whispercpp 1.3.6 → 1.3.8

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -10,12 +10,38 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 #if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
 #define K_PER_ITER 8
 #else
-#define K_PER_ITER 2
+#define K_PER_ITER 4
 #endif
 
 
 uint a_offset, b_offset, d_offset, y_offset;
 
+vec4 load_b(const uint j, const uint iybs, const uint iqs, const bool lastiter, out bool OOB_y, out bool OOB_z, out bool OOB_w) {
+    // Check if the latter elements are OOB, and don't fetch B or accumulate it.
+    OOB_y = lastiter && (iybs + iqs + y_offset >= p.ncols);
+    OOB_z = lastiter && (iybs + iqs + y_offset*2 >= p.ncols);
+    OOB_w = lastiter && (iybs + iqs + y_offset*3 >= p.ncols);
+
+    if (!OOB_w) {
+        return vec4(FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset*2]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset*3]));
+    } else if (!OOB_z) {
+        return vec4(FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset*2]),
+                 0);
+    } else if (!OOB_y) {
+        return vec4(FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]),
+                 FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]),
+                 0, 0);
+    } else {
+        return vec4(FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]),
+                 0, 0, 0);
+    }
+}
+
 void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
 {
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
@@ -25,6 +51,8 @@ void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const
 
 #if K_PER_ITER == 8
 #if QUANT_R == 2
+        // Note that we end up fetching bogus elements here, but its fine as they'll be
+        // within an accessible block.
         const vec4 bv02 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
         const vec4 bv13 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4]);
         const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
@@ -34,18 +62,11 @@ void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const
         const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]);
 #endif
 #else
-        // Check if the second of the pair of elements is OOB, and don't fetch B or
-        // accumulate it. We still fetch a pair of elements for A, which is fine for
-        // quantized formats since they'll be within the same block. We should
-        // probably skip fetching the second element for F16/F32, but as of now we
-        // still do.
-        const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
-
-        FLOAT_TYPE b0 = 0, b1 = 0;
-        b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]);
-        if (!OOB) {
-            b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]);
-        }
+        bool OOB_y;
+        bool OOB_z;
+        bool OOB_w;
+
+        const vec4 b = load_b(j, iybs, iqs, lastiter, OOB_y, OOB_z, OOB_w);
 #endif
         uint ibi = first_row*p.ncols;
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -71,22 +92,60 @@ void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const
 
             temp[j][n] += rowtmp;
 #else
-            const vec2 v = dequantize(ib, iqs, a_offset);
-
-            // matrix multiplication
-            temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]);
-            if (!OOB) {
-                temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]);
+            if (!OOB_w) {
+                const vec4 v = dequantize4(ib, iqs, a_offset);
+                temp[j][n] += dot(v, b);
+            } else if (!OOB_z) {
+                const vec2 v0 = dequantize(ib, iqs, a_offset);
+                const FLOAT_TYPE v1 = dequantize1(ib + 2/QUANT_R, iqs, a_offset);
+                const vec3 v = vec3(v0.x, v0.y, v1);
+                const vec3 b0 = vec3(b.x, b.y, b.z);
+                temp[j][n] += dot(v, b0);
+            } else if (!OOB_y) {
+                const vec2 v0 = dequantize(ib, iqs, a_offset);
+                const vec2 b0 = vec2(b.x, b.y);
+                temp[j][n] += dot(v0, b0);
+            } else {
+                const FLOAT_TYPE v = dequantize1(ib, iqs, a_offset);
+                temp[j][n] = fma(v, b.x, temp[j][n]);
             }
 #endif
         }
     }
 }
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
+void iter_aligned_nonquant(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i)
+{
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
+        const uint iqs = 0; // quant index
+        const uint iybs = col; // y block start index
+
+        const vec4 b = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4];
+
+        uint ibi = first_row*p.ncols;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib = (ibi + col)/QUANT_K; // block index
+            ibi += p.ncols;
+
+            const vec4 v = dequantize4_2aligned(ib, iqs, a_offset);
+
+            // matrix multiplication
+            temp[j][n] += dot(v, b);
+        }
+    }
+}
+#endif
+
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
+    const bool is_aligned_nonquant =
+        p.batch_stride_b % 4 == 0 && b_offset % 4 == 0 &&
+        p.ncols % 4 == 0 && BLOCK_SIZE % 4 == 0 &&
+        K_PER_ITER == 4;
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
@@ -105,17 +164,26 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     int unroll_count = 4;
     uint unrolled_iters = num_iters & ~(unroll_count - 1);
 
