whispercpp 1.3.3 → 1.3.5

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

@@ -1,33 +1,42 @@
 #version 450
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 void main() {
     const uint i00 = gl_GlobalInvocationID.x;
-    const uint i10 = gl_GlobalInvocationID.y;
-    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
-    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
     if (i00 >= p.ne00) {
         return;
     }
 
-    const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+    uint gid_z = gl_GlobalInvocationID.z;
+    while (gid_z < p.ne11 * p.ne12) {
+        uint gid_y = gl_GlobalInvocationID.y;
+        while (gid_y < p.ne10) {
+            const uint i10 = gid_y;
+            const uint i11 = gid_z / p.ne12;
+            const uint i12 = gid_z % p.ne12;
 
-    const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-    const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+            const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+
+            const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+            const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
 #if defined(DATA_A_BF16)
-    FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
+            TEMP_TYPE v = TEMP_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
 #else
-    FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]);
+            TEMP_TYPE v = TEMP_TYPE(data_a[a_offset + i00]);
 #endif
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[d_offset + i00] = D_TYPE(v);
+            data_d[d_offset + i00] = D_TYPE(v);
 #else
-    data_d[d_offset + i00] = D_TYPE(v);
+            data_d[d_offset + i00] = D_TYPE(v);
 #endif
+            gid_y += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+        gid_z += gl_WorkGroupSize.z * gl_NumWorkGroups.z;
+    }
 }

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

@@ -2,17 +2,14 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
-#include "types.comp"
-#include "generic_binary_head.comp"
-#include "dequant_funcs.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 void main() {
     const uint i00 = (gl_GlobalInvocationID.x)*2;
-    const uint i10 = gl_GlobalInvocationID.y;
-    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
-    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
@@ -22,20 +19,33 @@ void main() {
         return;
     }
 
-    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+    uint gid_z = gl_GlobalInvocationID.z;
+    while (gid_z < p.ne11 * p.ne12) {
+        uint gid_y = gl_GlobalInvocationID.y;
+        while (gid_y < p.ne10) {
+            const uint i10 = gid_y;
+            const uint i11 = gid_z / p.ne12;
+            const uint i12 = gid_z % p.ne12;
 
-    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+            const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-    const uint ib = a_offset + i00/QUANT_K; // block index
-    const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
-    const uint iybs = i00 - i00%QUANT_K; // dst block start index
-    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+            const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+            const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
-    vec2 v = dequantize(ib, iqs, 0);
-    const vec2 dm = get_dm(ib, 0);
-    v = v * dm.x + dm.y;
+            const uint ib = a_offset + i00/QUANT_K; // block index
+            const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
+            const uint iybs = i00 - i00%QUANT_K; // dst block start index
+            const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
-    data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
-    data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
+            vec2 v = dequantize(ib, iqs, 0);
+            const vec2 dm = get_dm(ib, 0);
+            v = v * dm.x + dm.y;
+
+            data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
+            data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
+
+            gid_y += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
+        gid_z += gl_WorkGroupSize.z * gl_NumWorkGroups.z;
+    }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{glu_head.comp → glu_head.glsl}

@@ -1,5 +1,7 @@
 #extension GL_EXT_shader_16bit_storage : require
 
+#include "rte.glsl"
+
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -12,4 +14,6 @@ layout (push_constant) uniform parameter
     uint ne00;
     uint ne20;
     uint mode;
+    float alpha;
+    float limit;
 } p;

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

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512

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

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, 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[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(min(1.0f, max(0.0f, (x + 3.0f) / 6.0f)));
+}

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

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.glsl"
+#include "types.glsl"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, 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[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * min(1.0f, max(0.0f, (x + 3.0f) / 6.0f)));
+}

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

@@ -1,15 +1,14 @@
 #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.glsl"
+#include "types.glsl"
 
 layout (push_constant) uniform parameter
 {
+    BDA_STORAGE_T dst_addr;
     uint batch_offset; uint offset_delta;
     uint IC;
     uint IW; uint IH;
@@ -20,10 +19,9 @@ layout (push_constant) uniform parameter
     int s0; int s1;
     int p0; int p1;
     int d0; int d1;
+    uint batch_IC;
 } p;
 
-#include "types.comp"
-
 layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
 const uint NUM_ITER = 512 / BLOCK_SIZE;
@@ -33,29 +31,31 @@ 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[];};
 
-void main() {
+#if BDA
+layout (buffer_reference) buffer D_ptr {D_TYPE d;};
+#endif
+
+void im2col(const uint y, const uint z) {
     const uint gidx = gl_GlobalInvocationID.x;
 
-    const uint oh = gl_GlobalInvocationID.y;
-    const uint batch = gl_GlobalInvocationID.z / p.IC;
-    const uint ic = gl_GlobalInvocationID.z % p.IC;
+    const uint oh = y;
+    const uint batch = z / p.IC;
+    const uint ic = 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 +94,23 @@ 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
     }
+}
 
+void main() {
+    uint y = gl_GlobalInvocationID.y;
+    while (y < p.OH) {
+        uint z = gl_GlobalInvocationID.z;
+        while (z < p.batch_IC) {
+            im2col(y, z);
+            z += gl_NumWorkGroups.z;
+        }
+        y += gl_NumWorkGroups.y;
+    }
 }

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

@@ -0,0 +1,125 @@
+#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.glsl"
+#include "types.glsl"
+
+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/l2_norm.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512

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

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

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

@@ -0,0 +1,18 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const float val = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(log(val));
+}

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

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 const uint num_threads = 256;
 

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

@@ -2,7 +2,8 @@
 
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "mul_mat_vec_base.comp"
+#include "mul_mat_vec_base.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -86,7 +87,6 @@ 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;
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;