_ld1(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afp v10 = buffer_ld1(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afp v11 = buffer_ld1(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afp v12 = buffer_ld1(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afp v13 = buffer_ld1(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afp v20 = buffer_ld1(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afp v21 = buffer_ld1(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afp v22 = buffer_ld1(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afp v23 = buffer_ld1(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afp v30 = buffer_ld1(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afp v31 = buffer_ld1(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afp v32 = buffer_ld1(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afp v33 = buffer_ld1(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
 
     // const float itm[2][4] = {
     //     {1.0f,  1.0f,  1.0f,  0.0f},

src/layer/vulkan/shader/convolution_3x3s1d1_winograd43_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfp bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 6x6
@@ -187,42 +187,42 @@ void main()
     v55 = m51 - m53 * afp(2.5) + m55;
 
     // store 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v04);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v05);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v10);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v11);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v12);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v13);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v14);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v15);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v20);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v21);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v22);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v23);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep), v24);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep), v25);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep), v30);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep), v31);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep), v32);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep), v33);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep), v34);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep), v35);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep), v40);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep), v41);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep), v42);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep), v43);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep), v44);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep), v45);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep), v50);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep), v51);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep), v52);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep), v53);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep), v54);
-    buffer_st1(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep), v55);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v04);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v05);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v10);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v11);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v12);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v13);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v14);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v15);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v20);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v21);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v22);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v23);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep) * c, v24);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep) * c, v25);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep) * c, v30);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep) * c, v31);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep) * c, v32);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep) * c, v33);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep) * c, v34);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep) * c, v35);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep) * c, v40);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep) * c, v41);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep) * c, v42);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep) * c, v43);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep) * c, v44);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep) * c, v45);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep) * c, v50);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep) * c, v51);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep) * c, v52);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep) * c, v53);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep) * c, v54);
+    buffer_st1(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep) * c, v55);
 }

src/layer/vulkan/shader/convolution_3x3s1d1_winograd43_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfp top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfp bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,48 +44,48 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afp v00 = buffer_ld1(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afp v01 = buffer_ld1(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afp v02 = buffer_ld1(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afp v03 = buffer_ld1(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afp v04 = buffer_ld1(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afp v05 = buffer_ld1(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afp v10 = buffer_ld1(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afp v11 = buffer_ld1(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afp v12 = buffer_ld1(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afp v13 = buffer_ld1(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afp v14 = buffer_ld1(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afp v15 = buffer_ld1(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afp v20 = buffer_ld1(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afp v21 = buffer_ld1(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afp v22 = buffer_ld1(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afp v23 = buffer_ld1(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
-    afp v24 = buffer_ld1(top_tm_blob_data, v_tm_offset + 16 * psc(cstep));
-    afp v25 = buffer_ld1(top_tm_blob_data, v_tm_offset + 17 * psc(cstep));
-    afp v30 = buffer_ld1(top_tm_blob_data, v_tm_offset + 18 * psc(cstep));
-    afp v31 = buffer_ld1(top_tm_blob_data, v_tm_offset + 19 * psc(cstep));
-    afp v32 = buffer_ld1(top_tm_blob_data, v_tm_offset + 20 * psc(cstep));
-    afp v33 = buffer_ld1(top_tm_blob_data, v_tm_offset + 21 * psc(cstep));
-    afp v34 = buffer_ld1(top_tm_blob_data, v_tm_offset + 22 * psc(cstep));
-    afp v35 = buffer_ld1(top_tm_blob_data, v_tm_offset + 23 * psc(cstep));
-    afp v40 = buffer_ld1(top_tm_blob_data, v_tm_offset + 24 * psc(cstep));
-    afp v41 = buffer_ld1(top_tm_blob_data, v_tm_offset + 25 * psc(cstep));
-    afp v42 = buffer_ld1(top_tm_blob_data, v_tm_offset + 26 * psc(cstep));
-    afp v43 = buffer_ld1(top_tm_blob_data, v_tm_offset + 27 * psc(cstep));
-    afp v44 = buffer_ld1(top_tm_blob_data, v_tm_offset + 28 * psc(cstep));
-    afp v45 = buffer_ld1(top_tm_blob_data, v_tm_offset + 29 * psc(cstep));
-    afp v50 = buffer_ld1(top_tm_blob_data, v_tm_offset + 30 * psc(cstep));
-    afp v51 = buffer_ld1(top_tm_blob_data, v_tm_offset + 31 * psc(cstep));
-    afp v52 = buffer_ld1(top_tm_blob_data, v_tm_offset + 32 * psc(cstep));
-    afp v53 = buffer_ld1(top_tm_blob_data, v_tm_offset + 33 * psc(cstep));
-    afp v54 = buffer_ld1(top_tm_blob_data, v_tm_offset + 34 * psc(cstep));
-    afp v55 = buffer_ld1(top_tm_blob_data, v_tm_offset + 35 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afp v00 = buffer_ld1(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afp v01 = buffer_ld1(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afp v02 = buffer_ld1(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afp v03 = buffer_ld1(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afp v04 = buffer_ld1(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afp v05 = buffer_ld1(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afp v10 = buffer_ld1(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afp v11 = buffer_ld1(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afp v12 = buffer_ld1(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afp v13 = buffer_ld1(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afp v14 = buffer_ld1(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afp v15 = buffer_ld1(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afp v20 = buffer_ld1(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afp v21 = buffer_ld1(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afp v22 = buffer_ld1(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afp v23 = buffer_ld1(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
+    afp v24 = buffer_ld1(top_tm_blob_data, v_tm_offset + 16 * psc(cstep) * c);
+    afp v25 = buffer_ld1(top_tm_blob_data, v_tm_offset + 17 * psc(cstep) * c);
+    afp v30 = buffer_ld1(top_tm_blob_data, v_tm_offset + 18 * psc(cstep) * c);
+    afp v31 = buffer_ld1(top_tm_blob_data, v_tm_offset + 19 * psc(cstep) * c);
+    afp v32 = buffer_ld1(top_tm_blob_data, v_tm_offset + 20 * psc(cstep) * c);
+    afp v33 = buffer_ld1(top_tm_blob_data, v_tm_offset + 21 * psc(cstep) * c);
+    afp v34 = buffer_ld1(top_tm_blob_data, v_tm_offset + 22 * psc(cstep) * c);
+    afp v35 = buffer_ld1(top_tm_blob_data, v_tm_offset + 23 * psc(cstep) * c);
+    afp v40 = buffer_ld1(top_tm_blob_data, v_tm_offset + 24 * psc(cstep) * c);
+    afp v41 = buffer_ld1(top_tm_blob_data, v_tm_offset + 25 * psc(cstep) * c);
+    afp v42 = buffer_ld1(top_tm_blob_data, v_tm_offset + 26 * psc(cstep) * c);
+    afp v43 = buffer_ld1(top_tm_blob_data, v_tm_offset + 27 * psc(cstep) * c);
+    afp v44 = buffer_ld1(top_tm_blob_data, v_tm_offset + 28 * psc(cstep) * c);
+    afp v45 = buffer_ld1(top_tm_blob_data, v_tm_offset + 29 * psc(cstep) * c);
+    afp v50 = buffer_ld1(top_tm_blob_data, v_tm_offset + 30 * psc(cstep) * c);
+    afp v51 = buffer_ld1(top_tm_blob_data, v_tm_offset + 31 * psc(cstep) * c);
+    afp v52 = buffer_ld1(top_tm_blob_data, v_tm_offset + 32 * psc(cstep) * c);
+    afp v53 = buffer_ld1(top_tm_blob_data, v_tm_offset + 33 * psc(cstep) * c);
+    afp v54 = buffer_ld1(top_tm_blob_data, v_tm_offset + 34 * psc(cstep) * c);
+    afp v55 = buffer_ld1(top_tm_blob_data, v_tm_offset + 35 * psc(cstep) * c);
 
 #define sq2 1.41421356237
 #define sq2_m2 1.41421356237*2

src/layer/vulkan/shader/convolution_3x3s1d1_winograd_gemm.comp

@@ -6,14 +6,13 @@
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfp bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfp top_tm_blob_data[]; };
@@ -21,11 +20,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfp weight_tm_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -41,7 +37,7 @@ void main()
     int gz = int(gl_GlobalInvocationID.z);
 
 #if !NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
@@ -50,21 +46,21 @@ void main()
     afp sum2 = afp(0.f);
     afp sum3 = afp(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = gz * psc(c) * psc(outc) + gy * psc(c);
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = gz * c * outc + gy * c;
 
 #if NCNN_shader_local_memory
     const int lx = int(gl_LocalInvocationID.x);
     const int ly = int(gl_LocalInvocationID.y);
 
     int z = 0;
-    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
+    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < c; z += LOCAL_MEMORY_UNROLL_INCH)
     {
         if (ly < 4)
         {
             for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(outw) + ly));
+                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(cstep) + ly));
             }
         }
 
@@ -93,21 +89,21 @@ void main()
             sum3 += v3 * k;
         }
 
-        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(outw);
+        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
         w_offset += LOCAL_MEMORY_UNROLL_INCH;
 
         barrier();
     }
 
-    if (z < psc(c))
+    if (z < c)
     {
-        const int remain = psc(c) - z;
+        const int remain = c - z;
 
         if (ly < 4)
         {
             for (int z4 = 0; z4 < remain; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(outw) + ly));
+                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(cstep) + ly));
             }
         }
 