-#if K_PER_ITER == 2
+    uint i = 0;
+
+#if K_PER_ITER == 4
     // If the K dimension is odd, we need lastiter==true on the last iteration
     // so OOB is computed correctly. Skip some unrolling to make that happen.
-    if ((p.ncols & 1) != 0 &&
+    if ((p.ncols & 3) != 0 &&
         unrolled_iters == num_iters &&
         unrolled_iters > 0) {
         unrolled_iters -= unroll_count;
     }
+    if (is_aligned_nonquant) {
+        while (i < unrolled_iters) {
+            // Manually partially unroll the loop
+            [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+                iter_aligned_nonquant(temp, first_row, num_rows, tid, i*K_PER_ITER);
+                i++;
+            }
+        }
+    } else {
 #endif
-
-    uint i = 0;
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
@@ -123,18 +191,30 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
             i++;
         }
     }
+#if K_PER_ITER == 4
+    }
+#endif
 
     unroll_count = 2;
     unrolled_iters = num_iters & ~(unroll_count - 1);
 
-#if K_PER_ITER == 2
-    if ((p.ncols & 1) != 0 &&
+#if K_PER_ITER == 4
+    if ((p.ncols & 3) != 0 &&
         unrolled_iters == num_iters &&
         unrolled_iters > 0) {
         unrolled_iters -= unroll_count;
     }
-#endif
 
+    if (is_aligned_nonquant) {
+        while (i < unrolled_iters && is_aligned_nonquant) {
+            // Manually partially unroll the loop
+            [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+                iter_aligned_nonquant(temp, first_row, num_rows, tid, i*K_PER_ITER);
+                i++;
+            }
+        }
+    } else {
+#endif
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
@@ -142,10 +222,25 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
             i++;
         }
     }
+#if K_PER_ITER == 4
+    }
+#endif
+
+#if K_PER_ITER == 4
+    if (is_aligned_nonquant) {
+        while (i < num_iters) {
+            iter_aligned_nonquant(temp, first_row, num_rows, tid, i*K_PER_ITER);
+            i++;
+        }
+    } else {
+#endif
     while (i < num_iters) {
         iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true);
         i++;
     }
+#if K_PER_ITER == 4
+    }
+#endif
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);
 }
@@ -164,6 +259,6 @@ void main() {
         if (first_row >= p.stride_d) {
             return;
         }
-        compute_outputs(first_row, p.stride_d - first_row);
+        compute_outputs(first_row, min(NUM_ROWS, p.stride_d - first_row));
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl

@@ -6,8 +6,8 @@
 #define MAT_VEC_FUSION_FLAGS_SCALE1 0x8
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-#if defined(A_TYPE_VEC4)
-layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
+#if defined(A_TYPEV4)
+layout (binding = 0) readonly buffer AV4 {A_TYPEV4 data_a_v4[];};
 #endif
 #if defined(A_TYPE_PACKED16)
 layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
@@ -17,11 +17,11 @@ layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32
 #endif
 
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-#ifdef B_TYPE_VEC2
-layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+#ifdef B_TYPEV2
+layout (binding = 1) readonly buffer BV2 {B_TYPEV2 data_b_v2[];};
 #endif
-#ifdef B_TYPE_VEC4
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#ifdef B_TYPEV4
+layout (binding = 1) readonly buffer BV4 {B_TYPEV4 data_b_v4[];};
 #endif
 
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -41,7 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
         const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
 
-        const FLOAT_TYPE_VEC2 dm = vec2(data_a[ib0 + i].dm);
+        const FLOAT_TYPEV2 dm = vec2(data_a[ib0 + i].dm);
 
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -14,7 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
-        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
+        const FLOAT_TYPEV2 dm = FLOAT_TYPEV2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -14,7 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
-        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
+        const FLOAT_TYPEV2 dm = FLOAT_TYPEV2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp

@@ -4,6 +4,7 @@
 #extension GL_EXT_integer_dot_product : require
 
 #define MMQ
+#define NEEDS_IQ1S_GRID_GPU
 #define B_TYPE block_q8_1_x4
 
 #include "mul_mat_vec_base.glsl"
@@ -27,13 +28,10 @@ vec2 cache_b_ds;
 
 #include "mul_mat_vecq_funcs.glsl"
 
-void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i) {
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint col, const uint b_qs_idx) {
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-        const uint col = i*BLOCK_SIZE + tid*K_PER_ITER;
-
         // Preload data_b block
         const uint b_block_idx = (j*p.batch_stride_b + col) / QUANT_K_Q8_1 + b_offset;
-        const uint b_qs_idx = tid % (32 / K_PER_ITER);
         const uint b_block_idx_outer = b_block_idx / 4;
         const uint b_block_idx_inner = b_block_idx % 4;
         cache_b_ds = vec2(data_b[b_block_idx_outer].ds[b_block_idx_inner]);
@@ -90,35 +88,35 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         }
     }
 
