Use mutex instead of atomics for vk_instance counters

* tests: Add im2col perf tests

* vulkan: optimize im2col, more elements per thread

* vulkan: increase small tile size for NV_coopmat2

* vulkan: change im2col to 512 elements per workgroup

examples/server/README.md

@@ -450,7 +450,7 @@ These words will not be included in the completion, so make sure to add them to
 
 `post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
 
-`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error.
+`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error. Note that fields with a slash will be unnested; for example, `generation_settings/n_predict` will move the field `n_predict` from the `generation_settings` object to the root of the response and give it a new name.
 
 **Response format**
 

examples/server/server.cpp

@@ -1856,6 +1856,8 @@ struct server_context {
                 result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
                 slot.n_sent_text += result.text_to_send.size();
                 // add the token to slot queue and cache
+            } else {
+                result.text_to_send = "";
             }
 
             slot.add_token(result);

ggml/include/ggml-vulkan.h

@@ -11,6 +11,7 @@ extern "C" {
 #define GGML_VK_MAX_DEVICES 16
 
 GGML_BACKEND_API void ggml_vk_instance_init(void);
+GGML_BACKEND_API void ggml_vk_instance_unload(void);
 
 // backend API
 GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);

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

@@ -21,9 +21,9 @@ void main() {
     get_indices(idx, i00, i01, i02, i03);
 
     if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
     } else {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]));
     }
 }
 

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

@@ -22,7 +22,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }

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

@@ -30,12 +30,12 @@ void main() {
     const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : data_b[get_boffset() + src1_idx]);
 #else
     if (is_src0) {
-        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+        data_d[get_doffset() + dst_idx] = data_a[get_aoffset() + src0_idx];
     } else {
-        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+        data_d[get_doffset() + dst_idx] = data_b[get_boffset() + src1_idx];
     }
 #endif
 }

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

@@ -19,9 +19,9 @@ void main() {
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }
@@ -32,9 +32,9 @@ void main() {
             }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }

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

@@ -13,8 +13,8 @@ void main() {
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
+    data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];
 #endif
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(cos(val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
 }

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

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

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

@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2; int param3;
 } p;
 
@@ -22,6 +22,10 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
+uint get_doffset() { return p.misalign_offsets & 0xFF; }
+
 // mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
 uint fastmod(uint a, uint b) {
     if ((b & (b-1)) == 0) {

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

@@ -6,7 +6,7 @@ layout (push_constant) uniform parameter
     uint ne;
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2;
 
     uint ne0_012mp; uint ne0_012L;
@@ -24,6 +24,9 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
 // see init_fastdiv_values in ggml-vulkan.cpp
 uint fastdiv(uint n, uint mp, uint L) {
     uint msbs, lsbs;

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

@@ -15,10 +15,10 @@ void main() {
         return;
     }
 
-    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+    const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+    const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
     data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -2,6 +2,7 @@
 
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_spirv_intrinsics: enable
+#extension GL_EXT_control_flow_attributes : require
 
 #if RTE16
 spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
@@ -23,40 +24,64 @@ layout (push_constant) uniform parameter
 
 #include "types.comp"
 
-#define BLOCK_SIZE 256
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+const uint NUM_ITER = 512 / BLOCK_SIZE;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
-    if (i >= p.pelements) {
-        return;
-    }
-
-    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
-    const uint kx = i / ksize;
-    const uint kd = kx * ksize;
-    const uint ky = (i - kd) / p.OW;
-    const uint ix = i % p.OW;
+    const uint gidx = gl_GlobalInvocationID.x;
 
     const uint oh = gl_GlobalInvocationID.y;
     const uint batch = gl_GlobalInvocationID.z / p.IC;
     const uint ic = gl_GlobalInvocationID.z % p.IC;
 
-    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
-    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
-
-    const uint offset_dst =
-        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
-        (ic * (p.KW * p.KH) + ky * p.KW + kx);
-
-    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
-        data_d[offset_dst] = D_TYPE(0.0f);
-    } else {
-        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
-        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    A_TYPE values[NUM_ITER];
+    uint offset_dst[NUM_ITER];
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+        values[idx] = A_TYPE(0);
     }
+
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
+        const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+        const uint kx = i / ksize;
+        const uint kd = kx * ksize;
+        const uint ky = (i - kd) / p.OW;
+        const uint ix = i % p.OW;
+
+        const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+        const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+        offset_dst[idx] =
+            ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+            (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        if (iih < p.IH && iiw < p.IW) {
+            const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+            values[idx] = data_a[offset_src + iih * p.IW + iiw];
+        }
+    }
+
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        data_d[offset_dst[idx]] = D_TYPE(values[idx]);
+    }
+
 }