@@ -137,7 +133,7 @@ void main()
         }
     }
 #else
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afp v0 = buffer_ld1(bottom_tm_blob_data, v_offset + 0);
         afp v1 = buffer_ld1(bottom_tm_blob_data, v_offset + 1);
@@ -151,17 +147,17 @@ void main()
         sum2 += v2 * k;
         sum3 += v3 * k;
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 1;
     }
 #endif
 
 #if NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st1(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st1(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_gemm_cm.comp

@@ -385,8 +385,6 @@ void main()
                 {
                     if (elempack == 1)
                     {
-                        const uint cstepd4 = psc(cstep) / 4;
-
                         const uint Md4_K_USGM_USGK = Md4 * K * UNROLL_SG_M * UNROLL_SG_K;
                         const uint Md4_K_USGM_USGK_d_subgroupsize = (Md4_K_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                         [[unroll]] for (uint q = 0; q < Md4_K_USGM_USGK_d_subgroupsize; q++)
@@ -401,8 +399,6 @@ void main()
                     }
                     else // if (elempack == 4)
                     {
-                        const uint inchd4 = inch / 4;
-
                         const uint Kd4_M_USGM_USGK = Kd4 * M * UNROLL_SG_M * UNROLL_SG_K;
                         const uint Kd4_M_USGM_USGK_d_subgroupsize = (Kd4_M_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                         [[unroll]] for (uint q = 0; q < Kd4_M_USGM_USGK_d_subgroupsize; q++)
@@ -419,11 +415,7 @@ void main()
 
                 // load weight
                 {
-                    const uint Nd4_K_USGN = Nd4 * K * UNROLL_SG_N;
                     const uint Nd4_K_USGN_USGK = Nd4 * K * UNROLL_SG_N * UNROLL_SG_K;
-
-                    const uint w_offset = ((wgni * kk) * UNROLL_WG_N) * Nd4_K_USGN + ((k / UNROLL_SG_K) * UNROLL_WG_N + sgni) * Nd4_K_USGN_USGK;
-
                     const uint Nd4_K_USGN_USGK_d_subgroupsize = (Nd4_K_USGN_USGK + (ncnn_subgroupSize * UNROLL_WG_M - 1)) / (ncnn_subgroupSize * UNROLL_WG_M);
                     [[unroll]] for (uint q = 0; q < Nd4_K_USGN_USGK_d_subgroupsize; q++)
                     {
@@ -629,29 +621,29 @@ void main()
                 {
                     if (elempack == 1)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         coopMatLoad(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Md4 * K), Md4, gl_CooperativeMatrixLayoutColumnMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         coopMatLoadNV(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Md4 * K), Md4, true);
-    #endif
+#endif
                     }
                     else // if (elempack == 4)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         coopMatLoad(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Kd4 * M), Kd4, gl_CooperativeMatrixLayoutRowMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         coopMatLoadNV(A[zm], tmp_v[sgmi], (zk * UNROLL_SG_M + zm) * (Kd4 * M), Kd4, false);
-    #endif
+#endif
                     }
                 }
 
                 [[unroll]] for (uint zn = 0; zn < UNROLL_SG_N; zn++)
                 {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                     coopMatLoad(B[zn], tmp_k[sgni], (zk * UNROLL_SG_N + zn) * (Nd4 * K), Nd4, gl_CooperativeMatrixLayoutRowMajor);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                     coopMatLoadNV(B[zn], tmp_k[sgni], (zk * UNROLL_SG_N + zn) * (Nd4 * K), Nd4, false);
-    #endif
+#endif
                 }
 
                 // sum += k * v
@@ -659,11 +651,11 @@ void main()
                 {
                     [[unroll]] for (uint zm = 0; zm < UNROLL_SG_M; zm++)
                     {
-    #if ncnn_VK_KHR_cooperative_matrix
+#if ncnn_VK_KHR_cooperative_matrix
                         sum[zn][zm] = coopMatMulAdd(A[zm], B[zn], sum[zn][zm]);
-    #elif ncnn_VK_NV_cooperative_matrix
+#elif ncnn_VK_NV_cooperative_matrix
                         sum[zn][zm] = coopMatMulAddNV(A[zm], B[zn], sum[zn][zm]);
-    #endif
+#endif
                     }
                 }
             }
@@ -691,8 +683,6 @@ void main()
                 }
                 else // if (elempack == 4)
                 {
-                    const uint inchd4 = inch / 4;
-
                     const uint Kd4_M_USGM_USGK = Kd4 * M * UNROLL_SG_M * UNROLL_SG_K;
                     const uint Kd4_M_USGM_USGK_d_subgroupsize = (Kd4_M_USGM_USGK + (ncnn_subgroupSize * UNROLL_WG_N - 1)) / (ncnn_subgroupSize * UNROLL_WG_N);
                     [[unroll]] for (uint q = 0; q < Kd4_M_USGM_USGK_d_subgroupsize; q++)
@@ -709,11 +699,7 @@ void main()
 
             // load weight
             {
-                const uint Nd4_K_USGN = Nd4 * K * UNROLL_SG_N;
                 const uint Nd4_K_USGN_USGK = Nd4 * K * UNROLL_SG_N * UNROLL_SG_K;
-
-                const uint w_offset = ((wgni * kk) * UNROLL_WG_N) * Nd4_K_USGN + ((k / UNROLL_SG_K) * UNROLL_WG_N + sgni) * Nd4_K_USGN_USGK;
-
                 const uint Nd4_K_USGN_USGK_d_subgroupsize = (Nd4_K_USGN_USGK + (ncnn_subgroupSize * UNROLL_WG_M - 1)) / (ncnn_subgroupSize * UNROLL_WG_M);
                 [[unroll]] for (uint q = 0; q < Nd4_K_USGN_USGK_d_subgroupsize; q++)
                 {

src/layer/vulkan/shader/convolution_pack1to4_3x3s1d1_winograd_gemm.comp

@@ -6,14 +6,13 @@
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfp bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
@@ -21,11 +20,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec4 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -41,7 +37,7 @@ void main()
     int gz = int(gl_GlobalInvocationID.z);
 
 #if !NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
@@ -50,21 +46,21 @@ void main()
     afpvec4 sum2 = afpvec4(0.f);
     afpvec4 sum3 = afpvec4(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = gz * psc(c) * psc(outc) + gy * psc(c);
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = gz * c * outc + gy * c;
 
 #if NCNN_shader_local_memory
     const int lx = int(gl_LocalInvocationID.x);
     const int ly = int(gl_LocalInvocationID.y);
 
     int z = 0;
-    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
+    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < c; z += LOCAL_MEMORY_UNROLL_INCH)
     {
         if (ly < 4)
         {
             for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(outw) + ly));
+                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(cstep) + ly));
             }
         }
 
@@ -93,21 +89,21 @@ void main()
             sum3 += v3 * k;
         }
 
-        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(outw);
+        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
         w_offset += LOCAL_MEMORY_UNROLL_INCH;
 
         barrier();
     }
 
-    if (z < psc(c))
+    if (z < c)
     {
-        const int remain = psc(c) - z;
+        const int remain = c - z;
 
         if (ly < 4)
         {
             for (int z4 = 0; z4 < remain; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(outw) + ly));
+                tmp_v[lx][z4][ly] = sfp2lfp(buffer_ld1(bottom_tm_blob_data, v_offset + z4 * psc(cstep) + ly));
             }
         }
 
@@ -137,7 +133,7 @@ void main()
         }
     }
 #else
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afp v0 = buffer_ld1(bottom_tm_blob_data, v_offset + 0);
         afp v1 = buffer_ld1(bottom_tm_blob_data, v_offset + 1);
@@ -151,17 +147,17 @@ void main()
         sum2 += v2 * k;
         sum3 += v3 * k;
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 1;
     }
 #endif
 
 #if NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st4(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack1to8_3x3s1d1_winograd_gemm.comp

@@ -4,14 +4,13 @@
 #version 450
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfp bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec8 top_tm_blob_data[]; };
@@ -19,11 +18,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec8 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -33,7 +29,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 
     afpvec8 sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));
@@ -41,10 +37,10 @@ void main()
     afpvec8 sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));
     afpvec8 sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = gz * psc(c) * psc(outc) + gy * psc(c);
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = gz * c * outc + gy * c;
 
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afp v0 = buffer_ld1(bottom_tm_blob_data, v_offset + 0);
         afp v1 = buffer_ld1(bottom_tm_blob_data, v_offset + 1);
@@ -66,11 +62,11 @@ void main()
         sum3[0] += v3 * k[0];
         sum3[1] += v3 * k[1];
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 1;
     }
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st8(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st8(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd23_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfpvec4 bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 4x4
@@ -115,22 +115,22 @@ void main()
     v33 = m33 - m31;
 
     // store 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v10);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v11);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v12);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v13);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v20);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v21);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v22);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v23);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v30);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v31);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v32);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v33);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v10);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v11);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v12);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v13);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v20);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v21);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v22);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v23);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v30);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v31);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v32);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v33);
 }

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd23_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,28 +44,28 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afpvec4 v00 = buffer_ld4(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afpvec4 v01 = buffer_ld4(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afpvec4 v02 = buffer_ld4(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afpvec4 v03 = buffer_ld4(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afpvec4 v10 = buffer_ld4(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afpvec4 v11 = buffer_ld4(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afpvec4 v12 = buffer_ld4(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afpvec4 v13 = buffer_ld4(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afpvec4 v20 = buffer_ld4(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afpvec4 v21 = buffer_ld4(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afpvec4 v22 = buffer_ld4(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afpvec4 v23 = buffer_ld4(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afpvec4 v30 = buffer_ld4(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afpvec4 v31 = buffer_ld4(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afpvec4 v32 = buffer_ld4(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afpvec4 v33 = buffer_ld4(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afpvec4 v00 = buffer_ld4(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afpvec4 v01 = buffer_ld4(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afpvec4 v02 = buffer_ld4(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afpvec4 v03 = buffer_ld4(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afpvec4 v10 = buffer_ld4(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afpvec4 v11 = buffer_ld4(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afpvec4 v12 = buffer_ld4(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afpvec4 v13 = buffer_ld4(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afpvec4 v20 = buffer_ld4(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afpvec4 v21 = buffer_ld4(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afpvec4 v22 = buffer_ld4(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afpvec4 v23 = buffer_ld4(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afpvec4 v30 = buffer_ld4(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afpvec4 v31 = buffer_ld4(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afpvec4 v32 = buffer_ld4(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afpvec4 v33 = buffer_ld4(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
 
