whispercpp 1.3.3 → 1.3.4

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -1,15 +1,15 @@
 #version 450
 
 #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
-#endif
+#include "rte.comp"
+
+#include "types.comp"
 
 layout (push_constant) uniform parameter
 {
+    BDA_STORAGE_T dst_addr;
     uint batch_offset; uint offset_delta;
     uint IC;
     uint IW; uint IH;
@@ -22,8 +22,6 @@ layout (push_constant) uniform parameter
     int d0; int d1;
 } p;
 
-#include "types.comp"
-
 layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
 const uint NUM_ITER = 512 / BLOCK_SIZE;
@@ -33,6 +31,10 @@ 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[];};
 
+#if BDA
+layout (buffer_reference) buffer D_ptr {D_TYPE d;};
+#endif
+
 void main() {
     const uint gidx = gl_GlobalInvocationID.x;
 
@@ -41,21 +43,19 @@ void main() {
     const uint ic = gl_GlobalInvocationID.z % p.IC;
 
     const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
-    const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH);
+    const BDA_OFFSET_T dst_base = ((BDA_OFFSET_T(batch) * p.OH + oh) * p.OW) * p.CHW + BDA_OFFSET_T(ic) * (p.KW * p.KH);
     const int oh_s1 = int(oh) * p.s1;
-    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint ksize = p.OW * p.KH;
 
     const uint base_linear_idx = gidx * NUM_ITER;
 
-    const uint max_ky = ksize / p.OW;
-
     uint current_kx = base_linear_idx / ksize;
     const uint rem = base_linear_idx - (current_kx * ksize);
     uint current_ky = rem / p.OW;
     uint current_ix = rem % p.OW;
 
     A_TYPE values[NUM_ITER];
-    uint offset_dst[NUM_ITER];
+    BDA_OFFSET_T offset_dst[NUM_ITER];
     [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
         values[idx] = A_TYPE(0);
     }
@@ -71,7 +71,7 @@ void main() {
         const uint iiw = current_ix * p.s0 + current_kx * p.d0 - p.p0;
         const uint iih = oh_s1 + current_ky * p.d1 - p.p1;
 
-        offset_dst[idx] = dst_base + current_ix * p.CHW + current_ky * p.KW + current_kx;
+        offset_dst[idx] = dst_base + BDA_OFFSET_T(current_ix) * p.CHW + current_ky * p.KW + current_kx;
 
         if ((iih < p.IH) && (iiw < p.IW)) {
             values[idx] = data_a[src_base + iih * p.IW + iiw];
@@ -79,7 +79,7 @@ void main() {
 
         if (++current_ix == p.OW) {
             current_ix = 0;
-            if (++current_ky == max_ky) {
+            if (++current_ky == p.KH) {
                 current_ky = 0;
                 current_kx++;
             }
@@ -94,7 +94,11 @@ void main() {
             continue;
         }
 
+#if BDA
+        D_ptr dst_addr = D_ptr(p.dst_addr + D_SIZE * offset_dst[idx]);
+        dst_addr.d = D_TYPE(values[idx]);
+#else
         data_d[offset_dst[idx]] = D_TYPE(values[idx]);
+#endif
     }
-
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp