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

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -6,21 +6,15 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
+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;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -38,15 +32,15 @@ void main() {
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = d.x;
-        const FLOAT_TYPE dmin = d.y;
-
         B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
         B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
         B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
@@ -56,58 +50,84 @@ void main() {
         B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
         B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
 
-        uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
-        uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = d.x;
+            const FLOAT_TYPE dmin = d.y;
 
-        uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
-        uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
-        uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
+            uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
 
-        uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
-        uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
-        uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
-        uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
+            uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
+            uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
+            uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
 
-        uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
-        uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
-        uvec2 qs0 =  uvec2(unpack8(qs0_u16));
-        uvec2 qs16 = uvec2(unpack8(qs16_u16));
+            uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
+            uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
+            uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
+            uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
 
-        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
-        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
-            sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
-                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
-                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
-                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
-                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
-                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
-                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
-                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
-            sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
-                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
-                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
-                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
-                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
-                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
-                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
-                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
+            uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
+            uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
+            uvec2 qs0 =  uvec2(unpack8(qs0_u16));
+            uvec2 qs16 = uvec2(unpack8(qs16_u16));
+
+            FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+            FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 2; ++l) {
+                sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
+                       fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
+                       fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
+                       fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
+                       fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
+                       fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
+                       fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
+                       fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
+                sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
+                       fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
+                       fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
+                       fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
+                       fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
+                       fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
+                       fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
+                       fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
+            }
+            temp[n] = fma(dall, sum1, fma(-dmin, sum2, temp[n]));
         }
-        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
-
     // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // 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/mul_mat_vec_q3_k.comp

@@ -6,21 +6,15 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
+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;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -35,19 +29,21 @@ void main() {
 
     const uint8_t m = uint8_t(1 << (4 * v_im));
 
-    const uint l0 = 2*v_in;                                // 0...15
+    const uint l0 = 2*v_in;                                 // 0...15
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     const uint s_shift = 4 * v_im;
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
         B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
         B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
         B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
@@ -57,44 +53,68 @@ void main() {
         B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
         B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
 
-        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
-        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
-        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
-        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
-        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
-        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
-        u8vec2 s0 = unpack8(s0_16);
-        u8vec2 s2 = unpack8(s2_16);
-        u8vec2 s4 = unpack8(s4_16);
-        u8vec2 s6 = unpack8(s6_16);
-        u8vec2 s8 = unpack8(s8_16);
-        u8vec2 s10 = unpack8(s10_16);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 2; ++l) {
-            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+            uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
+            uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
+            uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
+            uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
+            uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
+            uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
+            u8vec2 s0 = unpack8(s0_16);
+            u8vec2 s2 = unpack8(s2_16);
+            u8vec2 s4 = unpack8(s4_16);
+            u8vec2 s6 = unpack8(s6_16);
+            u8vec2 s8 = unpack8(s8_16);
+            u8vec2 s10 = unpack8(s10_16);
+
+            FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 2; ++l) {
+                sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                      fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+            }
+            temp[n] = fma(d, sum, temp[n]);
         }
-        temp = fma(d, sum, temp);
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
-
     // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // 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/mul_mat_vec_q4_k.comp

@@ -7,21 +7,15 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
+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;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -31,8 +25,8 @@ void main() {
 
     const uint step = 4;
 
-    const uint il = itid/step;                               // 0...3
-    const uint ir = itid - step*il;                          // 0...7 or 0...3
+    const uint il = itid/step;                      // 0...3
+    const uint ir = itid - step*il;                 // 0...7 or 0...3
     const uint n =  4;
 