     // const float itm[2][4] = {
     //     {1.0f,  1.0f,  1.0f,  0.0f},

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd43_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfpvec4 bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 6x6
@@ -187,42 +187,42 @@ void main()
     v55 = m51 - m53 * afp(2.5) + m55;
 
     // store 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v04);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v05);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v10);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v11);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v12);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v13);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v14);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v15);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v20);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v21);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v22);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v23);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep), v24);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep), v25);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep), v30);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep), v31);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep), v32);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep), v33);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep), v34);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep), v35);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep), v40);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep), v41);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep), v42);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep), v43);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep), v44);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep), v45);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep), v50);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep), v51);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep), v52);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep), v53);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep), v54);
-    buffer_st4(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep), v55);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v04);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v05);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v10);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v11);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v12);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v13);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v14);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v15);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v20);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v21);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v22);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v23);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep) * c, v24);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep) * c, v25);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep) * c, v30);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep) * c, v31);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep) * c, v32);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep) * c, v33);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep) * c, v34);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep) * c, v35);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep) * c, v40);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep) * c, v41);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep) * c, v42);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep) * c, v43);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep) * c, v44);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep) * c, v45);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep) * c, v50);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep) * c, v51);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep) * c, v52);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep) * c, v53);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep) * c, v54);
+    buffer_st4(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep) * c, v55);
 }

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd43_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfpvec4 bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,48 +44,48 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afpvec4 v00 = buffer_ld4(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afpvec4 v01 = buffer_ld4(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afpvec4 v02 = buffer_ld4(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afpvec4 v03 = buffer_ld4(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afpvec4 v04 = buffer_ld4(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afpvec4 v05 = buffer_ld4(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afpvec4 v10 = buffer_ld4(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afpvec4 v11 = buffer_ld4(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afpvec4 v12 = buffer_ld4(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afpvec4 v13 = buffer_ld4(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afpvec4 v14 = buffer_ld4(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afpvec4 v15 = buffer_ld4(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afpvec4 v20 = buffer_ld4(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afpvec4 v21 = buffer_ld4(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afpvec4 v22 = buffer_ld4(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afpvec4 v23 = buffer_ld4(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
-    afpvec4 v24 = buffer_ld4(top_tm_blob_data, v_tm_offset + 16 * psc(cstep));
-    afpvec4 v25 = buffer_ld4(top_tm_blob_data, v_tm_offset + 17 * psc(cstep));
-    afpvec4 v30 = buffer_ld4(top_tm_blob_data, v_tm_offset + 18 * psc(cstep));
-    afpvec4 v31 = buffer_ld4(top_tm_blob_data, v_tm_offset + 19 * psc(cstep));
-    afpvec4 v32 = buffer_ld4(top_tm_blob_data, v_tm_offset + 20 * psc(cstep));
-    afpvec4 v33 = buffer_ld4(top_tm_blob_data, v_tm_offset + 21 * psc(cstep));
-    afpvec4 v34 = buffer_ld4(top_tm_blob_data, v_tm_offset + 22 * psc(cstep));
-    afpvec4 v35 = buffer_ld4(top_tm_blob_data, v_tm_offset + 23 * psc(cstep));
-    afpvec4 v40 = buffer_ld4(top_tm_blob_data, v_tm_offset + 24 * psc(cstep));
-    afpvec4 v41 = buffer_ld4(top_tm_blob_data, v_tm_offset + 25 * psc(cstep));
-    afpvec4 v42 = buffer_ld4(top_tm_blob_data, v_tm_offset + 26 * psc(cstep));
-    afpvec4 v43 = buffer_ld4(top_tm_blob_data, v_tm_offset + 27 * psc(cstep));
-    afpvec4 v44 = buffer_ld4(top_tm_blob_data, v_tm_offset + 28 * psc(cstep));
-    afpvec4 v45 = buffer_ld4(top_tm_blob_data, v_tm_offset + 29 * psc(cstep));
-    afpvec4 v50 = buffer_ld4(top_tm_blob_data, v_tm_offset + 30 * psc(cstep));
-    afpvec4 v51 = buffer_ld4(top_tm_blob_data, v_tm_offset + 31 * psc(cstep));
-    afpvec4 v52 = buffer_ld4(top_tm_blob_data, v_tm_offset + 32 * psc(cstep));
-    afpvec4 v53 = buffer_ld4(top_tm_blob_data, v_tm_offset + 33 * psc(cstep));
-    afpvec4 v54 = buffer_ld4(top_tm_blob_data, v_tm_offset + 34 * psc(cstep));
-    afpvec4 v55 = buffer_ld4(top_tm_blob_data, v_tm_offset + 35 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afpvec4 v00 = buffer_ld4(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afpvec4 v01 = buffer_ld4(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afpvec4 v02 = buffer_ld4(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afpvec4 v03 = buffer_ld4(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afpvec4 v04 = buffer_ld4(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afpvec4 v05 = buffer_ld4(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afpvec4 v10 = buffer_ld4(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afpvec4 v11 = buffer_ld4(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afpvec4 v12 = buffer_ld4(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afpvec4 v13 = buffer_ld4(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afpvec4 v14 = buffer_ld4(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afpvec4 v15 = buffer_ld4(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afpvec4 v20 = buffer_ld4(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afpvec4 v21 = buffer_ld4(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afpvec4 v22 = buffer_ld4(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afpvec4 v23 = buffer_ld4(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
+    afpvec4 v24 = buffer_ld4(top_tm_blob_data, v_tm_offset + 16 * psc(cstep) * c);
+    afpvec4 v25 = buffer_ld4(top_tm_blob_data, v_tm_offset + 17 * psc(cstep) * c);
+    afpvec4 v30 = buffer_ld4(top_tm_blob_data, v_tm_offset + 18 * psc(cstep) * c);
+    afpvec4 v31 = buffer_ld4(top_tm_blob_data, v_tm_offset + 19 * psc(cstep) * c);
+    afpvec4 v32 = buffer_ld4(top_tm_blob_data, v_tm_offset + 20 * psc(cstep) * c);
+    afpvec4 v33 = buffer_ld4(top_tm_blob_data, v_tm_offset + 21 * psc(cstep) * c);
+    afpvec4 v34 = buffer_ld4(top_tm_blob_data, v_tm_offset + 22 * psc(cstep) * c);
+    afpvec4 v35 = buffer_ld4(top_tm_blob_data, v_tm_offset + 23 * psc(cstep) * c);
+    afpvec4 v40 = buffer_ld4(top_tm_blob_data, v_tm_offset + 24 * psc(cstep) * c);
+    afpvec4 v41 = buffer_ld4(top_tm_blob_data, v_tm_offset + 25 * psc(cstep) * c);
+    afpvec4 v42 = buffer_ld4(top_tm_blob_data, v_tm_offset + 26 * psc(cstep) * c);
+    afpvec4 v43 = buffer_ld4(top_tm_blob_data, v_tm_offset + 27 * psc(cstep) * c);
+    afpvec4 v44 = buffer_ld4(top_tm_blob_data, v_tm_offset + 28 * psc(cstep) * c);
+    afpvec4 v45 = buffer_ld4(top_tm_blob_data, v_tm_offset + 29 * psc(cstep) * c);
+    afpvec4 v50 = buffer_ld4(top_tm_blob_data, v_tm_offset + 30 * psc(cstep) * c);
+    afpvec4 v51 = buffer_ld4(top_tm_blob_data, v_tm_offset + 31 * psc(cstep) * c);
+    afpvec4 v52 = buffer_ld4(top_tm_blob_data, v_tm_offset + 32 * psc(cstep) * c);
+    afpvec4 v53 = buffer_ld4(top_tm_blob_data, v_tm_offset + 33 * psc(cstep) * c);
+    afpvec4 v54 = buffer_ld4(top_tm_blob_data, v_tm_offset + 34 * psc(cstep) * c);
+    afpvec4 v55 = buffer_ld4(top_tm_blob_data, v_tm_offset + 35 * psc(cstep) * c);
 
 #define sq2 1.41421356237
 #define sq2_m2 1.41421356237*2

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd_gemm.comp

@@ -6,14 +6,13 @@
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec4 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
@@ -21,11 +20,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec4 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -41,7 +37,7 @@ void main()
     int gz = int(gl_GlobalInvocationID.z);
 
 #if !NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
@@ -50,21 +46,21 @@ void main()
     afpvec4 sum2 = afpvec4(0.f);
     afpvec4 sum3 = afpvec4(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = (gz * psc(c) * psc(outc) + gy * psc(c)) * 4;
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = (gz * c * outc + gy * c) * 4;
 
 #if NCNN_shader_local_memory
     const int lx = int(gl_LocalInvocationID.x);
     const int ly = int(gl_LocalInvocationID.y);
 
     int z = 0;
-    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
+    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < c; z += LOCAL_MEMORY_UNROLL_INCH)
     {
         if (ly < 4)
         {
             for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(outw) + ly]);
+                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(cstep) + ly]);
             }
         }
 
