LIBXSMM_API_INTERN
void libxsmm_sparse_csc_asparse_innerloop_four_vector( libxsmm_generated_code* io_generated_code,
const libxsmm_gemm_descriptor* i_xgemm_desc,
const unsigned int i_k,
const unsigned int i_z,
const unsigned int* i_row_idx,
const unsigned int* i_column_idx ) {
char l_new_code[512];
int l_max_code_length = 511;
int l_code_length = 0;
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
unsigned int l_i;
unsigned int l_z = i_z;
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && defined(__AVX__)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m256d c%u_%u = _mm256_loadu_pd(&C[(l_n*%u)+%u]);\n", i_k, i_z, (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + i_z] );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m256d a%u_%u = _mm256_loadu_pd(&A[%u]);\n", i_k, i_z, i_column_idx[i_k] + i_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " c%u_%u = _mm256_add_pd(c%u_%u, _mm256_mul_pd(a%u_%u, b%u));\n", i_k, i_z, i_k, i_z, i_k, i_z, i_k );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " _mm256_storeu_pd(&C[(l_n*%u)+%u], c%u_%u);\n", (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + i_z], i_k, i_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#endif\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && !defined(__AVX__)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
for ( l_i = 0; l_i < 2; l_i++ ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m128d c%u_%u = _mm_loadu_pd(&C[(l_n*%u)+%u]);\n", i_k, l_z, (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + l_z] );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m128d a%u_%u = _mm_loadu_pd(&A[%u]);\n", i_k, l_z, i_column_idx[i_k] + l_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " c%u_%u = _mm_add_pd(c%u_%u, _mm_mul_pd(a%u_%u, b%u));\n", i_k, l_z, i_k, l_z, i_k, l_z, i_k );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " _mm_storeu_pd(&C[(l_n*%u)+%u], c%u_%u);\n", (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + l_z], i_k, l_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_z += 2;
}
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#endif\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
} else {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m128 c%u_%u = _mm_loadu_ps(&C[(l_n*%u)+%u]);\n", i_k, i_z, (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + i_z] );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " __m128 a%u_%u = _mm_loadu_ps(&A[%u]);\n", i_k, i_z, i_column_idx[i_k] + i_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " c%u_%u = _mm_add_ps(c%u_%u, _mm_mul_ps(a%u_%u, b%u));\n", i_k, i_z, i_k, i_z, i_k, i_z, i_k );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " _mm_storeu_ps(&C[(l_n*%u)+%u], c%u_%u);\n", (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[i_k] + i_z], i_k, i_z );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
}
LIBXSMM_API_INTERN
void libxsmm_generator_spgemm_csc_bsparse( libxsmm_generated_code* io_generated_code,
const libxsmm_gemm_descriptor* i_xgemm_desc,
const char* i_arch,
const unsigned int* i_row_idx,
const unsigned int* i_column_idx,
const double* i_values ) {
unsigned int l_n;
unsigned int l_z;
unsigned int l_column_elements;
unsigned int l_flop_count = 0;
char l_new_code[512];
int l_max_code_length = 511;
int l_code_length = 0;
LIBXSMM_UNUSED(i_values);
