vulkan: Coalesce Q4_K/Q5_K scale loads (ggml-org#21751)

Some SPIR-V compilers (notably mesa) don't handle the current
vulkan Q4_K/Q5_K scale load pattern in mul_mat particularly well.
While reading three `u8`s from the 12-byte scale array should (at
least on some hardware) result in loading the full 12 bytes in a
single LOAD followed by whatever extraction is needed, at least
the ANV Intel driver really can't practically perform this
optimization.

`mesa`'s unsigned upper bound logic doesn't handle tracking bounds
through ternary, resulting in the `(is < 4) ? ... : is - 4` having
an infinite upper bound (as it cannot prove `is - 4` doesn't
underflow). While this could still be rectified if mesa looked at
the array bounds, it currently doesn't and `glslc` currently emits
SPIR-V that doesn't allow for this optimization anyway (though
maybe it will at some point, see
KhronosGroup/glslang#4206).

In mul_mat_vecq we took a different approach to loading the same
fields. We read the first two bytes we needed from `scale` then
took a branch before deciding whether we needed to read a third
byte. In mesa this did, indeed, lead to a top-level branch with
conditional loads. As such these loads ended up not being
coalesced either (at least in the ANV driver) resulting in
additional instructions in our hot loop.

Instead, here, we go ahead and force loading the full 12 bytes and
extract the bits we need from the packed-u32s instead. In mul_mat
there's a few less ternaries and only one extra shift, so even on
drivers that did optimize the previous loads properly the only
material change should be pulling a few extra bytes into registers
(which on most hardware won't cost anything anyway, though
ironically on Intel it theoretically could). In mul_mat_vecq this
requires a bit of extra math and may read bytes from the u32 that
weren't needed, but it seems likely avoiding the branch is a win
on most platforms.

On Intel Xe2/mesa 26.0.4 with the optimizations from
https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15162,

for shader matmul_id_subgroup_q4_k_f32_f16acc_aligned_l:
 * Instruction Count: 2753 -> 2688
 * SEND Count: 269 -> 261
 * Cycle Count: 273976 -> 266138
 * Max live registers: 248 -> 246
 * Non SSA regs after NIR: 381 -> 382

for shader matmul_id_subgroup_q5_k_f32_f16acc_aligned_l:
 * Instruction Count: 2767 -> 2702
 * SEND Count: 271 -> 263
 * Cycle Count: 274140 -> 268144
 * Max live registers: 248 -> 246
 * Non SSA regs after NIR: 381 -> 382

for shader mul_mat_vec_id_q4_k_q8_1_f32:
 * Instruction Count: 1930 -> 1646
 * SEND Count: 116 -> 71
 * Cycle Count: 1348306 -> 843350
 * Max live registers: 78 -> 84
 * Non SSA regs after NIR: 300 -> 135

for shader mul_mat_vec_id_q5_k_q8_1_f32:
 * Instruction Count: 2207 -> 1922
 * SEND Count: 131 -> 86
 * Cycle Count: 1392012 -> 1037836
 * Max live registers: 90 -> 90
 * Non SSA regs after NIR: 300 -> 135

for shader mul_mat_vec_q4_k_q8_1_f32:
 * Instruction Count: 2029 -> 1749
 * SEND Count: 111 -> 66
 * Cycle Count: 1347278 -> 840118
 * Max live registers: 74 -> 80
 * Non SSA regs after NIR: 299 -> 134

for shader mul_mat_vec_q5_k_q8_1_f32:
 * Instruction Count: 2307 -> 2022
 * SEND Count: 126 -> 81
 * Cycle Count: 1379820 -> 954042
 * Max live registers: 86 -> 86
 * Non SSA regs after NIR: 299 -> 134

On one Arc Pro B60, unsloth/Qwen3.5-35B-A3B-GGUF:UD-Q4_K_XL:
 * pp512: 907.34 ± 9.28 -> 941.94 ± 10.53 (+4%)
 * pp2048: 897.95 ± 1.82 -> 931.55 ± 1.79 (+4%)
 * tg128: 49.49 ± 0.02 -> 49.86 ± 0.05 (+ <1%)

On one Arc Pro B60, unsloth/Qwen3.5-27B-GGUF:Q4_K_S:
 * pp512: 324.13 ± 10.52 -> 354.33 ± 6.81 (+9%)
 * pp2048: 329.80 ± 0.25 -> 357.10 ± 0.06 (+8%)
 * tg128: 17.11 ± 0.01 -> 18.11 ± 0.01 (+6%)

On four Arc Pro B60s, unsloth/Qwen3.5-122B-A10B-GGUF:Q5_K_S with
-sm layer (note that -sm tensor improvements will naturally be
less):
 * pp512: 264.55 ± 2.81 -> 280.45 ± 3.94 (+6%)
 * pp2048: 319.32 ± 2.72 -> 335.70 ± 3.48 (+5%)
 * tg128: 26.39 ± 0.01 -> 26.67 ± 0.01 (+1%)

