vulkan: initial support for IQ1_S and IQ1_M quantizations

.github/workflows/build.yml

@@ -401,7 +401,7 @@ jobs:
         run: |
           cd build
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 1800
+          ctest -L main --verbose --timeout 2700
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04

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

@@ -12,7 +12,7 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
     init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;

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

@@ -217,7 +217,7 @@ void quantize(uint dst_idx, uint src_idx)
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
     init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;

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

@@ -1,6 +1,9 @@
 #if !defined(DATA_A_F32) && !defined(DATA_A_F16)
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
 #endif
+#if defined(DATA_A_IQ1_M)
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
 
 #include "types.comp"
 
@@ -88,6 +91,83 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_IQ1_S)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+    const int i8 = int(iqs % 8);
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const float dl = float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint idxhi = bitfieldExtract(qh, 3 * int(ib8 & 3), 3);
+    const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+    // Signed bitfield extract.
+    const ivec2 gvec = ivec2(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2)
+    );
+    return dl * (vec2(gvec) + delta);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+    const int i8 = int(iqs % 8);
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const float dl = 2 * bitfieldExtract(qh, 12, 3) + 1;
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);
+    // Signed bitfield extract.
+    const ivec4 gvec = ivec4(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2),
+      bitfieldExtract(grid, 2 * (i8 + 2), 2),
+      bitfieldExtract(grid, 2 * (i8 + 3), 2)
+    );
+    return dl * (vec4(gvec) + delta);
+}
+#endif
+
+#if defined(DATA_A_IQ1_M)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib8 = iqs / 8;
+    const uint ib16 = iqs / 16;
+    const int i8 = int(iqs % 8);
+    const uint sc = data_a[a_offset + ib].scales[iqs / 64];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+    // Signed bitfield extract.
+    const ivec2 gvec = ivec2(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2)
+    );
+    return dl * (vec2(gvec) + delta);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib8 = iqs / 8;
+    const uint ib16 = iqs / 16;
+    const int i8 = int(iqs % 8);
+    const uint sc = data_a[a_offset + ib].scales[iqs / 64];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+    // Signed bitfield extract.
+    const ivec4 gvec = ivec4(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2),
+      bitfieldExtract(grid, 2 * (i8 + 2), 2),
+      bitfieldExtract(grid, 2 * (i8 + 3), 2)
+    );
+    return dl * (vec4(gvec) + delta);
+}
+#endif
+
 #if defined(DATA_A_IQ2_XXS)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint ib32 = iqs / 32;
@@ -357,7 +437,16 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_M)
+vec2 get_dm(uint ib, uint a_offset) {
+    const uint16_t[4] scales = data_a[a_offset + ib].scales;
+    const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+    const float d = float(uint16BitsToHalf(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)));
+    return vec2(d, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), 0);
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -301,6 +301,56 @@ float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2
     return ret;
 }
 
+#if defined(DATA_A_IQ1_S)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S {
+   block_iq1_s block;
+};
+
+float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+
+    const uint ib32 = idx / 32;
+    const uint ib8 = idx / 8;
+
+    const uint qh = bl.block.qh[ib32];
+    const uint qs = bl.block.qs[ib8];
+    const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)];
+
+    float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta));
+    return ret;
+}
+#endif
+
+#if defined(DATA_A_IQ1_M)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M {
+   block_iq1_m block;
+};
+
+float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const u16vec4 scales = u16vec4(bl.block.scales[0], bl.block.scales[1], bl.block.scales[2], bl.block.scales[3]) >> 12;
+    const float16_t d = uint16BitsToHalf(scales.x | (scales.y << 4) | (scales.z << 8) | (scales.w << 12));
+    const uint idx = coordInBlock[1];
+
+    const uint ib8 = idx / 8;
+    const uint ib16 = idx / 16;
+    const int i8 = int(idx % 8);
+    const uint sc = bl.block.scales[ib8 / 8];
+    const uint qs = bl.block.qs[ib8];
+    const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint grid = iq1s_grid[qs | ((qh & 7) << 8)];
+
+    float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta));
+    return ret;
+}
+#endif
+
 #if defined(DATA_A_IQ2_XXS)
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
    block_iq2_xxs block;
@@ -512,6 +562,10 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncQ5_K
 #elif defined(DATA_A_Q6_K)
 #define dequantFuncA dequantFuncQ6_K
+#elif defined(DATA_A_IQ1_S)
+#define dequantFuncA dequantFuncIQ1_S
+#elif defined(DATA_A_IQ1_M)
+#define dequantFuncA dequantFuncIQ1_M
 #elif defined(DATA_A_IQ2_XXS)
 #define dequantFuncA dequantFuncIQ2_XXS
 #elif defined(DATA_A_IQ2_XS)

ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp

@@ -0,0 +1,42 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq1_m data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint ib64 = ib32 / 2;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    const uint16_t[4] scales = data_a[ib].scales;
+    const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+    const float d = float(uint16BitsToHalf(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)));
+
+    const uint sc = data_a[ib].scales[ib64];
+    [[unroll]] for (int l = 0; l < 4; ++l) {
+        const uint ib16 = 2 * ib32 + l / 2;
+        const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1);
+        const uint qh = data_a[ib].qh[ib16] >> (4 * (l & 1));
+        const uint qs = data_a[ib].qs[4 * ib32 + l];
+        const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+        const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+        [[unroll]] for (int j = 0; j < 8; ++j) {
+            data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq1_s data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    uint qh = data_a[ib].qh[ib32];
+    const float d = float(data_a[ib].d);
+    const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        const uint qs = data_a[ib].qs[4 * ib32 + l];
+        const uint hi = bitfieldExtract(qh, 3 * int(l), 3);
+        const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]);
+        [[unroll]] for (int j = 0; j < 8; ++j) {
+            data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -104,7 +104,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

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

@@ -12,7 +12,7 @@ void main() {
     const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
     const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

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

@@ -133,7 +133,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -0,0 +1,82 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint16_t[4] scales = data_a[ibi].scales;
+        const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+        const float d = float(uint16BitsToHalf(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)));
+
+        const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
+            const uint qs = data_a[ibi].qs[4 * ib32 + l];
+            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+            const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
+
+            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int k = 0; k < 4; ++k) {
+                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
+                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
+                }
+                temp[j][n] = fma(dl, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}