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " unsigned int l_m = 0;\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* reset C if beta is zero */
if (0 != (LIBXSMM_GEMM_FLAG_BETA_0 & i_xgemm_desc->flags)) { /* Beta=0 */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " unsigned int l_n = 0;\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_n = 0; l_n < %u; l_n++) {\n", (unsigned int)i_xgemm_desc->n);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
if ( i_xgemm_desc->m > 1 ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma vector aligned\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_m = 0; l_m < %u; l_m++) { C[(l_n*%u)+l_m] = 0.0; }\n", (unsigned int)i_xgemm_desc->m, (unsigned int)i_xgemm_desc->ldc);
} else {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_m = 0; l_m < %u; l_m++) { C[(l_n*%u)+l_m] = 0.0f; }\n", (unsigned int)i_xgemm_desc->m, (unsigned int)i_xgemm_desc->ldc);
}
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " }\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* determine the correct simd pragma for each architecture */
if ( ( strcmp( i_arch, "noarch" ) == 0 ) ||
( strcmp( i_arch, "wsm" ) == 0 ) ||
( strcmp( i_arch, "snb" ) == 0 ) ||
( strcmp( i_arch, "hsw" ) == 0 ) ) {
if ( i_xgemm_desc->m > 7 ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd vectorlength(8)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
} else if ( i_xgemm_desc->m > 3 ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd vectorlength(4)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
} else if ( i_xgemm_desc->m > 1 ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd vectorlength(2)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
} else {}
if ( (i_xgemm_desc->m > 1) &&
((LIBXSMM_GEMM_FLAG_ALIGN_A & i_xgemm_desc->flags) != 0) &&
((LIBXSMM_GEMM_FLAG_ALIGN_C & i_xgemm_desc->flags) != 0) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma vector aligned\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
} else if ( ( strcmp( i_arch, "knl" ) == 0 ) ||
( strcmp( i_arch, "skx" ) == 0 ) ||
( strcmp( i_arch, "clx" ) == 0 ) ||
( strcmp( i_arch, "cpx" ) == 0 ) ) {
if ( (i_xgemm_desc->m > 1) &&
((LIBXSMM_GEMM_FLAG_ALIGN_A & i_xgemm_desc->flags) != 0) &&
((LIBXSMM_GEMM_FLAG_ALIGN_C & i_xgemm_desc->flags) != 0) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd vectorlength(32)\n #pragma vector aligned\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
} else {
LIBXSMM_HANDLE_ERROR( io_generated_code, LIBXSMM_ERR_ARCH );
return;
}
/* generate the actuel kernel */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_m = 0; l_m < %u; l_m++) {\n", (unsigned int)i_xgemm_desc->m);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
for ( l_n = 0; l_n < (unsigned int)i_xgemm_desc->n; l_n++ ) {
l_column_elements = i_column_idx[l_n+1] - i_column_idx[l_n];
for ( l_z = 0; l_z < l_column_elements; l_z++ ) {
/* check k such that we just use rows which actually need to be multiplied */
if ( i_row_idx[i_column_idx[l_n] + l_z] < (unsigned int)i_xgemm_desc->k ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " C[%u+l_m] += A[%u+l_m] * B[%u];\n", l_n * i_xgemm_desc->ldc, i_row_idx[i_column_idx[l_n] + l_z]*i_xgemm_desc->lda, i_column_idx[l_n] + l_z);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_flop_count += 2;
}
}
}
