whispercpp 1.3.2 → 1.3.4

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -19,21 +19,13 @@ if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     message(STATUS "Enabling bfloat16 glslc support")
 endif()
+if (GGML_VULKAN_SHADER_DEBUG_INFO)
+    add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+    message(STATUS "Enabling shader debug info")
+endif()
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
-
-# Configure output directories for MSVC builds
-if(MSVC)
-    # Get the main project's runtime output directory if possible
-    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
-        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
-            string(TOUPPER ${CONFIG} CONFIG)
-            set_target_properties(${TARGET} PROPERTIES
-                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-        endforeach()
-    endif()
-endif()

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

@@ -1,20 +1,34 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
+#if ADD_RMS
+#extension GL_KHR_shader_subgroup_arithmetic : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
 
 #include "types.comp"
 #include "generic_binary_head.comp"
 
 const uint num_threads = 256;
 
+layout (binding = 3, std430) buffer PartialBuf {float partial_sums[];};
+
 layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
 
+#if ADD_RMS
+// XXX TODO this could be sized based on number of subgroups, but that't not considered a constant
+shared FLOAT_TYPE sumsh[num_threads];
+#endif
+
 void main() {
     uint idx = get_idx();
+    uint orig_idx = idx;
 
     // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
     const uint num_iter = 2;
 
+    FLOAT_TYPE sum_sq = 0;
+
     [[unroll]] for (uint i = 0; i < num_iter; ++i) {
         if (idx >= p.ne) {
             continue;
@@ -22,8 +36,34 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
+        FLOAT_TYPE sum = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]);
+        sum_sq += sum*sum;
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
 
         idx += num_threads;
     }
+
+#if ADD_RMS
+    if (p.param3 != 0) {
+        // reduce the sum within each subgroup, then across subgroups
+        const uint NumSubgroups = num_threads / gl_SubgroupSize;
+        sum_sq = subgroupAdd(sum_sq);
+        if (gl_SubgroupInvocationID == 0) {
+            sumsh[gl_SubgroupID] = sum_sq;
+        }
+        barrier();
+        [[unroll]] for (uint s = NumSubgroups / 2; s > 0; s >>= 1) {
+            if (gl_SubgroupID < s && gl_SubgroupInvocationID == 0) {
+                sum_sq += sumsh[gl_SubgroupID + s];
+                sumsh[gl_SubgroupID] = sum_sq;
+            }
+            barrier();
+        }
+
+        if (gl_SubgroupID == 0 && gl_SubgroupInvocationID == 0) {
+            partial_sums[orig_idx / (num_iter * num_threads)] = sum_sq;
+        }
+    }
+#endif
 }

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

@@ -0,0 +1,42 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ne0;
+    uint ne1;
+    uint s01;
+    uint s02;
+    uint s11;
+    uint s21;
+} p;
+
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) readonly buffer Z {int32_t data_c[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i1 = gl_WorkGroupID.x;
+    const uint i2 = gl_WorkGroupID.y;
+
+    const uint i11 = data_c[i1 + i2 * p.s21];
+
+    const uint s1 = p.ne0;
+    const uint s2 = p.ne0 * p.ne1;
+
+    const uint d0 = i1 * s1 + i2 * s2;
+    const uint a0 = i1 * p.s01 + i2 * p.s02;
+    const uint b0 = i11 * p.s11;
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < p.ne0; i0 += BLOCK_SIZE) {
+        data_d[d0 + i0] = data_a[a0 + i0] + data_b[b0 + i0];
+    }
+}

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

@@ -5,6 +5,8 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
+#define FLT_MAX 3.402823466e+38F
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -19,19 +21,26 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
-    if (col >= p.KX) {
+    if (row >= p.KY) {
         return;
     }
-    A_TYPE amax = data_a[row*p.KX + col];
-    tmp[col] = col;
+
+    A_TYPE amax = -FLT_MAX;
+    uint acol = col;
+
+    if (col < p.KX) {
+        amax = data_a[row*p.KX + col];
+    }
 
     for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
         A_TYPE val = data_a[row*p.KX + i];
         if (val > amax) {
             amax = val;
-            tmp[col] = i;
+            acol = i;
         }
     }
+
+    tmp[col] = acol;
     tmpmax[col] = amax;
 
     barrier();