@@ -98,21 +94,21 @@ void main()
             sum3 += v3 * k;
         }
 
-        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(outw);
+        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
         w_offset += LOCAL_MEMORY_UNROLL_INCH * 4;
 
         barrier();
     }
 
-    if (z < psc(c))
+    if (z < c)
     {
-        const int remain = psc(c) - z;
+        const int remain = c - z;
 
         if (ly < 4)
         {
             for (int z4 = 0; z4 < remain; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(outw) + ly]);
+                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(cstep) + ly]);
             }
         }
 
@@ -147,7 +143,7 @@ void main()
         }
     }
 #else
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec4 v0 = buffer_ld4(bottom_tm_blob_data, v_offset + 0);
         afpvec4 v1 = buffer_ld4(bottom_tm_blob_data, v_offset + 1);
@@ -166,17 +162,17 @@ void main()
         sum2 += v2 * k;
         sum3 += v3 * k;
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 4;
     }
 #endif
 
 #if NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st4(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd_gemm_cm_16_16_16.comp

@@ -1,258 +0,0 @@
-// Copyright 2023 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#extension GL_KHR_memory_scope_semantics: require
-#extension GL_EXT_shader_explicit_arithmetic_types: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#if ncnn_VK_KHR_cooperative_matrix
-#extension GL_KHR_cooperative_matrix: require
-#elif ncnn_VK_NV_cooperative_matrix
-#extension GL_NV_cooperative_matrix: require
-#endif
-
-layout (constant_id = 0) const int batch = 1;
-
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
-
-layout (binding = 0) readonly buffer bottom_tm_blob { uvec2 bottom_tm_blob_data[]; };
-layout (binding = 1) writeonly buffer top_tm_blob { uvec2 top_tm_blob_data[]; };
-layout (binding = 2) readonly buffer weight_tm_blob { uvec2 weight_tm_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int c;
-    int cstep;
-
-    int outw;
-    int outc;
-    int outcstep;
-} p;
-
-#define UNROLL_INCH 2
-
-shared uvec2 tmp_v0[UNROLL_INCH * 16*4];
-shared uvec2 tmp_v1[UNROLL_INCH * 16*4];
-shared uvec2 tmp_k0[UNROLL_INCH * 16*4];
-shared uvec2 tmp_k1[UNROLL_INCH * 16*4];
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x) / 32 * 2 * 16;
-    int gy = int(gl_GlobalInvocationID.y) * 2 * 4;
-    int gz = int(gl_GlobalInvocationID.z);
-
-    const int lx = int(gl_LocalInvocationID.x);
-
-    const int lxd16 = lx / 16; // 0 1
-    const int lxm16 = lx % 16; // 0 1 2 3 .... 15
-
-#if ncnn_VK_KHR_cooperative_matrix
-    coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum0 = coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum1 = coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum2 = coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum3 = coopmat<afp, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(0.f);
-#elif ncnn_VK_NV_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum0 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum1 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum2 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum3 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(0.f);
-#else
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 16> sum0 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 16> sum1 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 16> sum2 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 16>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 16> sum3 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 16>(0.f);
-#endif
-#endif
-
-    const int N = psc(c) / 4;
-
-    int z = 0;
-    for (; z + (UNROLL_INCH - 1) < N; z += UNROLL_INCH)
-    {
-        {
-            for (int j = 0; j < 4; j++)
-            {
-                const int tmp_i = lxd16*16*4 + lxm16 * 4 + j;
-
-                const int v_offset = gz * psc(cstep) + ((z + lxd16) * 4 + j) * psc(outw) + (gx + lxm16);
-
-                tmp_v0[tmp_i] = (gx + lxm16) < psc(outw) ? bottom_tm_blob_data[v_offset] : uvec2(0);
-                tmp_v1[tmp_i] = (gx + lxm16 + 16) < psc(outw) ? bottom_tm_blob_data[v_offset + 16] : uvec2(0);
-
-                const int w_offset = gz * psc(outc) * psc(c) * 4 + gy * psc(c) * 4 + (z + lxd16) * 4 * 16 + (lxm16 * 4 + j);
-
-                tmp_k0[tmp_i] = weight_tm_data[w_offset];
-                tmp_k1[tmp_i] = weight_tm_data[w_offset + psc(c) * 16];
-            }
-        }
-
-        barrier();
-
-        for (int z4 = 0; z4 < UNROLL_INCH; z4++)
-        {
-#if ncnn_VK_KHR_cooperative_matrix
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> A0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> A1;
-            coopMatLoad(A0, tmp_v0, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(A1, tmp_v1, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseB> B0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseB> B1;
-            coopMatLoad(B0, tmp_k0, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B1, tmp_k1, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-
-            // sum += v * k
-            sum0 = coopMatMulAdd(A0, B0, sum0);
-            sum1 = coopMatMulAdd(A1, B0, sum1);
-            sum2 = coopMatMulAdd(A0, B1, sum2);
-            sum3 = coopMatMulAdd(A1, B1, sum3);
-#elif ncnn_VK_NV_cooperative_matrix
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> A0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> A1;
-            coopMatLoadNV(A0, tmp_v0, z4*16*4, 4, false);
-            coopMatLoadNV(A1, tmp_v1, z4*16*4, 4, false);
-
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> B0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> B1;
-            coopMatLoadNV(B0, tmp_k0, z4*16*4, 4, false);
-            coopMatLoadNV(B1, tmp_k1, z4*16*4, 4, false);
-
-            // sum += v * k
-            sum0 = coopMatMulAddNV(A0, B0, sum0);
-            sum1 = coopMatMulAddNV(A1, B0, sum1);
-            sum2 = coopMatMulAddNV(A0, B1, sum2);
-            sum3 = coopMatMulAddNV(A1, B1, sum3);
-#endif
-        }
-
-        barrier();
-    }
-
-    if (z < N)
-    {
-        const int remain = N - z;
-
-        if (lxd16 == 0)
-        {
-            for (int j = 0; j < 4; j++)
-            {
-                const int tmp_i = lxd16*16*4 + lxm16 * 4 + j;
-
-                const int v_offset = gz * psc(cstep) + ((z + lxd16) * 4 + j) * psc(outw) + (gx + lxm16);
-
-                tmp_v0[tmp_i] = (gx + lxm16) < psc(outw) ? bottom_tm_blob_data[v_offset] : uvec2(0);
-                tmp_v1[tmp_i] = (gx + lxm16 + 16) < psc(outw) ? bottom_tm_blob_data[v_offset + 16] : uvec2(0);
-
-                const int w_offset = gz * psc(outc) * psc(c) * 4 + gy * psc(c) * 4 + (z + lxd16) * 4 * 16 + (lxm16 * 4 + j);
-
-                tmp_k0[tmp_i] = weight_tm_data[w_offset];
-                tmp_k1[tmp_i] = weight_tm_data[w_offset + psc(c) * 16];
-            }
-        }
-
-        barrier();
-
-        for (int z4 = 0; z4 < remain; z4++)
-        {
-#if ncnn_VK_KHR_cooperative_matrix
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> A0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> A1;
-            coopMatLoad(A0, tmp_v0, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(A1, tmp_v1, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseB> B0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseB> B1;
-            coopMatLoad(B0, tmp_k0, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B1, tmp_k1, z4*16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-
-            // sum += v * k
-            sum0 = coopMatMulAdd(A0, B0, sum0);
-            sum1 = coopMatMulAdd(A1, B0, sum1);
-            sum2 = coopMatMulAdd(A0, B1, sum2);
-            sum3 = coopMatMulAdd(A1, B1, sum3);
-#elif ncnn_VK_NV_cooperative_matrix
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> A0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> A1;
-            coopMatLoadNV(A0, tmp_v0, z4*16*4, 4, false);
-            coopMatLoadNV(A1, tmp_v1, z4*16*4, 4, false);
-
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> B0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> B1;
-            coopMatLoadNV(B0, tmp_k0, z4*16*4, 4, false);
-            coopMatLoadNV(B1, tmp_k1, z4*16*4, 4, false);
-
-            // sum += v * k
-            sum0 = coopMatMulAddNV(A0, B0, sum0);
-            sum1 = coopMatMulAddNV(A1, B0, sum1);
-            sum2 = coopMatMulAddNV(A0, B1, sum2);
-            sum3 = coopMatMulAddNV(A1, B1, sum3);
-#endif
-        }
-
-        barrier();
-    }
-
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
-        return;
-
-#if ncnn_VK_KHR_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    coopMatStore(sum0, tmp_v0, 0, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum1, tmp_v1, 0, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum2, tmp_v0, 16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum3, tmp_v1, 16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-#else
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum0_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(sum0);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum1_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(sum1);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum2_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(sum2);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator> sum3_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>(sum3);
-
-    coopMatStore(sum0_fp16, tmp_v0, 0, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum1_fp16, tmp_v1, 0, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum2_fp16, tmp_v0, 16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum3_fp16, tmp_v1, 16*4, 4, gl_CooperativeMatrixLayoutRowMajor);
-#endif
-#elif ncnn_VK_NV_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    coopMatStoreNV(sum0, tmp_v0, 0, 4, false);
-    coopMatStoreNV(sum1, tmp_v1, 0, 4, false);
-    coopMatStoreNV(sum2, tmp_v0, 16*4, 4, false);
-    coopMatStoreNV(sum3, tmp_v1, 16*4, 4, false);
-#else
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum0_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(sum0);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum1_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(sum1);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum2_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(sum2);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 16> sum3_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 16>(sum3);
-
-    coopMatStoreNV(sum0_fp16, tmp_v0, 0, 4, false);
-    coopMatStoreNV(sum1_fp16, tmp_v1, 0, 4, false);
-    coopMatStoreNV(sum2_fp16, tmp_v0, 16*4, 4, false);
-    coopMatStoreNV(sum3_fp16, tmp_v1, 16*4, 4, false);
-#endif
-#endif
-
-    barrier();
-
-    {
-        for (int j = 0; j < 4; j++)
-        {
-            const int tmp_vi = lxm16 * 4 + j + lxd16*16*4;
-            const int gi = gz * psc(outcstep) + (gy + lxd16 * 4 + j) * psc(outw) + (gx + lxm16);
-
-            if (gy + lxd16 * 4 + j < psc(outc))
-            {
-                if (gx + lxm16 < psc(outw)) top_tm_blob_data[gi] = tmp_v0[tmp_vi];
-                if (gx + lxm16 + 16 < psc(outw)) top_tm_blob_data[gi + 16] = tmp_v1[tmp_vi];
-            }
-        }
-    }
-}

