revert

This reverts commit ee88a8999f517f60d9f079b7781fadf623cb1f72.

Revert "data b cache example, slower than original"

This reverts commit 97ec5a01403028f0b9c7eeaac858cb9652ff958d.

Author: netrunnereve

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

@@ -12,7 +12,6 @@ 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;
@@ -35,6 +34,7 @@ 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];
@@ -45,16 +45,12 @@ 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 = i0 * QUANT_K;
-        const int blim = min(int(num_blocks_per_row) - int(i0), 4);
+        const uint y_idx = i * QUANT_K + y_offset;
 
-        // 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();
+        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];
 
         uint ibi = first_row*num_blocks_per_row;
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -63,6 +59,7 @@ 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) {
@@ -86,14 +83,41 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
             const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-			FLOAT_TYPE dq[16];
-			FLOAT_TYPE sum = 0;
-            [[unroll]] for (uint l = 0; l < 16; ++l) {
-				dq[l] = bitfieldExtract(blkcache[ix].blk[l/2], 1, 4) | (bitfieldExtract(blkcache[ix].blk[64 + l], 2, 2) << 4);
-				sum = fma(FLOAT_TYPE(bycache[ix][16*itid + l]), dq[l], sum);
-			}
+            uint32_t ql0_u32 =  uint32_t(blkcache[ix].blk[ql_offset / 2]) | (uint32_t(blkcache[ix].blk[ql_offset / 2 + 1]) << 16);
+            uint32_t ql32_u32 = uint32_t(blkcache[ix].blk[ql_offset / 2 + 16]) | (uint32_t(blkcache[ix].blk[ql_offset / 2 + 17]) << 16);
+
+            uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+            uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+            uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+            uint32_t qh_u32 = uint32_t(blkcache[ix].blk[64 + qh_offset / 2]) | (uint32_t(blkcache[ix].blk[64 + qh_offset / 2 + 1]) << 16);
+            uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+            uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+            uint32_t qh4_u32 = (qh_u32 & 0x30303030);
+            uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+            uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+            uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+            uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+            uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+            uvec4 q0 = uvec4(unpack8(q0_u32));
+            uvec4 q1 = uvec4(unpack8(q1_u32));
+            uvec4 q2 = uvec4(unpack8(q2_u32));
+            uvec4 q3 = uvec4(unpack8(q3_u32));
+
+            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]);
+            }
 
-            temp[n] += sum * sccache[ix][itid] * d;
+            [[unroll]] for (uint l = 0; l < 4; ++l)
+                sum[l] *= sccache[ix][s_offset + l*2];
+            temp[n] += (sum[0] + sum[1] + sum[2] + sum[3]) * d;
         }
     }
 

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

@@ -269,8 +269,8 @@ struct block_q6_K_packed16
 {
     // blk contains the following:
     // uint16_t ql[QUANT_K_Q6_K/2/2];
-    // uint16_t qh[QUANT_K_Q6_K/4/2]; 64
-    // uint16_t scales[QUANT_K_Q6_K/8]; 96
+    // uint16_t qh[QUANT_K_Q6_K/4/2];
+    // uint16_t scales[QUANT_K_Q6_K/8];
     uint16_t blk[104];
     float16_t d;
 };