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

@@ -1,22 +1,24 @@
 #version 450
+#extension GL_EXT_control_flow_attributes : enable
 
 #include "types.comp"
 
-#define BLOCK_SIZE 1024
+layout(constant_id = 0) const int BLOCK_SIZE = 1024;
+layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
 #define ASC 0
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
-    uint ncols_pad;
     uint order;
 } p;
 
 shared int dst_row[BLOCK_SIZE];
+shared A_TYPE a_sh[BLOCK_SIZE];
 
 void swap(uint idx0, uint idx1) {
     int tmp = dst_row[idx0];
@@ -24,7 +26,7 @@ void swap(uint idx0, uint idx1) {
     dst_row[idx1] = tmp;
 }
 
-void main() {
+void argsort(bool needs_bounds_check) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
     const uint row = gl_WorkGroupID.y;
@@ -32,38 +34,46 @@ void main() {
     const uint row_offset = row * p.ncols;
 
     // initialize indices
-    if (col < p.ncols_pad) {
-        dst_row[col] = col;
-    }
+    dst_row[col] = col;
+    a_sh[col] = data_a[row_offset + col];
     barrier();
 
-    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
-        for (uint j = k / 2; j > 0; j /= 2) {
-            const uint ixj = col ^ j;
-            if (col < p.ncols_pad && ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= p.ncols ||
-                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                } else {
-                    if (dst_row[ixj] >= p.ncols ||
-                        (dst_row[col] < p.ncols && (p.order == ASC ?
-                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
-                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
-                    ) {
-                        swap(col, ixj);
-                    }
-                }
+    uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
+    [[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
+        uint num_inner_loop_iters = outer_idx + 1;
+        [[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
+            const int ixj = int(col ^ j);
+
+            int idx_0 = (col & k) == 0 ? col : ixj;
+            int idx_1 = (col & k) == 0 ? ixj : col;
+
+            int sh_idx_0 = dst_row[idx_0];
+            int sh_idx_1 = dst_row[idx_1];
+            bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
+            bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
+
+            if ((idx_0_oob ||
+                (!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
+                swap(idx_0, idx_1);
             }
+
             barrier();
         }
     }
 