src/layer/vulkan/shader/convolution_pack4_3x3s1d1_winograd_gemm_cm_16_8_8.comp

@@ -1,343 +0,0 @@
-// Copyright 2023 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#extension GL_KHR_memory_scope_semantics: require
-#extension GL_EXT_shader_explicit_arithmetic_types: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#if ncnn_VK_KHR_cooperative_matrix
-#extension GL_KHR_cooperative_matrix: require
-#elif ncnn_VK_NV_cooperative_matrix
-#extension GL_NV_cooperative_matrix: require
-#endif
-
-layout (constant_id = 0) const int batch = 1;
-
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
-
-layout (binding = 0) readonly buffer bottom_tm_blob { uvec2 bottom_tm_blob_data[]; };
-layout (binding = 1) writeonly buffer top_tm_blob { uvec2 top_tm_blob_data[]; };
-layout (binding = 2) readonly buffer weight_tm_blob { uvec2 weight_tm_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int c;
-    int cstep;
-
-    int outw;
-    int outc;
-    int outcstep;
-} p;
-
-#define UNROLL_INCH 4
-
-shared uvec2 tmp_v0[UNROLL_INCH * 16*2];
-shared uvec2 tmp_v1[UNROLL_INCH * 16*2];
-shared uvec2 tmp_k0[UNROLL_INCH * 8*2];
-shared uvec2 tmp_k1[UNROLL_INCH * 8*2];
-shared uvec2 tmp_k2[UNROLL_INCH * 8*2];
-shared uvec2 tmp_k3[UNROLL_INCH * 8*2];
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x) / 32 * 2 * 16;
-    int gy = int(gl_GlobalInvocationID.y) * 2 * 4;
-    int gz = int(gl_GlobalInvocationID.z);
-
-    const int lx = int(gl_LocalInvocationID.x);
-
-    const int lxd8 = lx / 8; // 0 1 2 3
-    const int lxm8 = lx % 8; // 0 1 2 3 .... 7
-
-#if ncnn_VK_KHR_cooperative_matrix
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum0 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum1 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum2 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum3 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum4 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum5 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum6 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-    coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum7 = coopmat<afp, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(0.f);
-#elif ncnn_VK_NV_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum0 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum1 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum2 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum3 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum4 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum5 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum6 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum7 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(0.f);
-#else
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum0 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum1 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum2 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum3 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum4 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum5 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum6 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-    fcoopmatNV<32, gl_ScopeSubgroup, 16, 8> sum7 = fcoopmatNV<32, gl_ScopeSubgroup, 16, 8>(0.f);
-#endif
-#endif
-
-    const int N = psc(c) / 2;
-
-    int z = 0;
-    for (; z + (UNROLL_INCH - 1) < N; z += UNROLL_INCH)
-    {
-        {
-            for (int j = 0; j < 2; j++)
-            {
-                const int tmp_vi = lxd8*16*2 + lxm8 * 2 + j;
-
-                int v_offset = gz * psc(cstep) + ((z + lxd8) * 2 + j) * psc(outw) + (gx + lxm8);
-
-                tmp_v0[tmp_vi] = (gx + lxm8) < psc(outw) ? bottom_tm_blob_data[v_offset] : uvec2(0);
-                tmp_v0[tmp_vi + 16] = (gx + lxm8 + 8) < psc(outw) ? bottom_tm_blob_data[v_offset + 8] : uvec2(0);
-                tmp_v1[tmp_vi] = (gx + lxm8 + 16) < psc(outw) ? bottom_tm_blob_data[v_offset + 16] : uvec2(0);
-                tmp_v1[tmp_vi + 16] = (gx + lxm8 + 24) < psc(outw) ? bottom_tm_blob_data[v_offset + 24] : uvec2(0);
-
-                const int tmp_ki = lxd8*8*2 + lxm8 * 2 + j;
-
-                int w_offset = gz * psc(outc) * psc(c) * 4 + gy * psc(c) * 4 + (z + lxd8) * 2 * 8 + (lxm8 * 2 + j);
-
-                tmp_k0[tmp_ki] = weight_tm_data[w_offset];
-                tmp_k1[tmp_ki] = weight_tm_data[w_offset + psc(c) * 8];
-                tmp_k2[tmp_ki] = weight_tm_data[w_offset + psc(c) * 16];
-                tmp_k3[tmp_ki] = weight_tm_data[w_offset + psc(c) * 24];
-            }
-        }
-
-        barrier();
-
-        for (int z4 = 0; z4 < UNROLL_INCH; z4++)
-        {
-#if ncnn_VK_KHR_cooperative_matrix
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseA> A0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseA> A1;
-            coopMatLoad(A0, tmp_v0, z4*16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(A1, tmp_v1, z4*16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B0;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B1;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B2;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B3;
-            coopMatLoad(B0, tmp_k0, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B1, tmp_k1, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B2, tmp_k2, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B3, tmp_k3, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-
-            // sum += v * k
-            sum0 = coopMatMulAdd(A0, B0, sum0);
-            sum1 = coopMatMulAdd(A1, B0, sum1);
-            sum2 = coopMatMulAdd(A0, B1, sum2);
-            sum3 = coopMatMulAdd(A1, B1, sum3);
-            sum4 = coopMatMulAdd(A0, B2, sum4);
-            sum5 = coopMatMulAdd(A1, B2, sum5);
-            sum6 = coopMatMulAdd(A0, B3, sum6);
-            sum7 = coopMatMulAdd(A1, B3, sum7);
-#elif ncnn_VK_NV_cooperative_matrix
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> A0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> A1;
-            coopMatLoadNV(A0, tmp_v0, z4*16*2, 2, false);
-            coopMatLoadNV(A1, tmp_v1, z4*16*2, 2, false);
-
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B1;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B2;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B3;
-            coopMatLoadNV(B0, tmp_k0, z4*8*2, 2, false);
-            coopMatLoadNV(B1, tmp_k1, z4*8*2, 2, false);
-            coopMatLoadNV(B2, tmp_k2, z4*8*2, 2, false);
-            coopMatLoadNV(B3, tmp_k3, z4*8*2, 2, false);
-
-            // sum += v * k
-            sum0 = coopMatMulAddNV(A0, B0, sum0);
-            sum1 = coopMatMulAddNV(A1, B0, sum1);
-            sum2 = coopMatMulAddNV(A0, B1, sum2);
-            sum3 = coopMatMulAddNV(A1, B1, sum3);
-            sum4 = coopMatMulAddNV(A0, B2, sum4);
-            sum5 = coopMatMulAddNV(A1, B2, sum5);
-            sum6 = coopMatMulAddNV(A0, B3, sum6);
-            sum7 = coopMatMulAddNV(A1, B3, sum7);
-#endif
-        }
-
-        barrier();
-    }
-
-    if (z < N)
-    {
-        const int remain = N - z;
-
-        if (lxd8 < remain)
-        {
-            for (int j = 0; j < 2; j++)
-            {
-                const int tmp_vi = lxd8*16*2 + lxm8 * 2 + j;
-
-                int v_offset = gz * psc(cstep) + ((z + lxd8) * 2 + j) * psc(outw) + (gx + lxm8);
-
-                tmp_v0[tmp_vi] = (gx + lxm8) < psc(outw) ? bottom_tm_blob_data[v_offset] : uvec2(0);
-                tmp_v0[tmp_vi + 16] = (gx + lxm8 + 8) < psc(outw) ? bottom_tm_blob_data[v_offset + 8] : uvec2(0);
-                tmp_v1[tmp_vi] = (gx + lxm8 + 16) < psc(outw) ? bottom_tm_blob_data[v_offset + 16] : uvec2(0);
-                tmp_v1[tmp_vi + 16] = (gx + lxm8 + 24) < psc(outw) ? bottom_tm_blob_data[v_offset + 24] : uvec2(0);
-
-                const int tmp_ki = lxd8*8*2 + lxm8 * 2 + j;
-
-                int w_offset = gz * psc(outc) * psc(c) * 4 + gy * psc(c) * 4 + (z + lxd8) * 2 * 8 + (lxm8 * 2 + j);
-
-                tmp_k0[tmp_ki] = weight_tm_data[w_offset];
-                tmp_k1[tmp_ki] = weight_tm_data[w_offset + psc(c) * 8];
-                tmp_k2[tmp_ki] = weight_tm_data[w_offset + psc(c) * 16];
-                tmp_k3[tmp_ki] = weight_tm_data[w_offset + psc(c) * 24];
-            }
-        }
-
-        barrier();
-
-        for (int z4 = 0; z4 < remain; z4++)
-        {
-#if ncnn_VK_KHR_cooperative_matrix
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseA> A0;
-            coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseA> A1;
-            coopMatLoad(A0, tmp_v0, z4*16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(A1, tmp_v1, z4*16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B0;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B1;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B2;
-            coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B3;
-            coopMatLoad(B0, tmp_k0, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B1, tmp_k1, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B2, tmp_k2, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-            coopMatLoad(B3, tmp_k3, z4*8*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-
-            // sum += v * k
-            sum0 = coopMatMulAdd(A0, B0, sum0);
-            sum1 = coopMatMulAdd(A1, B0, sum1);
-            sum2 = coopMatMulAdd(A0, B1, sum2);
-            sum3 = coopMatMulAdd(A1, B1, sum3);
-            sum4 = coopMatMulAdd(A0, B2, sum4);
-            sum5 = coopMatMulAdd(A1, B2, sum5);
-            sum6 = coopMatMulAdd(A0, B3, sum6);
-            sum7 = coopMatMulAdd(A1, B3, sum7);
-#elif ncnn_VK_NV_cooperative_matrix
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> A0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> A1;
-            coopMatLoadNV(A0, tmp_v0, z4*16*2, 2, false);
-            coopMatLoadNV(A1, tmp_v1, z4*16*2, 2, false);
-
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B0;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B1;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B2;
-            fcoopmatNV<16, gl_ScopeSubgroup, 8, 8> B3;
-            coopMatLoadNV(B0, tmp_k0, z4*8*2, 2, false);
-            coopMatLoadNV(B1, tmp_k1, z4*8*2, 2, false);
-            coopMatLoadNV(B2, tmp_k2, z4*8*2, 2, false);
-            coopMatLoadNV(B3, tmp_k3, z4*8*2, 2, false);
-
-            // sum += v * k
-            sum0 = coopMatMulAddNV(A0, B0, sum0);
-            sum1 = coopMatMulAddNV(A1, B0, sum1);
-            sum2 = coopMatMulAddNV(A0, B1, sum2);
-            sum3 = coopMatMulAddNV(A1, B1, sum3);
-            sum4 = coopMatMulAddNV(A0, B2, sum4);
-            sum5 = coopMatMulAddNV(A1, B2, sum5);
-            sum6 = coopMatMulAddNV(A0, B3, sum6);
-            sum7 = coopMatMulAddNV(A1, B3, sum7);
-#endif
-        }
-
-        barrier();
-    }
-
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
-        return;
-
-#if ncnn_VK_KHR_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    coopMatStore(sum0, tmp_v0, 0, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum1, tmp_v1, 0, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum2, tmp_v0, 16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum3, tmp_v1, 16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum4, tmp_v0, 16*4, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum5, tmp_v1, 16*4, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum6, tmp_v0, 16*6, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum7, tmp_v1, 16*6, 2, gl_CooperativeMatrixLayoutRowMajor);
-#else
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum0_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum0);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum1_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum1);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum2_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum2);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum3_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum3);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum4_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum4);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum5_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum5);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum6_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum6);
-    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> sum7_fp16 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(sum7);
-
-    coopMatStore(sum0_fp16, tmp_v0, 0, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum1_fp16, tmp_v1, 0, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum2_fp16, tmp_v0, 16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum3_fp16, tmp_v1, 16*2, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum4_fp16, tmp_v0, 16*4, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum5_fp16, tmp_v1, 16*4, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum6_fp16, tmp_v0, 16*6, 2, gl_CooperativeMatrixLayoutRowMajor);
-    coopMatStore(sum7_fp16, tmp_v1, 16*6, 2, gl_CooperativeMatrixLayoutRowMajor);
-#endif
-#elif ncnn_VK_NV_cooperative_matrix
-#if NCNN_fp16_arithmetic
-    coopMatStoreNV(sum0, tmp_v0, 0, 2, false);
-    coopMatStoreNV(sum1, tmp_v1, 0, 2, false);
-    coopMatStoreNV(sum2, tmp_v0, 16*2, 2, false);
-    coopMatStoreNV(sum3, tmp_v1, 16*2, 2, false);
-    coopMatStoreNV(sum4, tmp_v0, 16*4, 2, false);
-    coopMatStoreNV(sum5, tmp_v1, 16*4, 2, false);
-    coopMatStoreNV(sum6, tmp_v0, 16*6, 2, false);
-    coopMatStoreNV(sum7, tmp_v1, 16*6, 2, false);
-#else
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum0_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum0);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum1_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum1);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum2_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum2);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum3_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum3);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum4_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum4);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum5_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum5);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum6_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum6);
-    fcoopmatNV<16, gl_ScopeSubgroup, 16, 8> sum7_fp16 = fcoopmatNV<16, gl_ScopeSubgroup, 16, 8>(sum7);
-
-    coopMatStoreNV(sum0_fp16, tmp_v0, 0, 2, false);
-    coopMatStoreNV(sum1_fp16, tmp_v1, 0, 2, false);
-    coopMatStoreNV(sum2_fp16, tmp_v0, 16*2, 2, false);
-    coopMatStoreNV(sum3_fp16, tmp_v1, 16*2, 2, false);
-    coopMatStoreNV(sum4_fp16, tmp_v0, 16*4, 2, false);
-    coopMatStoreNV(sum5_fp16, tmp_v1, 16*4, 2, false);
-    coopMatStoreNV(sum6_fp16, tmp_v0, 16*6, 2, false);
-    coopMatStoreNV(sum7_fp16, tmp_v1, 16*6, 2, false);
-#endif
-#endif
-
-    barrier();
-
-    const int lxd16 = lx / 16; // 0 1
-    const int lxm16 = lx % 16; // 0 1 2 3 .... 15
-
-    {
-        for (int j = 0; j < 4; j++)
-        {
-            const int tmp_vi = lxm16 * 2 + lxd16 + j*16*2;
-            const int gi = gz * psc(outcstep) + (gy + lxd16 + j*2) * psc(outw) + (gx + lxm16);
-
-            if (gy + j * 2 + lxd16 < psc(outc))
-            {
-                if (gx + lxm16 < psc(outw)) top_tm_blob_data[gi] = tmp_v0[tmp_vi];
-                if (gx + lxm16 + 16 < psc(outw)) top_tm_blob_data[gi + 16] = tmp_v1[tmp_vi];
-            }
-        }
-    }
-}

