vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations

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

@@ -1,5 +1,7 @@
 #version 450
 
+#extension GL_EXT_control_flow_attributes : enable
+
 #include "types.comp"
 #include "generic_binary_head.comp"
 #include "dequant_funcs.comp"

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -0,0 +1,90 @@
+#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        const uint qh = data_a[ibi].qh[ib32];
+        const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]);
+        const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]);
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint8_t sign = sign16[l];
+            const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300);
+            const uvec2 grid = iq2s_grid[qs];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
+
+            [[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 =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, 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;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp

@@ -0,0 +1,87 @@
+#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint nibble_shift = 4 * (itid & 1);
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;
+        const float db = d * (0.5 + scale) * 0.25;
+
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs = data_a[ibi].qs[2 * itid + l];
+            const uint sign = qs >> 9;
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
+            const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
+
+            [[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 =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, 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;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp

@@ -0,0 +1,87 @@
+#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 itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 16 * itid;
+    const uint ib32 = itid / 2; // 0..7
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint signscale = pack32(u16vec2(
+            data_a_packed16[ibi].qs[4 * ib32 + 2],
+            data_a_packed16[ibi].qs[4 * ib32 + 3]));
+        const float db = d * 0.25 * (0.5 + (signscale >> 28));
+        [[unroll]] for (uint l = 0; l < 2; ++l) {
+            const uint qs = data_a[ibi].qs[8 * ib32 + 2 * (itid & 1) + l];
+            const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7);
+            const uint sign7 = bitCount(sign);
+            const vec4 grid0 = vec4(unpack8(iq2xxs_grid[qs].x));
+            const vec4 grid1 = vec4(unpack8(iq2xxs_grid[qs].y));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                      FLOAT_TYPE(0.0)))))))));
+                temp[j][n] = fma(db, 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;
+
+    // 16 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 16;  // 0...15
+    const uint ix = tid / 16;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

@@ -0,0 +1,90 @@
+#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 float d = float(data_a[ibi].d);
+        const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
+        const float dscale = d * (1 + 2 * scale);
+        const uint qh = data_a[ibi].qh[ib32];
+        FLOAT_TYPE sum[NUM_COLS];
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            sum[j] = 0.0;
+        }
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]);
+            const uint sign = data_a[ibi].signs[4 * ib32 + l];
+            const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
+            const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)]));
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                sum[j] =
+                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
+                      sum[j]))))))));
+            }
+        }
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            temp[j][n] = fma(dscale, sum[j], temp[j][n]);
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}