     if (col < p.ncols) {
-        data_d[row_offset + col] = dst_row[col];
+        if (p.order == ASC) {
+            data_d[row_offset + col] = dst_row[col];
+        } else {
+            data_d[row_offset + p.ncols - col - 1] = dst_row[col];
+        }
+    }
+}
+
+void main() {
+    if (p.ncols == BLOCK_SIZE) {
+        argsort(false);
+    } else {
+        argsort(true);
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -0,0 +1,349 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#ifdef COOPMAT2
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_KHR_memory_scope_semantics : enable
+#endif
+
+#ifdef USE_COLLECTIVES
+#    extension GL_KHR_shader_subgroup_shuffle : enable
+#endif
+
+#include "types.comp"
+
+// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
+layout(binding = 0) readonly buffer A {
+    A_TYPE knl_data[];
+};  // src0 - kernel:   [KW, KH, Cin, Cout] for conv_2d, [KW, KH, Cout, Cin] for conv_transposed_2d
+
+layout(binding = 1) readonly buffer B {
+    B_TYPE src_data[];
+};  // src1 - input:    [W, H, Cin, N] -- channel_first format
+
+layout(binding = 2) writeonly buffer D {
+    D_TYPE dst_data[];
+};  // dst - result:    [OW, OH, Cout, N]
+
+layout(push_constant) uniform parameter {
+    // I/O channels, batch size
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t N;
+
+    // Tensor spatial sizes: kernel, input, output
+    uint32_t KW;
+    uint32_t KH;
+    uint32_t W;
+    uint32_t H;
+    uint32_t OW;
+    uint32_t OH;
+
+    // Parameters: stride, padding, dilation - 0=y, 1=x
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t d0;
+    uint32_t d1;
+
+    // Strides in elements
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    // fastdiv helper values
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
+#ifdef TRANSPOSE
+    uint32_t s0mp; uint32_t s0L;
+    uint32_t s1mp; uint32_t s1L;
+#endif
+}
+
+p;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+// Blocktile sizes
+layout(constant_id = 1) const uint BS_K            = 128;
+layout(constant_id = 2) const uint BS_CRS          = 16;
+layout(constant_id = 3) const uint BS_NPQ          = 128;
+// Thread-tile sizes
+layout(constant_id = 4) const uint TS_K            = 8;
+layout(constant_id = 5) const uint use_collectives = 1;
+layout(constant_id = 6) const uint SHMEM_PAD       = 4;
+
+uint32_t       tid     = gl_LocalInvocationID.x;
+const uint32_t WG_SIZE = gl_WorkGroupSize.x;
+
+uint splitWork(uint work_size, uint block_size) {
+    return (block_size + work_size - 1) / block_size;
+}
+
+uint32_t K   = p.Cout;
+uint32_t CRS = p.Cin * p.KH * p.KW;
+uint32_t NPQ = p.N * p.OH * p.OW;
+
+uint32_t n_elems_out = K * NPQ;
+
+// Number of blocktiles per input
+uint32_t NB_CRS = splitWork(CRS, BS_CRS);
+
+#ifdef COOPMAT2
+#define SHMEM_TYPE float16_t
+#else
+#define SHMEM_TYPE float
+#endif
+
+const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
+const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
+
+const uint32_t Ash_numel = BS_K * BS_CRS;
+const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
+
+const uint32_t Ash_len = BS_K * Ash_stride;
+const uint32_t Bsh_len = BS_CRS * Bsh_stride;
+
+shared SHMEM_TYPE Ash[Ash_len];  // K x CRS
+shared SHMEM_TYPE Bsh[Bsh_len];  // CRS x NPQ
+
+// Threadtile sizes
+const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
+
+// Number of threadtiles per blocktile
+const uint32_t NT_K   = BS_K / TS_K;
+const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
+
+/*
+Compute
+KxCRS @ CRSxNPQ = K x NPQ
+K=Cout
+C=Cin
+R,S=KH,KW
+P,Q=OH,OW
+*/
+
+uint32_t B_idx_K   = gl_WorkGroupID.x;
+uint32_t B_idx_NPQ = gl_WorkGroupID.y;
+
+uint32_t T_y = tid / NT_NPQ;
+uint32_t T_x = tid % NT_NPQ;
+
+uint32_t       Ar    = tid / BS_CRS;
+uint32_t       Ac    = tid % BS_CRS;
+const uint32_t ArpWg = WG_SIZE / BS_CRS;
+
+uint32_t       Br    = tid / BS_NPQ;
+uint32_t       Bc    = tid % BS_NPQ;
+const uint32_t BrpWg = WG_SIZE / BS_NPQ;
+
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
+#ifdef COOPMAT2
+#define ACC_TYPE float16_t
+
+ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
+{
+    uint32_t K_idx   = B_idx_K * BS_K + r;
+    uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
+    uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+    uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+    uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+    uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+    if (K_idx < K && NPQ_idx < NPQ) {
+        dst_data[dst_idx] = D_TYPE(elem);
+    }
+    return elem;
+}
+#endif
+
+void main() {
+#ifdef COOPMAT2
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
+    matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
+#else
+    float regC[TS_K][TS_NPQ];
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            regC[T_ly][T_lx] = 0.0;
+        }
+    }
+#endif
+    /* Advance block in CRS dim */
+    for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+        uint32_t CRS_idx_a;
+        uint32_t Cin_idx_a;
+        uint32_t KH_idx_a;
+        uint32_t KW_idx_a;
+
+#ifdef USE_COLLECTIVES
+        uint32_t cached_CRS_idx;
+        uint32_t cached_Cin_idx;
+        uint32_t cached_KH_idx;
+        uint32_t cached_KW_idx;
+        if (use_collectives == 1) {
+            cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
+            cached_Cin_idx                = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
+            uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
+            cached_KH_idx                 = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
+            cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
+
+            CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
+            Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
+            KH_idx_a  = subgroupShuffle(cached_KH_idx, Ac);