-    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
-    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+    const uint col_stride = K_PER_ITER * BLOCK_SIZE;
+    uint num_iters = p.ncols / col_stride;
+    if (num_iters * col_stride + K_PER_ITER * tid < p.ncols) {
         num_iters++;
     }
-    int unroll_count = 4;
-    uint unrolled_iters = num_iters & ~(unroll_count - 1);
 
-    uint i = 0;
-    while (i < unrolled_iters) {
+    const uint b_qs_idx = tid % (32 / K_PER_ITER);
+    uint col = tid * K_PER_ITER;
+    while (num_iters >= 4) {
         // Manually partially unroll the loop
-        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-            i++;
+        [[unroll]] for (uint k = 0; k < 4; ++k) {
+            iter(temp, first_row, num_rows, col, b_qs_idx);
+            col += col_stride;
         }
-    }
 
-    unroll_count = 2;
-    unrolled_iters = num_iters & ~(unroll_count - 1);
+        num_iters -= 4;
+    }
 
-    while (i < unrolled_iters) {
+    if (num_iters >= 2) {
         // Manually partially unroll the loop
-        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-            i++;
-        }
+        iter(temp, first_row, num_rows, col, b_qs_idx);
+        col += col_stride;
+        iter(temp, first_row, num_rows, col, b_qs_idx);
+        col += col_stride;
+        num_iters -= 2;
     }
-    while (i < num_iters) {
-        iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
-        i++;
+
+    if (num_iters > 0) {
+        iter(temp, first_row, num_rows, col, b_qs_idx);
     }
 
     reduce_result(temp, d_offset, first_row, num_rows, tid);

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl

@@ -11,8 +11,8 @@ FLOAT_TYPE get_dm(uint ib) {
 #endif
 
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+FLOAT_TYPEV2 get_dm(uint ib) {
+    return FLOAT_TYPEV2(data_a_packed32[ib].dm);
 }
 #endif
 
@@ -23,9 +23,9 @@ FLOAT_TYPE get_dm(uint ib) {
 #endif
 
 #if defined(DATA_A_Q2_K)
-FLOAT_TYPE_VEC2 get_dm(uint ib) {
+FLOAT_TYPEV2 get_dm(uint ib) {
     const uint ib_k = ib / 8;
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm);
+    return FLOAT_TYPEV2(data_a_packed32[ib_k].dm);
 }
 #endif
 
@@ -212,28 +212,40 @@ i32vec4 repack4(uint ib, uint iqs) {
     const uint qs_shift = ((iqs_k % 32) / 8) * 2;
     const uint hm_shift = iqs_k / 8;
 
+    const uvec4 qs = uvec4( uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2    ]) |
+                           (uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 1]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 2]) |
+                           (uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 3]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 4]) |
+                           (uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 5]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 6]) |
+                           (uint32_t(data_a_packed16[ib_k].qs[qs_idx * 2 + 7]) << 16));
+
+    const uvec4 hmask = uvec4( uint32_t(data_a_packed16[ib_k].hmask[iqs * 2    ]) |
+                              (uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 1]) << 16),
+                               uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 2]) |
+                              (uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 3]) << 16),
+                               uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 4]) |
+                              (uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 5]) << 16),
+                               uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 6]) |
+                              (uint32_t(data_a_packed16[ib_k].hmask[iqs * 2 + 7]) << 16));
+
     // bitwise OR to add 4 if hmask is set, subtract later
-    const i8vec2 vals00 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2    ] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2    ] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals01 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 1] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 1] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals10 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 2] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 2] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals11 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 3] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 3] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals20 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 4] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 4] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals21 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 5] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 5] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals30 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 6] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 6] >> hm_shift) & uint16_t(0x0101)) << 2));
-    const i8vec2 vals31 = unpack8(int16_t((data_a_packed16[ib_k].qs[qs_idx  * 2 + 7] >> qs_shift) & uint16_t(0x0303))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].hmask[iqs * 2 + 7] >> hm_shift) & uint16_t(0x0101)) << 2));
-
-    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y) - int8_t(4)),
-                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y) - int8_t(4)),
-                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y) - int8_t(4)),
-                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y) - int8_t(4)));
+    const uint vals0 = ((    qs.x >> qs_shift) & 0x03030303) |
+                       (((hmask.x >> hm_shift) & 0x01010101) << 2);
+    const uint vals1 = ((    qs.y >> qs_shift) & 0x03030303) |
+                       (((hmask.y >> hm_shift) & 0x01010101) << 2);
+    const uint vals2 = ((    qs.z >> qs_shift) & 0x03030303) |
+                       (((hmask.z >> hm_shift) & 0x01010101) << 2);
+    const uint vals3 = ((    qs.w >> qs_shift) & 0x03030303) |
+                       (((hmask.w >> hm_shift) & 0x01010101) << 2);
+
+    // Subtract 4 by twiddling bits rather than using re-packing as mesa
+    // compiles repacking poorly.
+    return i32vec4(int32_t(((vals0 ^ 0x80808080) - 0x04040404) ^ 0x80808080),
+                   int32_t(((vals1 ^ 0x80808080) - 0x04040404) ^ 0x80808080),
+                   int32_t(((vals2 ^ 0x80808080) - 0x04040404) ^ 0x80808080),
+                   int32_t(((vals3 ^ 0x80808080) - 0x04040404) ^ 0x80808080));
 }
 