@@ -0,0 +1,126 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "rte.comp"
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    BDA_STORAGE_T dst_addr;
+    uint32_t nb10;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t s2;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t p2;
+    uint32_t d0;
+    uint32_t d1;
+    uint32_t d2;
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t IC;
+    uint32_t KW;
+    uint32_t OH;
+    uint32_t KD_KH_KW;
+    uint32_t KH_KW;
+    uint32_t IC_KD_KH_KW;
+    uint32_t N_OD_OH;
+    uint32_t OD_OH;
+    uint32_t OD_OH_OW_IC_KD_KH_KW;
+    uint32_t OH_OW_IC_KD_KH_KW;
+    uint32_t OW_IC_KD_KH_KW;
+    uint32_t misalign_offsets;
+} p;
+
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+
+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[];};
+
+#if BDA
+layout (buffer_reference) buffer D_ptr {D_TYPE d;};
+#endif
+
+void main() {
+    const uint32_t i = gl_GlobalInvocationID.x;
+
+    uint32_t nb10 = p.nb10;
+    uint32_t nb11 = p.nb11;
+    uint32_t nb12 = p.nb12;
+    uint32_t nb13 = p.nb13;
+    uint32_t s0 = p.s0;
+    uint32_t s1 = p.s1;
+    uint32_t s2 = p.s2;
+    uint32_t p0 = p.p0;
+    uint32_t p1 = p.p1;
+    uint32_t p2 = p.p2;
+    uint32_t d0 = p.d0;
+    uint32_t d1 = p.d1;
+    uint32_t d2 = p.d2;
+    uint32_t IW = p.IW;
+    uint32_t IH = p.IH;
+    uint32_t ID = p.ID;
+    uint32_t IC = p.IC;
+    uint32_t KW = p.KW;
+    uint32_t OH = p.OH;
+    uint32_t KD_KH_KW = p.KD_KH_KW;
+    uint32_t KH_KW = p.KH_KW;
+    uint32_t IC_KD_KH_KW = p.IC_KD_KH_KW;
+    uint32_t N_OD_OH = p.N_OD_OH;
+    uint32_t OD_OH = p.OD_OH;
+    uint32_t OD_OH_OW_IC_KD_KH_KW = p.OD_OH_OW_IC_KD_KH_KW;
+    uint32_t OH_OW_IC_KD_KH_KW = p.OH_OW_IC_KD_KH_KW;
+    uint32_t OW_IC_KD_KH_KW = p.OW_IC_KD_KH_KW;
+
+    if (i >= IC_KD_KH_KW) {
+        return;
+    }
+
+    const uint32_t iic = i / KD_KH_KW;
+    const uint32_t ikd = (i - iic * KD_KH_KW) / KH_KW;
+    const uint32_t ikh = (i - iic * KD_KH_KW - ikd * KH_KW) / KW;
+    const uint32_t ikw = i % KW;
+
+    const uint32_t iow = gl_GlobalInvocationID.y;
+    for (uint32_t iz = gl_GlobalInvocationID.z; iz < N_OD_OH; iz += gl_NumWorkGroups.z) {
+        const uint32_t in_ = iz / OD_OH;
+        const uint32_t iod = (iz - in_*OD_OH) / OH;
+        const uint32_t ioh = iz % OH;
+
+        const uint32_t iiw = iow * s0 + ikw * d0 - p0;
+        const uint32_t iih = ioh * s1 + ikh * d1 - p1;
+        const uint32_t iid = iod * s2 + ikd * d2 - p2;
+
+        const BDA_OFFSET_T offset_dst = BDA_OFFSET_T(in_)*OD_OH_OW_IC_KD_KH_KW + BDA_OFFSET_T(iod)*OH_OW_IC_KD_KH_KW + BDA_OFFSET_T(ioh)*OW_IC_KD_KH_KW + BDA_OFFSET_T(iow)*IC_KD_KH_KW + iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw;
+
+        const uint32_t offset_src = (in_*IC + iic)*nb13 + iid*nb12 + iih*nb11 + iiw*nb10;
+#if BDA
+        D_ptr dst_addr = D_ptr(p.dst_addr + D_SIZE * offset_dst);
+        if (iih >= IH || iiw >= IW || iid >= ID) {
+            dst_addr.d = D_TYPE(0.0f);
+        } else {
+            dst_addr.d = D_TYPE(data_a[offset_src + get_aoffset()]);
+        }
+#else
+        if (iih >= IH || iiw >= IW || iid >= ID) {
+            data_d[offset_dst + get_doffset()] = D_TYPE(0.0f);
+        } else {
+            data_d[offset_dst + get_doffset()] = D_TYPE(data_a[offset_src + get_aoffset()]);
+        }
+#endif
+    }
+}

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

@@ -2,16 +2,30 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
 
+#if USE_SUBGROUP_ADD || USE_SUBGROUP_ADD_NO_SHMEM
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#endif
+
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
 #endif
 
 #include "types.comp"
 
+#ifndef MMQ
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#else
+layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
+#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[];};
+#endif
+#ifdef B_TYPE_VEC4
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#endif
 
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
@@ -88,9 +102,57 @@ layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 layout (constant_id = 1) const uint NUM_ROWS = 1;
 layout (constant_id = 2) const uint NUM_COLS = 1;
 
+#ifdef USE_SUBGROUP_ADD_NO_SHMEM
+void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+}
+#else
 shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
 