+            KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
+        } else {
+            CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+            Cin_idx_a              = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+            KH_idx_a               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
+            KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
+        }
+#else
+        CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+        Cin_idx_a     = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
+        CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+        KH_idx_a      = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
+        KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
+#endif
+
+        /* Load kernel to A_block: (BS_K x BS_CRS)*/
+        for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+            uint32_t B_ly    = r_offset + Ar;
+            uint32_t B_lx    = Ac;
+            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+#ifdef TRANSPOSE
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + K_idx * p.nb02 + Cin_idx_a * p.nb03, K * CRS - 1);
+#else
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
+#endif
+            float    val     = knl_data[knl_idx];
+            if (K_idx >= K || CRS_idx_a >= CRS) {
+                val = 0.0;
+            }
+            Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        /* Load input to B_block: (BS_CRS x BS_NPQ) */
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+            uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
+            uint32_t B_lx          = Bc;
+            uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
+            uint32_t N_idx         = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+            uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
+            uint32_t OH_idx        = fastdiv(NPQ_remainder, p.OWmp, p.OWL); // divide by p.OW;
+            uint32_t OW_idx        = NPQ_remainder - OH_idx * p.OW;
+
+            uint32_t CRS_idx_b;
+            uint32_t Cin_idx_b;
+            uint32_t KH_idx_b;
+            uint32_t KW_idx_b;
+#ifdef USE_COLLECTIVES
+            if (use_collectives == 1) {
+                CRS_idx_b = subgroupShuffle(cached_CRS_idx, r_offset + Br);
+                Cin_idx_b = subgroupShuffle(cached_Cin_idx, r_offset + Br);
+                KH_idx_b  = subgroupShuffle(cached_KH_idx, r_offset + Br);
+                KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
+            } else {
+                CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+                Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
+                uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+                KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
+                KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+            }
+#else
+            CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+            Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+            KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
+            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+#endif
+
+#ifdef TRANSPOSE
+            uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * p.d1 + p.p1;
+            uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * p.d0 + p.p0;
+            uint32_t H_idx = fastdiv(H_idx_x_s1, p.s1mp, p.s1L);
+            uint32_t W_idx = fastdiv(W_idx_x_s0, p.s0mp, p.s0L);
+#else
+            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
+            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+#endif
+            uint32_t src_idx =
+                min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
+            float val = src_data[src_idx];
+            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
+                || H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
+#ifdef TRANSPOSE
+                || (H_idx_x_s1 - H_idx * p.s1 != 0) || (W_idx_x_s0 - W_idx * p.s0 != 0)
+#endif
+                ) {
+                val = 0.0;
+            }
+            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
+        }
+        barrier();
+#ifdef COOPMAT2
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
+        coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
+
+        coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
+        coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+        matC = coopMatMulAdd(matA, matB, matC);
+#else
+        if (T_y * TS_K < K) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                float regA[TS_K];
+                float regB[TS_NPQ];
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+                }
+                for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                    regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+                }
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                        regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    }
+                }
+            }
+        }
+#endif
+        barrier();
+    }
+    /* Save C* */
+#ifdef COOPMAT2
+    coopMatPerElementNV(matC, matC, perElemOpStore);
+#else
+    if (T_y * TS_K < K) {
+        for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+                uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+                uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+                uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+                uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+                uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+                if (K_idx < K && NPQ_idx < NPQ) {
+                    dst_data[dst_idx] = regC[T_ly][T_lx];
+                }
+            }
+        }
+    }
+#endif
+}