Merge branch 'master' into gemm-large

Author: nihui

.github/workflows/test-coverage.yml

@@ -59,6 +59,7 @@ jobs:
         token: ${{ secrets.CODECOV_TOKEN }}
         disable_search: true
         plugins: noop
+        binary: /data/action/.local/bin/codecov
         files: build/lcov.info
 
   linux-gcc-x64:

src/layer/vulkan/deconvolution_vulkan.cpp

@@ -37,6 +37,17 @@ class Deconvolution_vulkan : public Deconvolution
 
     Pipeline* pipeline_deconvolution_gemm;
     Pipeline* pipeline_deconvolution_col2im;
+
+    // cooperative matrix
+    bool use_cooperative_matrix;
+    int coopmat_M;
+    int coopmat_N;
+    int coopmat_K;
+    int UNROLL_SG_M;
+    int UNROLL_SG_N;
+    int UNROLL_SG_K;
+    int UNROLL_WG_M;
+    int UNROLL_WG_N;
 };
 
 } // namespace ncnn

src/layer/vulkan/deconvolution_vulkan.h

@@ -16,14 +16,15 @@ layout (constant_id = 6) const int bias_term = 0;
 layout (constant_id = 7) const int activation_type = 0;
 layout (constant_id = 8) const float activation_param_0 = 0;
 layout (constant_id = 9) const float activation_param_1 = 0;
+layout (constant_id = 10) const int outc = 0;
 
-#define shape_constant_id_offset 10
+#define shape_constant_id_offset 11
 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 cstep = 0;
 
-layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int outc = 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 col_blob { sfp col_blob_data[]; };
@@ -34,10 +35,10 @@ layout (push_constant) uniform parameter
 {
     int w;
     int h;
+    int cstep;
 
     int outw;
     int outh;
-    int outc;
     int outcstep;
 } p;
 
@@ -47,7 +48,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
     int gz = int(gl_GlobalInvocationID.z);
 
-    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
+    if (gx >= psc(outw) || gy >= psc(outh) || gz >= outc)
         return;
 
     afp sum;
@@ -84,7 +85,7 @@ void main()
                 h_k /= dilation_h;
                 w_k /= dilation_w;
 
-                const int gi = (gz * maxk + h_k * kernel_w + w_k) * psc(w) * psc(h) + sy * psc(w) + sx;
+                const int gi = (gz * maxk + h_k * kernel_w + w_k) * psc(cstep) + sy * psc(w) + sx;
 
                 sum += buffer_ld1(col_blob_data, gi);
             }

src/layer/vulkan/shader/deconvolution_col2im.comp

@@ -13,8 +13,8 @@ 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 + 4) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outc = 0;
 
 layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
 layout (binding = 1) writeonly buffer col_blob { sfp col_blob_data[]; };
@@ -27,8 +27,8 @@ layout (push_constant) uniform parameter
     int c;
     int cstep;
 
-    int outw;
-    int outh;
+    int outcstep;
+    int outc;
 } p;
 
 #if NCNN_shader_local_memory
@@ -42,7 +42,7 @@ void main()
     int gy = int(gl_GlobalInvocationID.y);
 
 #if !NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outh))
+    if (gx >= psc(outcstep) || gy >= psc(outc))
         return;
 #endif
 
@@ -158,14 +158,14 @@ void main()
 #endif
 
 #if NCNN_shader_local_memory
-    if (gx >= psc(outw) || gy >= psc(outh))
+    if (gx >= psc(outcstep) || gy >= psc(outc))
         return;
 #endif
 
-    const int gi = gy * psc(outw) + gx;
+    const int gi = gy * psc(outcstep) + gx;
 
     buffer_st1(col_blob_data, gi, sum0);
-    if (gx + 1 < psc(outw)) buffer_st1(col_blob_data, gi + 1, sum1);
-    if (gx + 2 < psc(outw)) buffer_st1(col_blob_data, gi + 2, sum2);
-    if (gx + 3 < psc(outw)) buffer_st1(col_blob_data, gi + 3, sum3);
+    if (gx + 1 < psc(outcstep)) buffer_st1(col_blob_data, gi + 1, sum1);
+    if (gx + 2 < psc(outcstep)) buffer_st1(col_blob_data, gi + 2, sum2);
+    if (gx + 3 < psc(outcstep)) buffer_st1(col_blob_data, gi + 3, sum3);
 }