-void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // subgroupAdd is probably faster on devices that support it,
+    // particularly when the workgroup has more than one subgroup
+#if USE_SUBGROUP_ADD
+    // sum up partial sums within a subgroup
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    // Go through shared memory to sum partials across subgroups
+    if (gl_SubgroupInvocationID == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[j][n][gl_SubgroupID] = temp[j][n];
+            }
+        }
+    }
+    barrier();
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                temp[j][n] = FLOAT_TYPE(0);
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    temp[j][n] += tmpsh[j][n][s];
+                }
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+#else
     // sum up partial sums and write back result
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -115,4 +177,6 @@ void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32
             }
         }
     }
+#endif
 }
+#endif

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

@@ -26,6 +26,9 @@ layout (push_constant) uniform parameter
     uint ne12;
     uint b_offset;
     uint d_offset;
+    uint nb03;
+    uint nb13;
+    uint nb23;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -34,6 +37,7 @@ void main() {
     const uint tid       = gl_LocalInvocationID.x;
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
+    const uint i3        = gl_WorkGroupID.x;
     const uint channel_x = channel / p.channel_x_divisor;
     const uint channel_y = channel % p.ne12;
 
@@ -41,7 +45,7 @@ void main() {
     const uint nrows_dst = p.nrows_x;
     const uint row_dst   = row_x;
 
-    const uint idst = channel*nrows_dst + row_dst;
+    const uint idst = i3*p.nb23 + channel*nrows_dst + row_dst;
 
     FLOAT_TYPE temp = 0.0f;
 
@@ -58,8 +62,8 @@ void main() {
 
                 const uint row_y = col_x;
 
-                const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel_y*p.channel_stride_y + row_y;
+                const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+                const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -74,8 +78,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -91,8 +95,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

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

@@ -0,0 +1,140 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_integer_dot_product : require
+
+#define MMQ
+#define B_TYPE block_q8_1_x4
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+#define K_PER_ITER 8
+
+#include "mul_mmq_funcs.comp"
+
+uint a_offset, b_offset, d_offset;
+
+int32_t cache_b_qs[2];
+vec2 cache_b_ds;
+
+void iter(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 + 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 % 4;
+        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]);
+
+#if QUANT_R == 2
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx + 4];
+#else
+        cache_b_qs[0] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2];
+        cache_b_qs[1] = data_b[b_block_idx_outer].qs[b_block_idx_inner * 8 + b_qs_idx * 2 + 1];
+#endif
+
+        uint ibi = first_row*p.ncols;
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint a_block_idx = (ibi + col)/QUANT_K + a_offset;
+            ibi += p.ncols;
+
+            int32_t q_sum = 0;
+#if QUANT_R == 2
+            const i32vec2 data_a_qs = repack(a_block_idx, b_qs_idx);
+            q_sum += dotPacked4x8EXT(data_a_qs.x,
+                                     cache_b_qs[0]);
+            q_sum += dotPacked4x8EXT(data_a_qs.y,
+                                     cache_b_qs[1]);
+#else
+            int32_t data_a_qs = repack(a_block_idx, b_qs_idx * 2);
+            q_sum += dotPacked4x8EXT(data_a_qs,
+                                     cache_b_qs[0]);
+            data_a_qs = repack(a_block_idx, b_qs_idx * 2 + 1);
+            q_sum += dotPacked4x8EXT(data_a_qs,
+                                     cache_b_qs[1]);
+#endif
+
+#if QUANT_AUXF == 1
+            temp[j][n] += mul_q8_1(q_sum,  get_d(a_block_idx), cache_b_ds, 4);
+#else
+            temp[j][n] += mul_q8_1(q_sum, get_dm(a_block_idx), cache_b_ds, 4);
+#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);
+    a_offset /= QUANT_K;
+    b_offset /= QUANT_K_Q8_1;
+
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = FLOAT_TYPE(0.0f);
+        }
+    }
+
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + 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) {
+        // 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_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+
+#if K_PER_ITER == 2
+    if ((p.ncols & 1) != 0 &&
+        unrolled_iters == num_iters &&
+        unrolled_iters > 0) {
+        unrolled_iters -= unroll_count;
+    }
+#endif
+
+    while (i < unrolled_iters) {
+        // 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++;
+        }
+    }
+    while (i < num_iters) {
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER);
+        i++;
+    }
+
+    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);
+
+    // 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);
+    }
+}