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " }\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* add flop counter */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "\n#ifndef NDEBUG\n#ifdef _OPENMP\n#pragma omp atomic\n#endif\nlibxsmm_num_total_flops += %u;\n#endif\n", l_flop_count * (unsigned int)i_xgemm_desc->m);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
LIBXSMM_API_INTERN
void libxsmm_generator_spgemm_csc_bsparse_soa_avx256_512( libxsmm_generated_code* io_generated_code,
const libxsmm_gemm_descriptor* i_xgemm_desc,
const char* i_arch,
const unsigned int* i_row_idx,
const unsigned int* i_column_idx,
const void* i_values ) {
unsigned int l_n = 0;
unsigned int l_k = 0;
unsigned int l_soa_width = 0;
unsigned int l_max_cols = 0;
unsigned int l_n_processed = 0;
unsigned int l_n_limit = 0;
unsigned int l_n_chunks = 0;
unsigned int l_n_chunksize = 0;
unsigned int l_found_mul = 0;
unsigned int l_max_reg_block = 0;
libxsmm_micro_kernel_config l_micro_kernel_config;
libxsmm_loop_label_tracker l_loop_label_tracker;
libxsmm_gp_reg_mapping l_gp_reg_mapping;
LIBXSMM_UNUSED(i_values);
/* select soa width */
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
if ( strcmp(i_arch, "knl") == 0 ||
strcmp(i_arch, "knm") == 0 ||
strcmp(i_arch, "skx") == 0 ||
strcmp(i_arch, "clx") == 0 ||
strcmp(i_arch, "cpx") == 0 ) {
l_soa_width = 8;
l_max_reg_block = 28;
} else {
l_soa_width = 4;
l_max_reg_block = 14;
}
} else {
if ( strcmp(i_arch, "knl") == 0 ||
strcmp(i_arch, "knm") == 0 ||
strcmp(i_arch, "skx") == 0 ||
strcmp(i_arch, "clx") == 0 ||
strcmp(i_arch, "cpx") == 0 ) {
l_soa_width = 16;
l_max_reg_block = 28;
} else {
l_soa_width = 8;
l_max_reg_block = 14;
}
}
/* define gp register mapping */
libxsmm_reset_x86_gp_reg_mapping( &l_gp_reg_mapping );
/* matching calling convention on Linux */
#if defined(_WIN32) || defined(__CYGWIN__)
l_gp_reg_mapping.gp_reg_a = LIBXSMM_X86_GP_REG_RCX;
l_gp_reg_mapping.gp_reg_b = LIBXSMM_X86_GP_REG_RDX;
l_gp_reg_mapping.gp_reg_c = LIBXSMM_X86_GP_REG_R8;
/* TODO: full support for Windows calling convention */
l_gp_reg_mapping.gp_reg_a_prefetch = LIBXSMM_X86_GP_REG_RDI;
l_gp_reg_mapping.gp_reg_b_prefetch = LIBXSMM_X86_GP_REG_RSI;
#else /* match calling convention on Linux */
l_gp_reg_mapping.gp_reg_a = LIBXSMM_X86_GP_REG_RDI;
l_gp_reg_mapping.gp_reg_b = LIBXSMM_X86_GP_REG_RSI;
l_gp_reg_mapping.gp_reg_c = LIBXSMM_X86_GP_REG_RDX;
l_gp_reg_mapping.gp_reg_a_prefetch = LIBXSMM_X86_GP_REG_RCX;
l_gp_reg_mapping.gp_reg_b_prefetch = LIBXSMM_X86_GP_REG_R8;
#endif
l_gp_reg_mapping.gp_reg_c_prefetch = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_mloop = LIBXSMM_X86_GP_REG_R12;
l_gp_reg_mapping.gp_reg_nloop = LIBXSMM_X86_GP_REG_R13;
l_gp_reg_mapping.gp_reg_kloop = LIBXSMM_X86_GP_REG_R14;
l_gp_reg_mapping.gp_reg_help_0 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_1 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_2 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_3 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_4 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_5 = LIBXSMM_X86_GP_REG_UNDEF;
/* define loop_label_tracker */
libxsmm_reset_loop_label_tracker( &l_loop_label_tracker );
/* define the micro kernel code gen properties */
libxsmm_generator_gemm_init_micro_kernel_config_fullvector( &l_micro_kernel_config, i_xgemm_desc, i_arch, 0 );
/* get max column in C */
