Compute shader shared mem not fast, and usampler2D,ssbo

[xin]

1. It takes about the same time between not use shared mem and use shared mem​. i measured the time as below.

t0=getTime();
glDispatchCompute();
glFinish();
printf("DispatchCompute:%.3f ms\n",getTime()-t0);

2. opengles3.2 compute shader support usampler2D ? i use texelFetch and get all 0 of output as below.

static const char gComputeShader[] = 
    "#version 320 es\n"
    "layout(local_size_x = 12,local_size_y = 12,local_size_z = 1) in;\n"
    "layout(location = 0) uniform highp usampler2D input0;\n"
    "layout(binding = 2, rgba32ui )writeonly uniform highp uimage2D output0;\n"
    "void main()\n"
    "{\n"
" ivec2 loc = ivec2(gl_GlobalInvocationID.xy);\n"
" uvec4 input0_data = texelFetch(input0,loc,0);\n"
        "    imageStore(output0,loc,input0_data);\n"
    "}\n";
     
       glGenTextures(1, &input0_id); 
       glBindTexture(GL_TEXTURE_2D, input0_id);
       glTexImage2D(GL_TEXTURE_2D, 0, GL_R8UI,width, height, 0, GL_RED_INTEGER, GL_UNSIGNED_BYTE, f0);
       int texId = 0;
       glActiveTexture(GL_TEXTURE0 + texId);
       glUniform1i(0, texId);

3.can create an ssbo(Shader Storage Buffer Object ) for compute shader backed by EGL image that accomplished without ever copying ？

please advise

thanks

[Alfonse_Reinheart]

@xin

It takes about the same time between not use shared mem and use shared mem​. i measured the time as below.

That is a terrible way to profile strictly on-GPU operations. You’re only profiling API overhead on the CPU, not the actual GPU execution time.

I don’t know if ES 3.2 has timer queries, but if it does, use them.

@xin

can create an ssbo(Shader Storage Buffer Object ) for compute shader backed by EGL image that accomplished without ever copying

SSBOs are buffers, not images. So no.

[xin]

when use ssbo，only copy data to ssbo from cpu buffer？ have way to remove copy？

thanks