 float get_d_scale(uint ib, uint iqs) {
@@ -296,15 +308,24 @@ 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));
-    }
 
-    return FLOAT_TYPE_VEC2(data_a_packed32[ib_k].dm) * FLOAT_TYPE_VEC2(scale_dm);
+    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);
 }
 
 FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {
@@ -334,27 +355,39 @@ i32vec4 repack4(uint ib, uint iqs) {
     const uint qh_idx = (iqs_k / 32) * 8 + iqs;
     const uint qh_shift = ((iqs_k % 32) / 8) * 2;
 
-    const i8vec2 vals00 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2    ] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2    ] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals01 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 1] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 1] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals10 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 2] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 2] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals11 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 3] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 3] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals20 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 4] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 4] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals21 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 5] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 5] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals30 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 6] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 6] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-    const i8vec2 vals31 = (unpack8(int16_t((data_a_packed16[ib_k].ql[ql_idx * 2 + 7] >> ql_shift) & uint16_t(0x0F0F))) |
-                          unpack8(int16_t(((data_a_packed16[ib_k].qh[qh_idx * 2 + 7] >> qh_shift) & uint16_t(0x0303)) << 4))) - int8_t(32);
-
-    return i32vec4(pack32(i8vec4(vals00.x, vals00.y, vals01.x, vals01.y)),
-                   pack32(i8vec4(vals10.x, vals10.y, vals11.x, vals11.y)),
-                   pack32(i8vec4(vals20.x, vals20.y, vals21.x, vals21.y)),
-                   pack32(i8vec4(vals30.x, vals30.y, vals31.x, vals31.y)));
+    const uvec4 ql = uvec4( uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2    ]) |
+                           (uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 1]) << 16),
+                            uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 2]) |
+                           (uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 3]) << 16),
+                            uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 4]) |
+                           (uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 5]) << 16),
+                            uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 6]) |
+                           (uint32_t(data_a_packed16[ib_k].ql[ql_idx * 2 + 7]) << 16));
+
+    const uvec4 qh = uvec4( uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2    ]) |
+                           (uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 1]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 2]) |
+                           (uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 3]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 4]) |
+                           (uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 5]) << 16),
+                            uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 6]) |
+                           (uint32_t(data_a_packed16[ib_k].qh[qh_idx * 2 + 7]) << 16));
+
+    const uint vals0 = (( ql.x >> ql_shift) & 0x0F0F0F0F) |
+                       (((qh.x >> qh_shift) & 0x03030303) << 4);
+    const uint vals1 = (( ql.y >> ql_shift) & 0x0F0F0F0F) |
+                       (((qh.y >> qh_shift) & 0x03030303) << 4);
+    const uint vals2 = (( ql.z >> ql_shift) & 0x0F0F0F0F) |
+                       (((qh.z >> qh_shift) & 0x03030303) << 4);
+    const uint vals3 = (( ql.w >> ql_shift) & 0x0F0F0F0F) |
+                       (((qh.w >> qh_shift) & 0x03030303) << 4);
+
+    // Subtract 32 by twiddling bits rather than using re-packing as mesa
+    // compiles repacking poorly.
+    return i32vec4(int32_t(((vals0 ^ 0x80808080) - 0x20202020) ^ 0x80808080),
+                   int32_t(((vals1 ^ 0x80808080) - 0x20202020) ^ 0x80808080),
+                   int32_t(((vals2 ^ 0x80808080) - 0x20202020) ^ 0x80808080),
+                   int32_t(((vals3 ^ 0x80808080) - 0x20202020) ^ 0x80808080));
 }
 
 float get_d_scale(uint ib, uint iqs) {
@@ -422,7 +455,7 @@ vec2 get_dm(uint ib, uint iqs) {
     const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
 
     // the -1 cancels out the bias in iq1s_grid_gpu
-    return FLOAT_TYPE_VEC2(dl, dl * (delta - 1));
+    return FLOAT_TYPEV2(dl, dl * (delta - 1));
 }
 
 FLOAT_TYPE mmvq_dot_product(const uint ib_a, const uint iqs) {