l_max_cols = i_xgemm_desc->n;
for ( l_n = 0; l_n < i_xgemm_desc->n; l_n++ ) {
if ( i_column_idx[l_n] == i_column_idx[i_xgemm_desc->n] ) {
l_max_cols = l_n+1;
}
}
/* calculate the chunk size of current columns to work on */
l_n_chunks = ( (l_max_cols % l_max_reg_block) == 0 ) ? (l_max_cols / l_max_reg_block) : (l_max_cols / l_max_reg_block) + 1;
assert(0 != l_n_chunks); /* mute static analysis (division-by-zero); such invalid input must be caught upfront */
l_n_chunksize = ( (l_max_cols % l_n_chunks) == 0 ) ? (l_max_cols / l_n_chunks) : (l_max_cols / l_n_chunks) + 1;
/* open asm */
libxsmm_x86_instruction_open_stream( io_generated_code, &l_gp_reg_mapping, i_arch, i_xgemm_desc->prefetch );
/* m loop */
libxsmm_x86_instruction_register_jump_back_label( io_generated_code, &l_loop_label_tracker );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_mloop, 1 );
/* loop over n-blocks */
l_n_processed = 0;
l_n_limit = l_n_chunksize;
while ( l_n_processed < l_max_cols ) {
#if 0
printf("l_max_cols: %i, l_n_processed: %i, l_n_limit: %i\n", l_max_cols, l_n_processed, l_n_limit);
#endif
/* load C accumulator */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
if (0 != (LIBXSMM_GEMM_FLAG_BETA_0 & i_xgemm_desc->flags)) { /* Beta=0 */
libxsmm_x86_instruction_vec_compute_reg( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vxor_instruction,
l_micro_kernel_config.vector_name,
l_n, l_n, l_n );
} else {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.c_vmove_instruction,
l_gp_reg_mapping.gp_reg_c,
LIBXSMM_X86_GP_REG_UNDEF, 0,
(l_n_processed + l_n)*l_soa_width*l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_n, 0, 1, 0 );
}
}
/* do dense soa times sparse multiplication */
for ( l_k = 0; l_k < (unsigned int)i_xgemm_desc->k; l_k++ ) {
unsigned int l_found_qmadd = 0;
unsigned int l_col_k = 0;
unsigned int l_column_active[28];
int l_nnz_idx[28][4];
/* reset helpers */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
l_column_active[l_n] = 0;
l_nnz_idx[l_n][0] = -1; l_nnz_idx[l_n][1] = -1; l_nnz_idx[l_n][2] = -1; l_nnz_idx[l_n][3] = -1;
}
l_found_mul = 0;
/* let's figure out if we can apply qmadd when being sin F32 setting and on KNM */
if ( (l_k < ((unsigned int)i_xgemm_desc->k - 3)) &&
(l_micro_kernel_config.instruction_set == LIBXSMM_X86_AVX512_KNM) &&
(LIBXSMM_GEMM_PRECISION_F32 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) ) {
/* loop over the columns of B/C */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
unsigned int l_found = 0;
unsigned int l_acol_k = 0;
unsigned int l_col_elements = i_column_idx[l_n_processed+l_n+1] - i_column_idx[l_n_processed+l_n];
unsigned int l_cur_column = i_column_idx[l_n_processed+l_n];
for ( l_col_k = 0; l_col_k < l_col_elements; l_col_k++ ) {
for ( l_acol_k = l_found; l_acol_k < 4; l_acol_k++ ) {
if ( (l_k + l_acol_k) == i_row_idx[l_cur_column + l_col_k] ) {
l_nnz_idx[l_n][l_acol_k] = l_cur_column + l_col_k;
l_found = l_acol_k+1;
}
if (l_found == 4) {
l_col_k = l_col_elements;
}
}
}
/* let's check if we can apply qmadd in col l_n */
if ( (l_nnz_idx[l_n][0] != -1) && (l_nnz_idx[l_n][1] != -1) && (l_nnz_idx[l_n][2] != -1) && (l_nnz_idx[l_n][3] != -1) ) {
l_column_active[l_n] = 2;
l_found_qmadd = 1;
l_found_mul = 1;
} else {
/* let's check if we have at least one entry in the column that matches one of the four entries */