     const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
@@ -42,90 +36,116 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-        uvec4 scale0 = uvec4(unpack8(scale0_u32));
-        uvec4 scale4 = uvec4(unpack8(scale4_u32));
-        uvec4 scale8 = uvec4(unpack8(scale8_u32));
-
-        const uint32_t sc0 = (  scale0.x       & 0x3f);
-        const uint32_t sc1 = (  scale0.y       & 0x3f);
-        const uint32_t sc2 = (  scale4.x       & 0x3f);
-        const uint32_t sc3 = (  scale4.y       & 0x3f);
-        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
-
-        uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
-        uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
-
-        uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
-        uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
-        uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
-
-        uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
-        uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
-        uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
-        uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
-
-        const uint32_t q4_0  = qs0_lo4.x;
-        const uint32_t q4_1  = qs0_lo4.y;
-        const uint32_t q4_2  = qs0_lo4.z;
-        const uint32_t q4_3  = qs0_lo4.w;
-        const uint32_t q4_4  = qs0_hi4.x;
-        const uint32_t q4_5  = qs0_hi4.y;
-        const uint32_t q4_6  = qs0_hi4.z;
-        const uint32_t q4_7  = qs0_hi4.w;
-        const uint32_t q4_8  = qs64_lo4.x;
-        const uint32_t q4_9  = qs64_lo4.y;
-        const uint32_t q4_10 = qs64_lo4.z;
-        const uint32_t q4_11 = qs64_lo4.w;
-        const uint32_t q4_12 = qs64_hi4.x;
-        const uint32_t q4_13 = qs64_hi4.y;
-        const uint32_t q4_14 = qs64_hi4.z;
-        const uint32_t q4_15 = qs64_hi4.w;
-
         B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
         B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
         B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
         B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
 
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
-            fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
-            fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
-            fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+            uvec4 scale0 = uvec4(unpack8(scale0_u32));
+            uvec4 scale4 = uvec4(unpack8(scale4_u32));
+            uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+            const uint32_t sc0 = (  scale0.x       & 0x3f);
+            const uint32_t sc1 = (  scale0.y       & 0x3f);
+            const uint32_t sc2 = (  scale4.x       & 0x3f);
+            const uint32_t sc3 = (  scale4.y       & 0x3f);
+            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+            uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+            uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
+
+            uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
+            uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
+            uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
+
+            uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
+            uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
+            uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
+            uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
+
+            const uint32_t q4_0  = qs0_lo4.x;
+            const uint32_t q4_1  = qs0_lo4.y;
+            const uint32_t q4_2  = qs0_lo4.z;
+            const uint32_t q4_3  = qs0_lo4.w;
+            const uint32_t q4_4  = qs0_hi4.x;
+            const uint32_t q4_5  = qs0_hi4.y;
+            const uint32_t q4_6  = qs0_hi4.z;
+            const uint32_t q4_7  = qs0_hi4.w;
+            const uint32_t q4_8  = qs64_lo4.x;
+            const uint32_t q4_9  = qs64_lo4.y;
+            const uint32_t q4_10 = qs64_lo4.z;
+            const uint32_t q4_11 = qs64_lo4.w;
+            const uint32_t q4_12 = qs64_hi4.x;
+            const uint32_t q4_13 = qs64_hi4.y;
+            const uint32_t q4_14 = qs64_hi4.z;
+            const uint32_t q4_15 = qs64_hi4.w;
+
+            const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
+            const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
+            const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
+            const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
+            const FLOAT_TYPE smin =
+                fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
+                fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
+                fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
+                fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
+            temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
+        }
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
-
     // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // 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/mul_mat_vec_q5_k.comp

@@ -7,21 +7,15 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
+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;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -39,74 +33,16 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        f16vec2 d = data_a[ib0 + i].d;
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
-
-        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
-        uvec4 scale0 = uvec4(unpack8(scale0_u32));
-        uvec4 scale4 = uvec4(unpack8(scale4_u32));
-        uvec4 scale8 = uvec4(unpack8(scale8_u32));
-
-        const uint32_t sc0 = (  scale0.x       & 0x3f);
-        const uint32_t sc1 = (  scale0.y       & 0x3f);
-        const uint32_t sc2 = (  scale4.x       & 0x3f);
-        const uint32_t sc3 = (  scale4.y       & 0x3f);
-        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
-        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
-        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
-        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
-
-        uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
-        uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
-
-        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
-        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
-        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
-
-        uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
-
-        uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
-        uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
-        uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
-        uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
-
-        qs0_16_u32_lo4 += qs0_16_lo4_offset16;
-        qs0_16_u32_hi4 += qs0_16_hi4_offset16;
-        qs64_80_u32_lo4 += qs64_80_lo4_offset16;
-        qs64_80_u32_hi4 += qs64_80_hi4_offset16;
-
-        uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
-        uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
-        uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
-        uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
-
-        const uint32_t q4_0  = qs0_16_lo4.x;
-        const uint32_t q4_1  = qs0_16_lo4.y;
-        const uint32_t q4_2  = qs0_16_lo4.z;
-        const uint32_t q4_3  = qs0_16_lo4.w;
-        const uint32_t q4_4  = qs0_16_hi4.x;
-        const uint32_t q4_5  = qs0_16_hi4.y;
-        const uint32_t q4_6  = qs0_16_hi4.z;
-        const uint32_t q4_7  = qs0_16_hi4.w;
-        const uint32_t q4_8  = qs64_80_lo4.x;
-        const uint32_t q4_9  = qs64_80_lo4.y;
-        const uint32_t q4_10 = qs64_80_lo4.z;
-        const uint32_t q4_11 = qs64_80_lo4.w;
-        const uint32_t q4_12 = qs64_80_hi4.x;
-        const uint32_t q4_13 = qs64_80_hi4.y;
-        const uint32_t q4_14 = qs64_80_hi4.z;
-        const uint32_t q4_15 = qs64_80_hi4.w;
-
         B_TYPE_VEC2 by10 =  data_b_v2[(b_offset + y1_idx) / 2];
         B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
         B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
@@ -116,45 +52,129 @@ void main() {
         B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
         B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
 