src/layer/vulkan/shader/convolution_pack4to1_3x3s1d1_winograd_gemm.comp

@@ -6,14 +6,13 @@
 #define LOCAL_MEMORY_UNROLL_INCH 8
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec4 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfp top_tm_blob_data[]; };
@@ -21,11 +20,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec4 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -41,7 +37,7 @@ void main()
     int gz = int(gl_GlobalInvocationID.z);
 
 #if !NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
@@ -50,21 +46,21 @@ void main()
     afp sum2 = afp(0.f);
     afp sum3 = afp(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = gz * psc(c) * psc(outc) + gy * psc(c);
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = gz * c * outc + gy * c;
 
 #if NCNN_shader_local_memory
     const int lx = int(gl_LocalInvocationID.x);
     const int ly = int(gl_LocalInvocationID.y);
 
     int z = 0;
-    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
+    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < c; z += LOCAL_MEMORY_UNROLL_INCH)
     {
         if (ly < 4)
         {
             for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(outw) + ly]);
+                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(cstep) + ly]);
             }
         }
 
@@ -93,21 +89,21 @@ void main()
             sum3 += dot(v3, k);
         }
 
-        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(outw);
+        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
         w_offset += LOCAL_MEMORY_UNROLL_INCH;
 
         barrier();
     }
 
-    if (z < psc(c))
+    if (z < c)
     {
-        const int remain = psc(c) - z;
+        const int remain = c - z;
 
         if (ly < 4)
         {
             for (int z4 = 0; z4 < remain; z4++)
             {
-                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(outw) + ly]);
+                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_tm_blob_data[v_offset + z4 * psc(cstep) + ly]);
             }
         }
 
@@ -137,7 +133,7 @@ void main()
         }
     }
 #else
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec4 v0 = buffer_ld4(bottom_tm_blob_data, v_offset + 0);
         afpvec4 v1 = buffer_ld4(bottom_tm_blob_data, v_offset + 1);
@@ -151,17 +147,17 @@ void main()
         sum2 += dot(v2, k);
         sum3 += dot(v3, k);
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 1;
     }
 #endif
 
 #if NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 #endif
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st1(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st1(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack4to8_3x3s1d1_winograd_gemm.comp

@@ -4,14 +4,13 @@
 #version 450
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec4 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec8 top_tm_blob_data[]; };
@@ -19,11 +18,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec4 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -33,7 +29,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 
     afpvec8 sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));
@@ -41,10 +37,10 @@ void main()
     afpvec8 sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));
     afpvec8 sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = (gz * psc(c) * psc(outc) + gy * psc(c)) * 8;
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = (gz * c * outc + gy * c) * 8;
 
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec4 v0 = buffer_ld4(bottom_tm_blob_data, v_offset + 0);
         afpvec4 v1 = buffer_ld4(bottom_tm_blob_data, v_offset + 1);
@@ -97,11 +93,11 @@ void main()
         sum3[1].b += dot(v3, k6);
         sum3[1].a += dot(v3, k7);
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 8;
     }
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st8(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st8(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd23_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 4x4
@@ -115,22 +115,22 @@ void main()
     v33 = m33 - m31;
 
     // store 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v10);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v11);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v12);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v13);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v20);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v21);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v22);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v23);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v30);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v31);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v32);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v33);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v10);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v11);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v12);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v13);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v20);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v21);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v22);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v23);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v30);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v31);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v32);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v33);
 }

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd23_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfpvec8 top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfpvec8 bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,28 +44,28 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= psc(c))
+    if (gx >= psc(block_x) || gy >= psc(block_y) || gz >= c)
         return;
 