if ( (l_nnz_idx[l_n][0] != -1) || (l_nnz_idx[l_n][1] != -1) || (l_nnz_idx[l_n][2] != -1) || (l_nnz_idx[l_n][3] != -1) ) {
l_column_active[l_n] = 1;
l_found_mul = 1;
} else {
l_column_active[l_n] = 0;
}
}
}
}
if ( l_found_qmadd == 0 ) {
/* loop over the columns of B/C */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
unsigned int l_col_elements = i_column_idx[l_n_processed+l_n+1] - i_column_idx[l_n_processed+l_n];
unsigned int l_cur_column = i_column_idx[l_n_processed+l_n];
/* search for entries matching that k */
for ( l_col_k = 0; l_col_k < l_col_elements; l_col_k++ ) {
if ( l_k == i_row_idx[l_cur_column + l_col_k] ) {
l_nnz_idx[l_n][0] = l_cur_column + l_col_k;
l_col_k = l_col_elements;
}
}
/* let's check if we have an entry in the column that matches the k from A */
if ( (l_nnz_idx[l_n][0] != -1) ) {
l_column_active[l_n] = 1;
l_found_mul = 1;
} else {
l_column_active[l_n] = 0;
}
}
}
/* First case: we can use qmadd */
if ( l_found_qmadd != 0 ) {
unsigned int l_lcl_k = 0;
for ( l_lcl_k = 0; l_lcl_k < 4; l_lcl_k++ ) {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.a_vmove_instruction,
l_gp_reg_mapping.gp_reg_a,
LIBXSMM_X86_GP_REG_UNDEF, 0,
(l_k+l_lcl_k)*l_soa_width*l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_max_reg_block+l_lcl_k, 0, 1, 0 );
}
/* loop over the columns of B/C */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
/* issue a qmadd */
if ( l_column_active[l_n] == 2 ) {
libxsmm_x86_instruction_vec_compute_qfma( io_generated_code,
l_micro_kernel_config.instruction_set,
LIBXSMM_X86_INSTR_V4FMADDPS,
l_gp_reg_mapping.gp_reg_b,
LIBXSMM_X86_GP_REG_UNDEF,
0,
l_nnz_idx[l_n][0] * l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_max_reg_block,
l_n );
} else if ( l_column_active[l_n] == 1 ) {
for ( l_lcl_k = 0; l_lcl_k < 4; l_lcl_k++ ) {
if ( l_nnz_idx[l_n][l_lcl_k] != -1 ) {
libxsmm_x86_instruction_vec_compute_mem( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vmul_instruction,
1,
l_gp_reg_mapping.gp_reg_b,
LIBXSMM_X86_GP_REG_UNDEF,
0,
l_nnz_idx[l_n][l_lcl_k] * l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_max_reg_block+l_lcl_k,
l_n );
}
}
}
}
/* increment by additional 3 columns */
l_k += 3;
} else if ( l_found_mul != 0 ) {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.a_vmove_instruction,
l_gp_reg_mapping.gp_reg_a,
LIBXSMM_X86_GP_REG_UNDEF, 0,
l_k*l_soa_width*l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_max_reg_block, 0, 1, 0 );
/* loop over the columns of B/C */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
if ( l_nnz_idx[l_n][0] != -1 ) {
if ( strcmp(i_arch, "knl") == 0 ||
strcmp(i_arch, "knm") == 0 ||
strcmp(i_arch, "skx") == 0 ||
strcmp(i_arch, "clx") == 0 ||
strcmp(i_arch, "cpx") == 0 ) {
libxsmm_x86_instruction_vec_compute_mem( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vmul_instruction,
1,
l_gp_reg_mapping.gp_reg_b,
LIBXSMM_X86_GP_REG_UNDEF,
0,
l_nnz_idx[l_n][0] * l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_max_reg_block,
l_n );
} else if ( strcmp(i_arch, "hsw") == 0 ) {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.b_vmove_instruction,
l_gp_reg_mapping.gp_reg_b,
LIBXSMM_X86_GP_REG_UNDEF, 0,
l_nnz_idx[l_n][0] * l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
15, 0, 1, 0 );
libxsmm_x86_instruction_vec_compute_reg( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vmul_instruction,
l_micro_kernel_config.vector_name,
