kompute: op_mul_mat_q4_k permutted support

Signed-off-by: Sergio Lopez <[email protected]>

Author: slp

ggml/src/ggml-kompute/ggml-kompute.cpp

@@ -1079,19 +1079,26 @@ static void ggml_vk_mul_mat_q4_k(
     const std::shared_ptr<kp::Tensor>& inB,
     const std::shared_ptr<kp::Tensor>& out,
     uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10,
-    int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0,
-    int32_t ne1, int32_t r2, int32_t r3
+    int32_t ne00, int32_t ne01, int32_t ne02,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t ne0, int32_t ne1,
+    uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    uint32_t nb11, uint32_t nb12, uint32_t nb13,
+    uint32_t r2, uint32_t r3
 ) {
     const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
         kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
 
     struct PushConstants {
         uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3;
+        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
+        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
+        uint32_t r2, r3;
     } pushConsts {
-        0, 0, 0,
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3
+        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
+        nb01, nb02, nb03, nb11, nb12, nb13,
+        r2, r3
     };
 
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
@@ -1705,7 +1712,8 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                             case GGML_TYPE_Q4_K:
                                 ggml_vk_mul_mat_q4_k(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q6_K:

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp

@@ -24,8 +24,14 @@ layout (push_constant) uniform parameter {
     int ne01;
     int ne02;
     int ne12;
-    int r2;
-    int r3;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
 } pcs;
 
 void main() {
@@ -50,10 +56,11 @@ void main() {
     const uint i12 = im%pcs.ne12;
     const uint i13 = im/pcs.ne12;
 
-    const uint offset0 = (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
 
-    const uint xblk = ib_row + offset0 + pcs.inAOff;
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff;
+    const uint xblk = offset0 + pcs.inAOff;
+    const uint y = (offset1 / 4) + pcs.inBOff;
 
     float yl[16];
     float yh[16];
@@ -74,7 +81,7 @@ void main() {
         }
 
         for (int row = 0; row < N_DST; row++) {
-            uint row_idx = row * nb;
+            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
 
             uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
             uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);