Missing barrier added to shader.

Number of additional tests reduced to 108.

Author: etasnadi

ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp

@@ -29,7 +29,7 @@ const uint32_t bs = gl_WorkGroupSize.x;
 uint32_t tid = gl_LocalInvocationID.x;
 // Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
 uint32_t tmp_len = bs*p.s0+p.K;
-shared D_TYPE tmp[1024]; 
+shared D_TYPE tmp[4096]; 
 
 uint splitWork(uint workSize){
     return (bs + workSize -1) / bs;
@@ -46,6 +46,7 @@ void main(){
     uint32_t L_blocks = splitWork(p.L);
     for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
         if(L_block_id > 0){
+            barrier();
             // Shift values in tmp to the current processing window
             for(int i = 0; i < splitWork(tmp_len); i++){
                 uint32_t idx = i*bs+tid;

tests/test-backend-ops.cpp

@@ -4031,8 +4031,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     for(uint32_t Cout : {1, 9}){
         for(uint32_t Cin : {1, 7}){
-            for(uint32_t K : {1, 2, 3, 5, 6, 8, 9, 28}){
-                for(uint32_t L : {1, 2, 3, 5, 15, 16, 60, 100, 111, 127, 128, 157, 255, 376, 1024, 1173}){
+            for(uint32_t K : {1, 3, 1337}){
+                for(uint32_t L : {1, 2, 2173}){
                     for(uint32_t s0: {1, 2, 3}){
                         test_cases.emplace_back(new test_conv_transpose_1d({L,Cin,1,1}, {K,Cout,Cin,1}, s0, 0, 1));
                     }