l_max_reg_block,
15,
l_n );
} else {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.b_vmove_instruction,
l_gp_reg_mapping.gp_reg_b,
LIBXSMM_X86_GP_REG_UNDEF, 0,
l_nnz_idx[l_n][0] * l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
15, 0, 1, 0 );
libxsmm_x86_instruction_vec_compute_reg( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vmul_instruction,
l_micro_kernel_config.vector_name,
l_max_reg_block,
15,
15 );
libxsmm_x86_instruction_vec_compute_reg( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.vadd_instruction,
l_micro_kernel_config.vector_name,
15,
l_n,
l_n );
}
}
}
} else {
/* shouldn't happen */
}
}
/* store C accumulator */
for ( l_n = 0; l_n < l_n_limit - l_n_processed; l_n++ ) {
libxsmm_x86_instruction_vec_move( io_generated_code,
l_micro_kernel_config.instruction_set,
l_micro_kernel_config.c_vmove_instruction,
l_gp_reg_mapping.gp_reg_c,
LIBXSMM_X86_GP_REG_UNDEF, 0,
(l_n_processed + l_n)*l_soa_width*l_micro_kernel_config.datatype_size,
l_micro_kernel_config.vector_name,
l_n, 0, 0, 1 );
}
/* adjust n progression */
l_n_processed += l_n_chunksize;
l_n_limit = LIBXSMM_MIN(l_n_processed + l_n_chunksize, l_max_cols);
}
/* advance C pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_c,
l_micro_kernel_config.datatype_size*l_soa_width*i_xgemm_desc->ldc);
/* advance A pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_a,
l_micro_kernel_config.datatype_size*l_soa_width*i_xgemm_desc->lda);
/* close m loop */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_cmp_instruction, l_gp_reg_mapping.gp_reg_mloop, i_xgemm_desc->m );
libxsmm_x86_instruction_jump_back_to_label( io_generated_code, l_micro_kernel_config.alu_jmp_instruction, &l_loop_label_tracker );
/* close asm */
libxsmm_x86_instruction_close_stream( io_generated_code, &l_gp_reg_mapping, i_arch, i_xgemm_desc->prefetch );
}
LIBXSMM_API_INTERN
void libxsmm_generator_spgemm_csc_asparse( libxsmm_generated_code* io_generated_code,
const libxsmm_gemm_descriptor* i_xgemm_desc,
const char* i_arch,
const unsigned int* i_row_idx,
const unsigned int* i_column_idx,
const double* i_values ) {
char l_new_code[512];
int l_max_code_length = 511;
int l_code_length = 0;
unsigned int l_k;
unsigned int l_flop_count = 0;
LIBXSMM_UNUSED(i_arch);
LIBXSMM_UNUSED(i_values);
/* loop over columns in C in generated code, we fully unroll inside each column */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " unsigned int l_n = 0;\n #pragma nounroll_and_jam\n for ( l_n = 0; l_n < %u; l_n++) {\n", (unsigned int)i_xgemm_desc->n);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* reset the current column in C if needed */
if (0 != (LIBXSMM_GEMM_FLAG_BETA_0 & i_xgemm_desc->flags)) { /* Beta=0 */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " unsigned int l_m = 0;\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
if ( i_xgemm_desc->m > 1 ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " #pragma simd\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_m = 0; l_m < %u; l_m++) {\n C[(l_n*%u)+l_m] = 0.0;\n }\n", (unsigned int)i_xgemm_desc->m, (unsigned int)i_xgemm_desc->ldc);
} else {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " for ( l_m = 0; l_m < %u; l_m++) {\n C[(l_n*%u)+l_m] = 0.0f;\n }\n", (unsigned int)i_xgemm_desc->m, (unsigned int)i_xgemm_desc->ldc);
}
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