Author: TheBlueMatt

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl

@@ -296,13 +296,22 @@ vec2 get_dm_scale(uint ib, uint iqs) {
     const uint ib_k = ib / 8;
     const uint iqs_k = (ib % 8) * 8 + iqs;
     const uint is = iqs_k / 8;
-    u8vec2 scale_dm;
-    if (is < 4) {
-        scale_dm = u8vec2(data_a[ib_k].scales[is] & 0x3F, data_a[ib_k].scales[is + 4] & 0x3F);
-    } else {
-        scale_dm = u8vec2((data_a[ib_k].scales[is+4] & 0xF) | ((data_a[ib_k].scales[is-4] & 0xC0) >> 2),
-                          (data_a[ib_k].scales[is+4] >>  4) | ((data_a[ib_k].scales[is  ] & 0xC0) >> 2));
-    }
+
+    const uvec3 scales = uvec3(data_a_packed32[ib_k].scales[0],
+                               data_a_packed32[ib_k].scales[1],
+                               data_a_packed32[ib_k].scales[2]);
+    const uint scalesoffs = (is & 3) * 8;
+
+    const uint scidx0 = (is < 4) ? 0 : 2;
+    const uint scidxshift0 = scalesoffs;
+    const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+    const uint mbidx0 = (is < 4) ? 1 : 2;
+    const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
+    const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+
+    const uint8_t sc    = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
+    const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
+    u8vec2 scale_dm = u8vec2(sc, mbyte);
 
     return FLOAT_TYPEV2(data_a_packed32[ib_k].dm) * FLOAT_TYPEV2(scale_dm);
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -201,19 +201,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 
             const vec2 loadd = vec2(data_a[ib].dm);
 
-            const uint scidx0 = (is < 4) ? is : (is + 4);
-            const uint scidx1 = (is < 4) ? is : (is - 4);
-            const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            const uint scidxshift1 = (is < 4) ? 0 : 2;
-            const uint mbidx0 = is + 4;
-            const uint mbidx1 = (is < 4) ? is + 4 : is;
-            const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-            const uint mbidxshift0 = (is < 4) ? 0 : 4;
-            const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            const uint mbidxshift1 = (is < 4) ? 0 : 2;
-
-            const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-            const uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+            const uvec3 scales = uvec3(data_a_packed32[ib].scales[0],
+                                       data_a_packed32[ib].scales[1],
+                                       data_a_packed32[ib].scales[2]);
+            const uint scalesoffs = (is & 3) * 8;
+
+            const uint scidx0 = (is < 4) ? 0 : 2;
+            const uint scidxshift0 = scalesoffs;
+            const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+            const uint mbidx0 = (is < 4) ? 1 : 2;
+            const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
+            const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+
+            const uint8_t sc    = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
+            const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
 
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
@@ -237,19 +238,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 
             const vec2 loadd = vec2(data_a[ib].dm);
 
-            const uint scidx0 = (is < 4) ? is : (is + 4);
-            const uint scidx1 = (is < 4) ? is : (is - 4);
-            const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            const uint scidxshift1 = (is < 4) ? 0 : 2;
-            const uint mbidx0 = is + 4;
-            const uint mbidx1 = (is < 4) ? is + 4 : is;
-            const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
-            const uint mbidxshift0 = (is < 4) ? 0 : 4;
-            const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
-            const uint mbidxshift1 = (is < 4) ? 0 : 2;
-
-            const uint8_t sc    = uint8_t((data_a[ib].scales[scidx0] & 0xF)                         | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
-            const uint8_t mbyte = uint8_t(((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+            const uvec3 scales = uvec3(data_a_packed32[ib].scales[0],
+                                       data_a_packed32[ib].scales[1],
+                                       data_a_packed32[ib].scales[2]);
+            const uint scalesoffs = (is & 3) * 8;
+
+            const uint scidx0 = (is < 4) ? 0 : 2;
+            const uint scidxshift0 = scalesoffs;
+            const uint scidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+            const uint mbidx0 = (is < 4) ? 1 : 2;
+            const uint mbidxshift0 = (is < 4) ? scalesoffs : scalesoffs + 4;
+            const uint mbidxshift1 = (is < 4) ? scalesoffs : scalesoffs + 2;
+
+            const uint8_t sc    = uint8_t(((scales[scidx0] >> scidxshift0) & 0xF) | ((scales[0] >> scidxshift1) & 0x30));
+            const uint8_t mbyte = uint8_t(((scales[mbidx0] >> mbidxshift0) & 0xF) | ((scales[1] >> mbidxshift1) & 0x30));
 
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;