     // load 16
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afpvec8 v00 = buffer_ld8(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afpvec8 v01 = buffer_ld8(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afpvec8 v02 = buffer_ld8(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afpvec8 v03 = buffer_ld8(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afpvec8 v10 = buffer_ld8(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afpvec8 v11 = buffer_ld8(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afpvec8 v12 = buffer_ld8(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afpvec8 v13 = buffer_ld8(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afpvec8 v20 = buffer_ld8(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afpvec8 v21 = buffer_ld8(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afpvec8 v22 = buffer_ld8(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afpvec8 v23 = buffer_ld8(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afpvec8 v30 = buffer_ld8(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afpvec8 v31 = buffer_ld8(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afpvec8 v32 = buffer_ld8(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afpvec8 v33 = buffer_ld8(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afpvec8 v00 = buffer_ld8(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afpvec8 v01 = buffer_ld8(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afpvec8 v02 = buffer_ld8(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afpvec8 v03 = buffer_ld8(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afpvec8 v10 = buffer_ld8(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afpvec8 v11 = buffer_ld8(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afpvec8 v12 = buffer_ld8(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afpvec8 v13 = buffer_ld8(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afpvec8 v20 = buffer_ld8(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afpvec8 v21 = buffer_ld8(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afpvec8 v22 = buffer_ld8(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afpvec8 v23 = buffer_ld8(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afpvec8 v30 = buffer_ld8(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afpvec8 v31 = buffer_ld8(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afpvec8 v32 = buffer_ld8(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afpvec8 v33 = buffer_ld8(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
 
     // const float itm[2][4] = {
     //     {1.0f,  1.0f,  1.0f,  0.0f},

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd43_transform_input.comp

@@ -3,16 +3,17 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
+layout (constant_id = 0) const int c = 0;
+
+#define shape_constant_id_offset 1
 layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
 layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outcstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 5) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int block_y = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
@@ -21,7 +22,6 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
-    int c;
     int cstep;
 
     int outcstep;
@@ -36,7 +36,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.block_x || gy >= p.block_y || gz >= psc(c))
+    if (gx >= p.block_x || gy >= p.block_y || gz >= c)
         return;
 
     // load 6x6
@@ -187,42 +187,42 @@ void main()
     v55 = m51 - m53 * afp(2.5) + m55;
 
     // store 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep), v00);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep), v01);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep), v02);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep), v03);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep), v04);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep), v05);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep), v10);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep), v11);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep), v12);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep), v13);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep), v14);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep), v15);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep), v20);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep), v21);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep), v22);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep), v23);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep), v24);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep), v25);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep), v30);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep), v31);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep), v32);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep), v33);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep), v34);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep), v35);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep), v40);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep), v41);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep), v42);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep), v43);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep), v44);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep), v45);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep), v50);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep), v51);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep), v52);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep), v53);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep), v54);
-    buffer_st8(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep), v55);
+    int v_tm_offset = gz * psc(outcstep) + gy * psc(block_x) + gx;
+
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 0 * psc(outcstep) * c, v00);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 1 * psc(outcstep) * c, v01);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 2 * psc(outcstep) * c, v02);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 3 * psc(outcstep) * c, v03);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 4 * psc(outcstep) * c, v04);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 5 * psc(outcstep) * c, v05);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 6 * psc(outcstep) * c, v10);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 7 * psc(outcstep) * c, v11);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 8 * psc(outcstep) * c, v12);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 9 * psc(outcstep) * c, v13);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 10 * psc(outcstep) * c, v14);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 11 * psc(outcstep) * c, v15);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 12 * psc(outcstep) * c, v20);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 13 * psc(outcstep) * c, v21);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 14 * psc(outcstep) * c, v22);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 15 * psc(outcstep) * c, v23);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 16 * psc(outcstep) * c, v24);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 17 * psc(outcstep) * c, v25);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 18 * psc(outcstep) * c, v30);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 19 * psc(outcstep) * c, v31);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 20 * psc(outcstep) * c, v32);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 21 * psc(outcstep) * c, v33);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 22 * psc(outcstep) * c, v34);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 23 * psc(outcstep) * c, v35);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 24 * psc(outcstep) * c, v40);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 25 * psc(outcstep) * c, v41);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 26 * psc(outcstep) * c, v42);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 27 * psc(outcstep) * c, v43);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 28 * psc(outcstep) * c, v44);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 29 * psc(outcstep) * c, v45);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 30 * psc(outcstep) * c, v50);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 31 * psc(outcstep) * c, v51);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 32 * psc(outcstep) * c, v52);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 33 * psc(outcstep) * c, v53);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 34 * psc(outcstep) * c, v54);
+    buffer_st8(bottom_tm_blob_data, v_tm_offset + 35 * psc(outcstep) * c, v55);
 }

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd43_transform_output.comp

@@ -10,17 +10,17 @@ layout (constant_id = 0) const int bias_term = 0;
 layout (constant_id = 1) const int activation_type = 0;
 layout (constant_id = 2) const float activation_param_0 = 0;
 layout (constant_id = 3) const float activation_param_1 = 0;
+layout (constant_id = 4) const int c = 0;
 
-#define shape_constant_id_offset 4
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
+#define shape_constant_id_offset 5
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int block_x = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int block_y = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int block_x = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int block_y = 0;
 
-layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer top_tm_blob { sfpvec8 top_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
@@ -28,7 +28,6 @@ layout (binding = 2) readonly buffer bias_blob { sfpvec8 bias_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
 
     int block_x;
@@ -45,48 +44,48 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= p.block_x || gy >= p.block_y || gz >= psc(c))
+    if (gx >= p.block_x || gy >= p.block_y || gz >= c)
         return;
 
