subgroup 64 version with subgroup add. 15% faster

scalable version

tested for subgroup sizes 16-128

Author: netrunnereve

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -4,9 +4,11 @@
 
 #include "mul_mat_vec_base.comp"
 
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-shared FLOAT_TYPE tmp[32];
+layout (constant_id = 0) const uint SUBGROUP_SIZE = 32;
+
+shared FLOAT_TYPE tmp[SUBGROUP_SIZE];
 
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
@@ -21,21 +23,19 @@ void main() {
     const uint num_blocks_per_row = p.ncols / QUANT_K;
     const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+    uint it_size = gl_WorkGroupSize.x/16;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid/it_size;  // 0...16
+    const uint ix  = tid%it_size;
 
-    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+    const uint step = 8;
 
-    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
 
-#if K_QUANTS_PER_ITERATION == 1
-    const uint l0 = v_in;                                   // 0...15
-    const uint is = 0;
-#else
     const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
     const uint is = v_in / 4;
-#endif
 
     const uint ql_offset = 64*v_im + l0;
     const uint qh_offset = 32*v_im + l0;
@@ -44,7 +44,7 @@ void main() {
 
     FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx   = i * QUANT_K + y_offset;
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
@@ -95,10 +95,10 @@ void main() {
     }
 
     tmp[gl_LocalInvocationID.x] = temp;
-
     // sum up partial sums and write back result
+
     barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
         }