opencl: use broadcast semantic for mul_mv_id_mxfp4_f32_flat

Author: lhez

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -7228,20 +7228,21 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
             CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
             CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
             CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
-            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
-            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
-            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
-            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
-            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
-            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
-            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb13));
-            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
-            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
-            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
-            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
-            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
-            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &r2));
-            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
 #else // GGML_OPENCL_SOA_Q
             kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
 

ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl

@@ -79,8 +79,9 @@ kernel void kernel_mul_mv_id_mxfp4_f32_flat(
     global uchar * dst,
     ulong         offsetd,
     int           ne00,
-    int           ne01,
-    int           ne02,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
     int           ne11,
     int           ne12,
     ulong         nb11,
@@ -105,7 +106,10 @@ kernel void kernel_mul_mv_id_mxfp4_f32_flat(
 
     int nb = ne00 / QK_MXFP4;
 
-    src0_e = src0_e + i02 * nb * ne01;
+    uint src0_off = i02*nb02;
+    src0_off /= 17; // 17 = sizeof(block_mxfp4)
+
+    src0_e = src0_e + src0_off;
 
     dst = dst + (idx * ne0 + iid1 * ne1 * ne0) * sizeof(float);
 
@@ -114,11 +118,12 @@ kernel void kernel_mul_mv_id_mxfp4_f32_flat(
 
     int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
 
-    ulong offset_src0 = first_row * nb;
+    uint offset_src0 = first_row*nb01;
+    offset_src0 /= 17; // 17 = sizeof(block_mxfp4)
 #ifdef SRC0Q_IMG
-    ulong offset_q = i02 * nb * ne01 + offset_src0;
+    ulong offset_q = src0_off + offset_src0;
 #else
-    src0_q = src0_q + i02 * nb * 16 * ne01;
+    src0_q = src0_q + src0_off*16;
     global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
 #endif
     global uchar * x_e = src0_e + offset_src0;