-        const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(by10.x), q4_0,
-          fma(FLOAT_TYPE(by10.y), q4_1,
-          fma(FLOAT_TYPE(by116.x), q4_2,
-             FLOAT_TYPE(by116.y) * q4_3)));
-        const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(by132.x), q4_4,
-          fma(FLOAT_TYPE(by132.y), q4_5,
-          fma(FLOAT_TYPE(by148.x), q4_6,
-             FLOAT_TYPE(by148.y) * q4_7)));
-        const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(by20.x), q4_8,
-          fma(FLOAT_TYPE(by20.y), q4_9,
-          fma(FLOAT_TYPE(by216.x), q4_10,
-             FLOAT_TYPE(by216.y) * q4_11)));
-        const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(by232.x), q4_12,
-          fma(FLOAT_TYPE(by232.y), q4_13,
-          fma(FLOAT_TYPE(by248.x), q4_14,
-             FLOAT_TYPE(by248.y) * q4_15)));
-        const FLOAT_TYPE smin =
-          fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
-          fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
-          fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
-              (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            f16vec2 d = data_a[ib0 + i].d;
+            const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+            const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+            uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+            uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+            uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+            uvec4 scale0 = uvec4(unpack8(scale0_u32));
+            uvec4 scale4 = uvec4(unpack8(scale4_u32));
+            uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+            const uint32_t sc0 = (  scale0.x       & 0x3f);
+            const uint32_t sc1 = (  scale0.y       & 0x3f);
+            const uint32_t sc2 = (  scale4.x       & 0x3f);
+            const uint32_t sc3 = (  scale4.y       & 0x3f);
+            const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+            const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+            const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+            const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+            uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+            uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+
+            uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+            uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+            uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+
+            uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
+
+            uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
+            uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
+            uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
+            uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
+
+            qs0_16_u32_lo4 += qs0_16_lo4_offset16;
+            qs0_16_u32_hi4 += qs0_16_hi4_offset16;
+            qs64_80_u32_lo4 += qs64_80_lo4_offset16;
+            qs64_80_u32_hi4 += qs64_80_hi4_offset16;
+
+            uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
+            uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
+            uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
+            uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
+
+            const uint32_t q4_0  = qs0_16_lo4.x;
+            const uint32_t q4_1  = qs0_16_lo4.y;
+            const uint32_t q4_2  = qs0_16_lo4.z;
+            const uint32_t q4_3  = qs0_16_lo4.w;
+            const uint32_t q4_4  = qs0_16_hi4.x;
+            const uint32_t q4_5  = qs0_16_hi4.y;
+            const uint32_t q4_6  = qs0_16_hi4.z;
+            const uint32_t q4_7  = qs0_16_hi4.w;
+            const uint32_t q4_8  = qs64_80_lo4.x;
+            const uint32_t q4_9  = qs64_80_lo4.y;
+            const uint32_t q4_10 = qs64_80_lo4.z;
+            const uint32_t q4_11 = qs64_80_lo4.w;
+            const uint32_t q4_12 = qs64_80_hi4.x;
+            const uint32_t q4_13 = qs64_80_hi4.y;
+            const uint32_t q4_14 = qs64_80_hi4.z;
+            const uint32_t q4_15 = qs64_80_hi4.w;
+
+            const FLOAT_TYPE sx =
+              fma(FLOAT_TYPE(by10.x), q4_0,
+              fma(FLOAT_TYPE(by10.y), q4_1,
+              fma(FLOAT_TYPE(by116.x), q4_2,
+                 FLOAT_TYPE(by116.y) * q4_3)));
+            const FLOAT_TYPE sy =
+              fma(FLOAT_TYPE(by132.x), q4_4,
+              fma(FLOAT_TYPE(by132.y), q4_5,
+              fma(FLOAT_TYPE(by148.x), q4_6,
+                 FLOAT_TYPE(by148.y) * q4_7)));
+            const FLOAT_TYPE sz =
+              fma(FLOAT_TYPE(by20.x), q4_8,
+              fma(FLOAT_TYPE(by20.y), q4_9,
+              fma(FLOAT_TYPE(by216.x), q4_10,
+                 FLOAT_TYPE(by216.y) * q4_11)));
+            const FLOAT_TYPE sw =
+              fma(FLOAT_TYPE(by232.x), q4_12,
+              fma(FLOAT_TYPE(by232.y), q4_13,
+              fma(FLOAT_TYPE(by248.x), q4_14,
+                 FLOAT_TYPE(by248.y) * q4_15)));
+            const FLOAT_TYPE smin =
+              fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+              fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+              fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+                  (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+            temp[n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[n]));
+        }
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
-
     // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // 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/mul_mat_vec_q6_k.comp

