vulkan: add dedicated iq4_xs mat-vec shader (#28426)

* vulkan: add dedicated iq4_xs mat-vec shader

Dedicated mul_mat_vec_iq4_xs for the dmmv path, replacing the generic fallback. ~+6-17% token generation on RDNA4 depending on model.

Assisted-by: Pi agent with Qwen3.8 27B

* vulkan iq4_xs: remove dead n_it unroll branch

Remove the n_it <= 8 experimental branch that attempted to fully unroll
the block loop. Since n_it is a runtime value, [[unroll]] is ignored by
the compiler, making both branches equivalent. Kept the simple loop
matching mul_mat_vec_iq3_s.comp.

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq4_xs.comp

@@ -0,0 +1,97 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.glsl"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+// dedicated iq4_xs mat-vec, mirrors mul_mat_vec_iq3_s.comp
+// one packed32 word per l, so the 6-bit subblock scale is hoisted to a single fma after register accumulation
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint sl = (data_a[ibi].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
+        const uint sh = (data_a[ibi].scales_h >> (2 * ib32)) & 3;
+        const float dscale = d * float(int(sl | (sh << 4)) - 32);
+
+        FLOAT_TYPE sum[NUM_COLS];
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            sum[j] = FLOAT_TYPE(0);
+        }
+
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint w = data_a_packed32[ibi].qs[4 * ib32 + l];
+            const u8vec4 q0 = unpack8(w & 0x0F0F0F0F);
+            const u8vec4 q1 = unpack8((w >> 4) & 0x0F0F0F0F);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + l]);
+                const vec4 b1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 4 + l]);
+
+                sum[j] = fma(FLOAT_TYPE(b0.x), FLOAT_TYPE(kvalues_iq4nl[q0.x]),
+                        fma(FLOAT_TYPE(b0.y), FLOAT_TYPE(kvalues_iq4nl[q0.y]),
+                        fma(FLOAT_TYPE(b0.z), FLOAT_TYPE(kvalues_iq4nl[q0.z]),
+                        fma(FLOAT_TYPE(b0.w), FLOAT_TYPE(kvalues_iq4nl[q0.w]),
+                        fma(FLOAT_TYPE(b1.x), FLOAT_TYPE(kvalues_iq4nl[q1.x]),
+                        fma(FLOAT_TYPE(b1.y), FLOAT_TYPE(kvalues_iq4nl[q1.y]),
+                        fma(FLOAT_TYPE(b1.z), FLOAT_TYPE(kvalues_iq4nl[q1.z]),
+                        fma(FLOAT_TYPE(b1.w), FLOAT_TYPE(kvalues_iq4nl[q1.w]),
+                        sum[j]))))))));
+            }
+        }
+
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            temp[j][n] = fma(dscale, sum[j], temp[j][n]);
+        }
+
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -735,7 +735,7 @@ void process_shaders() {
     for (const auto& tname : type_names) {
         // mul mat vec
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0" || tname == "tq1_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
+        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "iq4_xs" || tname == "tq2_0" || tname == "tq1_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
 
         string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
         string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}));