     // load 36
-    int v_tm_offset = gz * psc(block_x) * psc(block_y) + gy * psc(block_x) + gx;
-
-    afpvec8 v00 = buffer_ld8(top_tm_blob_data, v_tm_offset + 0 * psc(cstep));
-    afpvec8 v01 = buffer_ld8(top_tm_blob_data, v_tm_offset + 1 * psc(cstep));
-    afpvec8 v02 = buffer_ld8(top_tm_blob_data, v_tm_offset + 2 * psc(cstep));
-    afpvec8 v03 = buffer_ld8(top_tm_blob_data, v_tm_offset + 3 * psc(cstep));
-    afpvec8 v04 = buffer_ld8(top_tm_blob_data, v_tm_offset + 4 * psc(cstep));
-    afpvec8 v05 = buffer_ld8(top_tm_blob_data, v_tm_offset + 5 * psc(cstep));
-    afpvec8 v10 = buffer_ld8(top_tm_blob_data, v_tm_offset + 6 * psc(cstep));
-    afpvec8 v11 = buffer_ld8(top_tm_blob_data, v_tm_offset + 7 * psc(cstep));
-    afpvec8 v12 = buffer_ld8(top_tm_blob_data, v_tm_offset + 8 * psc(cstep));
-    afpvec8 v13 = buffer_ld8(top_tm_blob_data, v_tm_offset + 9 * psc(cstep));
-    afpvec8 v14 = buffer_ld8(top_tm_blob_data, v_tm_offset + 10 * psc(cstep));
-    afpvec8 v15 = buffer_ld8(top_tm_blob_data, v_tm_offset + 11 * psc(cstep));
-    afpvec8 v20 = buffer_ld8(top_tm_blob_data, v_tm_offset + 12 * psc(cstep));
-    afpvec8 v21 = buffer_ld8(top_tm_blob_data, v_tm_offset + 13 * psc(cstep));
-    afpvec8 v22 = buffer_ld8(top_tm_blob_data, v_tm_offset + 14 * psc(cstep));
-    afpvec8 v23 = buffer_ld8(top_tm_blob_data, v_tm_offset + 15 * psc(cstep));
-    afpvec8 v24 = buffer_ld8(top_tm_blob_data, v_tm_offset + 16 * psc(cstep));
-    afpvec8 v25 = buffer_ld8(top_tm_blob_data, v_tm_offset + 17 * psc(cstep));
-    afpvec8 v30 = buffer_ld8(top_tm_blob_data, v_tm_offset + 18 * psc(cstep));
-    afpvec8 v31 = buffer_ld8(top_tm_blob_data, v_tm_offset + 19 * psc(cstep));
-    afpvec8 v32 = buffer_ld8(top_tm_blob_data, v_tm_offset + 20 * psc(cstep));
-    afpvec8 v33 = buffer_ld8(top_tm_blob_data, v_tm_offset + 21 * psc(cstep));
-    afpvec8 v34 = buffer_ld8(top_tm_blob_data, v_tm_offset + 22 * psc(cstep));
-    afpvec8 v35 = buffer_ld8(top_tm_blob_data, v_tm_offset + 23 * psc(cstep));
-    afpvec8 v40 = buffer_ld8(top_tm_blob_data, v_tm_offset + 24 * psc(cstep));
-    afpvec8 v41 = buffer_ld8(top_tm_blob_data, v_tm_offset + 25 * psc(cstep));
-    afpvec8 v42 = buffer_ld8(top_tm_blob_data, v_tm_offset + 26 * psc(cstep));
-    afpvec8 v43 = buffer_ld8(top_tm_blob_data, v_tm_offset + 27 * psc(cstep));
-    afpvec8 v44 = buffer_ld8(top_tm_blob_data, v_tm_offset + 28 * psc(cstep));
-    afpvec8 v45 = buffer_ld8(top_tm_blob_data, v_tm_offset + 29 * psc(cstep));
-    afpvec8 v50 = buffer_ld8(top_tm_blob_data, v_tm_offset + 30 * psc(cstep));
-    afpvec8 v51 = buffer_ld8(top_tm_blob_data, v_tm_offset + 31 * psc(cstep));
-    afpvec8 v52 = buffer_ld8(top_tm_blob_data, v_tm_offset + 32 * psc(cstep));
-    afpvec8 v53 = buffer_ld8(top_tm_blob_data, v_tm_offset + 33 * psc(cstep));
-    afpvec8 v54 = buffer_ld8(top_tm_blob_data, v_tm_offset + 34 * psc(cstep));
-    afpvec8 v55 = buffer_ld8(top_tm_blob_data, v_tm_offset + 35 * psc(cstep));
+    int v_tm_offset = gz * psc(cstep) + gy * psc(block_x) + gx;
+
+    afpvec8 v00 = buffer_ld8(top_tm_blob_data, v_tm_offset + 0 * psc(cstep) * c);
+    afpvec8 v01 = buffer_ld8(top_tm_blob_data, v_tm_offset + 1 * psc(cstep) * c);
+    afpvec8 v02 = buffer_ld8(top_tm_blob_data, v_tm_offset + 2 * psc(cstep) * c);
+    afpvec8 v03 = buffer_ld8(top_tm_blob_data, v_tm_offset + 3 * psc(cstep) * c);
+    afpvec8 v04 = buffer_ld8(top_tm_blob_data, v_tm_offset + 4 * psc(cstep) * c);
+    afpvec8 v05 = buffer_ld8(top_tm_blob_data, v_tm_offset + 5 * psc(cstep) * c);
+    afpvec8 v10 = buffer_ld8(top_tm_blob_data, v_tm_offset + 6 * psc(cstep) * c);
+    afpvec8 v11 = buffer_ld8(top_tm_blob_data, v_tm_offset + 7 * psc(cstep) * c);
+    afpvec8 v12 = buffer_ld8(top_tm_blob_data, v_tm_offset + 8 * psc(cstep) * c);
+    afpvec8 v13 = buffer_ld8(top_tm_blob_data, v_tm_offset + 9 * psc(cstep) * c);
+    afpvec8 v14 = buffer_ld8(top_tm_blob_data, v_tm_offset + 10 * psc(cstep) * c);
+    afpvec8 v15 = buffer_ld8(top_tm_blob_data, v_tm_offset + 11 * psc(cstep) * c);
+    afpvec8 v20 = buffer_ld8(top_tm_blob_data, v_tm_offset + 12 * psc(cstep) * c);
+    afpvec8 v21 = buffer_ld8(top_tm_blob_data, v_tm_offset + 13 * psc(cstep) * c);
+    afpvec8 v22 = buffer_ld8(top_tm_blob_data, v_tm_offset + 14 * psc(cstep) * c);
+    afpvec8 v23 = buffer_ld8(top_tm_blob_data, v_tm_offset + 15 * psc(cstep) * c);
+    afpvec8 v24 = buffer_ld8(top_tm_blob_data, v_tm_offset + 16 * psc(cstep) * c);
+    afpvec8 v25 = buffer_ld8(top_tm_blob_data, v_tm_offset + 17 * psc(cstep) * c);
+    afpvec8 v30 = buffer_ld8(top_tm_blob_data, v_tm_offset + 18 * psc(cstep) * c);
+    afpvec8 v31 = buffer_ld8(top_tm_blob_data, v_tm_offset + 19 * psc(cstep) * c);
+    afpvec8 v32 = buffer_ld8(top_tm_blob_data, v_tm_offset + 20 * psc(cstep) * c);
+    afpvec8 v33 = buffer_ld8(top_tm_blob_data, v_tm_offset + 21 * psc(cstep) * c);
+    afpvec8 v34 = buffer_ld8(top_tm_blob_data, v_tm_offset + 22 * psc(cstep) * c);
+    afpvec8 v35 = buffer_ld8(top_tm_blob_data, v_tm_offset + 23 * psc(cstep) * c);
+    afpvec8 v40 = buffer_ld8(top_tm_blob_data, v_tm_offset + 24 * psc(cstep) * c);
+    afpvec8 v41 = buffer_ld8(top_tm_blob_data, v_tm_offset + 25 * psc(cstep) * c);
+    afpvec8 v42 = buffer_ld8(top_tm_blob_data, v_tm_offset + 26 * psc(cstep) * c);
+    afpvec8 v43 = buffer_ld8(top_tm_blob_data, v_tm_offset + 27 * psc(cstep) * c);
+    afpvec8 v44 = buffer_ld8(top_tm_blob_data, v_tm_offset + 28 * psc(cstep) * c);
+    afpvec8 v45 = buffer_ld8(top_tm_blob_data, v_tm_offset + 29 * psc(cstep) * c);
+    afpvec8 v50 = buffer_ld8(top_tm_blob_data, v_tm_offset + 30 * psc(cstep) * c);
+    afpvec8 v51 = buffer_ld8(top_tm_blob_data, v_tm_offset + 31 * psc(cstep) * c);
+    afpvec8 v52 = buffer_ld8(top_tm_blob_data, v_tm_offset + 32 * psc(cstep) * c);
+    afpvec8 v53 = buffer_ld8(top_tm_blob_data, v_tm_offset + 33 * psc(cstep) * c);
+    afpvec8 v54 = buffer_ld8(top_tm_blob_data, v_tm_offset + 34 * psc(cstep) * c);
+    afpvec8 v55 = buffer_ld8(top_tm_blob_data, v_tm_offset + 35 * psc(cstep) * c);
 
 #define sq2 1.41421356237
 #define sq2_m2 1.41421356237*2

src/layer/vulkan/shader/convolution_pack8_3x3s1d1_winograd_gemm.comp

@@ -4,14 +4,13 @@
 #version 450
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec8 top_tm_blob_data[]; };
@@ -19,11 +18,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec8 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -33,7 +29,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 
     afpvec8 sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));
@@ -41,10 +37,10 @@ void main()
     afpvec8 sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));
     afpvec8 sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = (gz * psc(c) * psc(outc) + gy * psc(c)) * 8;
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = (gz * c * outc + gy * c) * 8;
 
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec8 v0 = buffer_ld8(bottom_tm_blob_data, v_offset + 0);
         afpvec8 v1 = buffer_ld8(bottom_tm_blob_data, v_offset + 1);
@@ -97,11 +93,11 @@ void main()
         sum3[1].b += dot(v3[0], k6[0]) + dot(v3[1], k6[1]);
         sum3[1].a += dot(v3[0], k7[0]) + dot(v3[1], k7[1]);
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 8;
     }
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st8(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st8(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack8to1_3x3s1d1_winograd_gemm.comp

@@ -4,14 +4,13 @@
 #version 450
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfp top_tm_blob_data[]; };
@@ -19,11 +18,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec8 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -33,7 +29,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 
     afp sum0 = afp(0.f);
@@ -41,10 +37,10 @@ void main()
     afp sum2 = afp(0.f);
     afp sum3 = afp(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = gz * psc(c) * psc(outc) + gy * psc(c);
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = gz * c * outc + gy * c;
 
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec8 v0 = buffer_ld8(bottom_tm_blob_data, v_offset + 0);
         afpvec8 v1 = buffer_ld8(bottom_tm_blob_data, v_offset + 1);
@@ -59,11 +55,11 @@ void main()
         sum2 += dot(v2[0], k[0]) + dot(v2[1], k[1]);
         sum3 += dot(v3[0], k[0]) + dot(v3[1], k[1]);
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 1;
     }
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st1(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st1(top_tm_blob_data, gi + 1, sum1);

src/layer/vulkan/shader/convolution_pack8to4_3x3s1d1_winograd_gemm.comp

@@ -4,14 +4,13 @@
 #version 450
 
 layout (constant_id = 0) const int batch = 1;
+layout (constant_id = 1) const int c = 0;
+layout (constant_id = 2) const int outc = 0;
 
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+#define shape_constant_id_offset 3
+layout (constant_id = shape_constant_id_offset + 0) const int cstep = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int outcstep = 0;
 
 layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
 layout (binding = 1) writeonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
@@ -19,11 +18,8 @@ layout (binding = 2) readonly buffer weight_tm_blob { sfpvec8 weight_tm_data[];
 
 layout (push_constant) uniform parameter
 {
-    int c;
     int cstep;
-
     int outw;
-    int outc;
     int outcstep;
 } p;
 
@@ -33,7 +29,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
+    if (gx >= psc(outw) || gy >= outc || gz >= batch)
         return;
 
     afpvec4 sum0 = afpvec4(0.f);
@@ -41,10 +37,10 @@ void main()
     afpvec4 sum2 = afpvec4(0.f);
     afpvec4 sum3 = afpvec4(0.f);
 
-    int v_offset = gz * psc(cstep) + gx;
-    int w_offset = (gz * psc(c) * psc(outc) + gy * psc(c)) * 4;
+    int v_offset = gz * c * psc(cstep) + gx;
+    int w_offset = (gz * c * outc + gy * c) * 4;
 
-    for (int z = 0; z < psc(c); z++)
+    for (int z = 0; z < c; z++)
     {
         afpvec8 v0 = buffer_ld8(bottom_tm_blob_data, v_offset + 0);
         afpvec8 v1 = buffer_ld8(bottom_tm_blob_data, v_offset + 1);
@@ -77,11 +73,11 @@ void main()
         sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
         sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
 
-        v_offset += psc(outw);
+        v_offset += psc(cstep);
         w_offset += 4;
     }
 
-    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
+    int gi = gz * outc * psc(outcstep) + gy * psc(outcstep) + gx;
 
     buffer_st4(top_tm_blob_data, gi + 0, sum0);
     if (gx + 1 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 1, sum1);