@@ -7,21 +7,15 @@
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
+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;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -42,69 +36,95 @@ void main() {
     const uint s_offset  =  8*v_im + is;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
-        const uint y_idx   = i * QUANT_K + y_offset;
-
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
-        FLOAT_TYPE scales[4];
-        scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
-        scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
-        scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
-        scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
-
-        uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
-        uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
-
-        uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
-        uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
-        uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
-        uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
-
-        uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
-        uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
-        uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
-        uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
-        uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
-
-        uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
-        uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
-        uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
-        uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
-
-        uvec4 q0 = uvec4(unpack8(q0_u32));
-        uvec4 q1 = uvec4(unpack8(q1_u32));
-        uvec4 q2 = uvec4(unpack8(q2_u32));
-        uvec4 q3 = uvec4(unpack8(q3_u32));
+        const uint y_idx = i * QUANT_K + y_offset;
 
         B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
         B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
         B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
         B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
 
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
-                  fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
-                  fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
-                  fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+            const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+            FLOAT_TYPE scales[4];
+            scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
+            scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
+            scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
+            scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
+
+            uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+            uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+
+            uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+            uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+            uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+            uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+            uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+            uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+            uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+            uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
+            uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+            uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+            uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+            uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+            uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+            uvec4 q0 = uvec4(unpack8(q0_u32));
+            uvec4 q1 = uvec4(unpack8(q1_u32));
+            uvec4 q2 = uvec4(unpack8(q2_u32));
+            uvec4 q3 = uvec4(unpack8(q3_u32));
+
+            FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+            [[unroll]] for (int l = 0; l < 4; ++l) {
+                sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
+                      fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
+                      fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
+                      fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
+            }
+            temp[n] += sum * d;
         }
-        temp += sum * d;
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
     // sum up partial sums and write back result
-
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // 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/pad.comp

@@ -24,5 +24,5 @@ void main() {
 
     const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
 
-    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : 0.0f);
 }

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

@@ -22,5 +22,5 @@ void main() {
         return;
     }
 
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx_mod(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx_mod(idx)]);
 }

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

@@ -18,7 +18,7 @@ void main() {
             continue;
         }
 
-        data_d[p.d_offset + idx] = D_TYPE(FLOAT_TYPE(data_a[idx]) * FLOAT_TYPE(p.param1));
+        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) * FLOAT_TYPE(p.param1));
         idx += num_threads;
     }
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(sin(val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(sin(val));
 }

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

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val * val);
 }

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

@@ -2,7 +2,7 @@
 
 layout (push_constant) uniform parameter
 {
-    uint ne; uint d_offset;
+    uint ne; uint a_offset; uint d_offset;
     uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -32,5 +32,5 @@ void main() {
     const uint i02 = uint(i12 / p.sf2);
     const uint i03 = uint(i13 / p.sf3);
 
-    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+    data_d[p.d_offset + idx] = D_TYPE(data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
 }

tests/test-backend-ops.cpp

@@ -3945,6 +3945,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    for (int K : {3, 5}) {
+        for (int IC : {256, 2560}) {
+            for (int IW_IH : {32, 64, 256}) {
+                if (IC == 2560 && IW_IH == 256) {
+                    // too big
+                    continue;
+                }
+                test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {IW_IH, IW_IH, IC, 1}, {K, K, IC, 1}, 1, 1, 1, 1, 1, 1, true));
+            }
+        }
+    }
+
     return test_cases;
 }