data b cache example, slower than original

Author: netrunnereve

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

@@ -12,6 +12,7 @@ layout (constant_id = 1) const uint NUM_ROWS = 1;
 shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16];
 shared block_q6_K_packed16 blkcache[BLOCK_SIZE/16];
+shared B_TYPE bycache[BLOCK_SIZE/16][QUANT_K];
 
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
@@ -34,7 +35,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint ql_offset = 64*v_im + l0;
     const uint qh_offset = 32*v_im + l0;
     const uint s_offset = 8*v_im + is;
-    const uint y_offset = 128*v_im + l0;
     const uint bcs_offset = (itid%2 == 1) ? 8 : 0;
 
     FLOAT_TYPE temp[NUM_ROWS];
@@ -45,12 +45,16 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
     [[unroll]] for (uint i0 = 0; i0 < num_blocks_per_row; i0 += it_size) {
         uint i = i0 + ix; // 16 thread group specific counter
-        const uint y_idx = i * QUANT_K + y_offset;
+        const uint y_idx = i0 * QUANT_K;
+        const int blim = min(int(num_blocks_per_row) - int(i0), 4);
 
-        B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
-        B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
-        B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
-        B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
+        // assume 64 threads
+        [[unroll]] for (int n = 0; n < blim; ++n) {
+            [[unroll]] for (int l = 0; l < 4; ++l) {
+                bycache[n][tid + 64*l] = data_b[b_offset + y_idx + QUANT_K*n + tid + 64*l];
+            }
+        }
+        barrier();
 
         uint ibi = first_row*num_blocks_per_row;
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -59,7 +63,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
             // cache full superblock into shared memory with coalesced reads
             // we assume 64 threads here!
-            const int blim = min(int(num_blocks_per_row) - int(i0), 4);
             // this is required as this loop is super sensitive to unrolling with hardcoded 4
             if (blim == 4) {
                 if (tid < 52) {
@@ -109,10 +112,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
             FLOAT_TYPE sum[4] = {0, 0, 0, 0};
             [[unroll]] for (uint l = 0; l < 4; ++l) {
-                sum[0] = fma(FLOAT_TYPE(by0[l]), FLOAT_TYPE(int8_t(q0[l]) - 32), sum[0]);
-                sum[1] = fma(FLOAT_TYPE(by32[l]), FLOAT_TYPE(int8_t(q1[l]) - 32), sum[1]);
-                sum[2] = fma(FLOAT_TYPE(by64[l]), FLOAT_TYPE(int8_t(q2[l]) - 32), sum[2]);
-                sum[3] = fma(FLOAT_TYPE(by96[l]), FLOAT_TYPE(int8_t(q3[l]) - 32), sum[3]);
+                sum[0] = fma(FLOAT_TYPE(bycache[ix][128*v_im + l0 + l]), FLOAT_TYPE(int8_t(q0[l]) - 32), sum[0]);
+                sum[1] = fma(FLOAT_TYPE(bycache[ix][128*v_im + l0 + 32 + l]), FLOAT_TYPE(int8_t(q1[l]) - 32), sum[1]);
+                sum[2] = fma(FLOAT_TYPE(bycache[ix][128*v_im + l0 + 64 + l]), FLOAT_TYPE(int8_t(q2[l]) - 32), sum[2]);
+                sum[3] = fma(FLOAT_TYPE(bycache[ix][128*v_im + l0 + 96 + l]), FLOAT_TYPE(int8_t(q3[l]) - 32), sum[3]);
             }
 
             [[unroll]] for (uint l = 0; l < 4; ++l)