vulkan: mul_mat_id coopmat2 optimizations

Add a path for when the tile fits in BN/2, similar to what we have for mul_mat.

Only call fetch_scales/store_scales once per QUANT_K block, and once at the
beginning in case start_k is not aligned.

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -2214,7 +2214,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
-        l_warptile_mmqid = { 256, 128, 128, 16, 0, device->subgroup_size };
+        l_warptile_mmqid = { 256, 128, 128, 16, 1, device->subgroup_size };
         m_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
         s_warptile_mmqid = { 256, 128, 64, 16, 0, device->subgroup_size };
         l_mmqid_wg_denoms = { 128, 128, 1 };

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -441,18 +441,64 @@ void main() {
 
         tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1);
 
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum;
-        sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0);
-
         uint k_iters = (end_k - start_k + BK - 1) / BK;
 
         fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, false);
+        store_scales(tid);
+
+#ifdef MUL_MAT_ID
+        if (enable_smaller_matrices && ic * BN + BNover2 >= _ne1) {
+            coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> sum;
+            sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(0.0);
+
+            [[dont_unroll]]
+            for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {
+
+                if ((block_k % QUANT_K) == 0) {
+                    store_scales(tid);
+                }
+                if (block_k + BK < end_k && ((block_k + BK) % QUANT_K) == 0) {
+                    fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
+                }
+
+                if ((ir + 1) * BM <= p.M && block_k + BK <= end_k) {
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+
+                    coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+
+                    sum = coopMatMulAdd(mat_a, mat_b, sum);
+                } else {
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+
+                    coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+
+                    sum = coopMatMulAdd(mat_a, mat_b, sum);
+                }
+            }
+
+            // Convert from ACC_TYPE to D_TYPE
+            coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator> mat_d;
+            mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BNover2, gl_MatrixUseAccumulator>(sum);
+
+            // Call callback to store each element, remapping row through shared memory
+            coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic);
+            return;
+        }
+#endif
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum;
+        sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0);
 
         [[dont_unroll]]
         for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {
 
-            store_scales(tid);
-            if (block_k + BK < end_k) {
+            if ((block_k % QUANT_K) == 0) {
+                store_scales(tid);
+            }
+            if (block_k + BK < end_k && ((block_k + BK) % QUANT_K) == 0) {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }