Add clra_gemv one_thread_per_row_impl

Signed-off-by: Shaun Ren <[email protected]>

nengo_ocl/clra_gemv.py

@@ -1126,6 +1126,166 @@ def block_impl(p, items):
     return [plan, plan_reduce]
 
 
+def one_thread_per_row_impl(p, num_items):
+    """Each thread computes the dot product over one row.
+       The rows are grouped consecutively when y_len is small.
+    """
+    if p.clra_alpha is not None:
+        raise NotImplementedError()
+    if p.clra_gamma is not None:
+        raise NotImplementedError()
+    if p.clra_beta is not None:
+        raise NotImplementedError()
+    if p.cl_alpha is not None:
+        raise NotImplementedError()
+    if p.cl_gamma is not None:
+        raise NotImplementedError()
+    if not all(s == 1 for s in p.A.stride0s):
+        raise NotImplementedError()
+
+    assert num_items > 0
+    items = range(num_items)
+
+    assert p.float_alpha is not None
+    assert p.float_gamma is not None
+
+    if p.A_js is None:
+        # -- easy probably, but not done
+        raise NotImplementedError()
+
+    cl_gstructure, textconf = p.cl_geometry_and_textconf(items,
+                                                         stride=1)
+
+    max_y_len = max(p.geometry[ii].y_len for ii in items)
+    max_n_dots = max(len(p.geometry[ii].dots) for ii in items)
+    all_same_y_len = len(set(p.geometry[ii].y_len for ii in items)) == 1
+
+
+    # Compute in consecutive threads if max_y_len is small,
+    # and all y_len are the same.
+    # TODO autotune
+    if all_same_y_len and max_y_len < 64:
+        consecutive = True
+        gsize = (max_y_len * num_items,)
+    else:
+        consecutive = False
+        gsize = (max_y_len, num_items)
+
+    textconf.update({
+        'max_n_dots': max_n_dots,
+        'max_y_len': max_y_len,
+        'consecutive': consecutive,
+        'all_same_y_len': all_same_y_len,
+    })
+    textconf.update(p.__dict__)
+
+    """
+    int i, item_i;
+% for ix, sprev, s, item_i in zip(range(len(items)), num_rows, num_rows[1:], items):
+    % if ix == 0:
+         if (global_i < ${s})
+    % else:
+         else if (global_i < ${s})
+    % endif
+         {
+            item_i = ${item_i};
+            i = global_i - ${sprev};
+         }
+    % endfor
+    """
+
+    text = """
+    __kernel void gemv(
+                const __global int* gstructure,
+                const __global ${A.cl_buf.ctype}* A_data,
+                const __global ${X.cl_buf.ctype}* X_data,
+            % if cl_beta is not None:
+                const __global ${cl_beta.ctype}* betas,
+            % endif
+                const __global ${Y_in.cl_buf.ctype}* Y_in_data,
+                __global ${Y.cl_buf.ctype}* Y_data)
+    {
+    % if consecutive:
+        const int global_i = get_global_id(0);
+
+        const int item_i = global_i / ${max_y_len};     // Item index
+        const int i = global_i - item_i * ${max_y_len}; // Row index within item
+    % else:
+        const int item_i = get_global_id(1);  // Item index
+        const int i = get_global_id(0);       // Row index within item
+    % endif
+
+        const __global int* lstructure =
+            gstructure + item_i * ${structure_vars_stride};
+
+    % if not all_same_y_len:
+        if (i >= ${y_len}) return;
+    % endif
+
+        ${Y.cl_buf.ctype} sum = 0;
+
+    % if max_n_dots > 1:
+        for (int ii = 0; ii < ${n_dot_products}; ii++)
+        {
+    % else:
+        const int ii = 0;
+    % endif
+
+            const int a_s1 = ${a_s1};
+            const int n_i = ${N_i};
+
+            const __global ${A.cl_buf.ctype}* a = A_data + ${a_starts} + i;
+            const __global ${X.cl_buf.ctype}* x = X_data + ${x_starts};
+
+            for (int j=0;j<n_i;j++)
+                sum += a[a_s1*j] * x[j];
+
+    % if max_n_dots > 1:
+        }
+    % endif
+
+    % if float_gamma is not None:
+        const ${Y.cl_buf.ctype} gamma = ${float_gamma};
+    % else:
+        const ${Y.cl_buf.ctype} gamma = 0;
+    % endif
+
+    % if float_beta is not None and float_beta != 0 :
+        Y_data[${y_offset} + i] = ${float_alpha} * sum + ${float_beta} * Y_in_data[${y_in_starts} + i] + gamma;
+    % elif cl_beta is not None:
+        Y_data[${y_offset} + i] = ${float_alpha} * sum + betas[${bb}] * Y_in_data[${y_in_starts} + i] + gamma;
+    % else:
+        Y_data[${y_offset} + i] = ${float_alpha} * sum + gamma;
+    % endif
+    }
+    """
+
+    text = as_ascii(Template(text, output_encoding='ascii').render(**textconf))
+
+    fn = cl.Program(p.queue.context, text).build().gemv
+
+    full_args = [
+        cl_gstructure,
+        p.A.cl_buf,
+        p.X.cl_buf,
+    ] + ([p.cl_beta] if p.cl_beta is not None else []) + [
+        p.Y_in.cl_buf,
+        p.Y.cl_buf,
+    ]
+
+    fn.set_args(*[arr.data for arr in full_args])
+    rval = Plan(p.queue, fn, gsize, None,
+                name='clra_gemv.one_thread_per_row_impl',
+                tag=p.tag,
+                bw_per_call=bw_from_geometry(p.geometry, items),
+                flops_per_call=flops_from_geometry(p.geometry, items),
+                )
+    rval.full_args = full_args  # prevent GC the args
+    rval.description = p.geometry.summary(items)
+    return rval
+
+
+
 class plan_ref_gemv(gemv_prog):
     def choose_plans(self):
         return [ref_impl(self, range(len(self.Y)))]
@@ -1191,3 +1351,7 @@ def choose_plans(self):
             plans.append(remaining_plan)
 
         return plans
+
+class plan_one_thread_per_row(gemv_prog):
+    def choose_plans(self):
+        return [one_thread_per_row_impl(self, len(self.Y))]