assert(0 != i_column_idx);
/* loop over columns in A, rows in B and fully unroll */
for ( l_k = 0; l_k < (unsigned int)i_xgemm_desc->k; l_k++ ) {
unsigned int l_column_elements = i_column_idx[l_k + 1] - i_column_idx[l_k];
unsigned int l_z = 0;
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) || defined(__AVX__)\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
if ( l_column_elements > 0 ) {
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && defined(__AVX__)\n __m256d b%u = _mm256_broadcast_sd(&B[(l_n*%u)+%u]);\n#endif\n", l_k, (unsigned int)i_xgemm_desc->ldb, l_k);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && !defined(__AVX__)\n __m128d b%u = _mm_loaddup_pd(&B[(l_n*%u)+%u]);\n#endif\n", l_k, (unsigned int)i_xgemm_desc->ldb, l_k);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
} else {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && defined(__AVX__)\n __m128 b%u = _mm_broadcast_ss(&B[(l_n*%u)+%u]);\n#endif\n", l_k, (unsigned int)i_xgemm_desc->ldb, l_k);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#if defined(__SSE3__) && !defined(__AVX__)\n __m128 b%u = _mm_load_ss(&B[(l_n*%u)+%u]); b%u = _mm_shuffle_ps(b%u, b%u, 0x00);\n#endif\n", l_k, (unsigned int)i_xgemm_desc->ldb, l_k, l_k, l_k, l_k);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
}
/* loop over the columns of A and look for vectorization potential */
for ( l_z = 0; l_z < l_column_elements; l_z++ ) {
assert(0 != i_row_idx);
/* 4 element vector might be possible */
if ( (l_z < (l_column_elements - 3)) && (l_column_elements > 3) ) {
/* check for 256bit vector instruction */
if ((i_row_idx[i_column_idx[l_k] + l_z] + 1 == i_row_idx[i_column_idx[l_k] + l_z + 1]) &&
(i_row_idx[i_column_idx[l_k] + l_z] + 2 == i_row_idx[i_column_idx[l_k] + l_z + 2]) &&
(i_row_idx[i_column_idx[l_k] + l_z] + 3 == i_row_idx[i_column_idx[l_k] + l_z + 3]) &&
(i_row_idx[i_column_idx[l_k] + l_z + 3] < (unsigned int)i_xgemm_desc->m)) {
libxsmm_sparse_csc_asparse_innerloop_four_vector(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
l_z += 3;
/* check for 128bit vector instruction */
} else if ((i_row_idx[i_column_idx[l_k] + l_z] + 1 == i_row_idx[i_column_idx[l_k] + l_z + 1]) &&
(i_row_idx[i_column_idx[l_k] + l_z + 1] < (unsigned int)i_xgemm_desc->m) ) {
libxsmm_sparse_csc_asparse_innerloop_two_vector(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
l_z++;
/* scalar instruction */
} else {
if ( (i_row_idx[i_column_idx[l_k] + l_z] < (unsigned int)i_xgemm_desc->m) ) {
libxsmm_sparse_csc_asparse_innerloop_scalar(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
}
}
/* 2 element vector might be possible */
} else if ( (l_z < (l_column_elements - 1)) && (l_column_elements > 1)) {
/* check for 128bit vector instruction */
if ((i_row_idx[i_column_idx[l_k] + l_z] + 1 == i_row_idx[i_column_idx[l_k] + l_z + 1]) &&
(i_row_idx[i_column_idx[l_k] + l_z + 1] < (unsigned int)i_xgemm_desc->m) ) {
libxsmm_sparse_csc_asparse_innerloop_two_vector(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
l_z++;
/* scalar instruction */
} else {
if ( (i_row_idx[i_column_idx[l_k] + l_z] < (unsigned int)i_xgemm_desc->m) ) {
libxsmm_sparse_csc_asparse_innerloop_scalar(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
}
}
/* scalar anyways */
} else {
if ( (i_row_idx[i_column_idx[l_k] + l_z] < (unsigned int)i_xgemm_desc->m) ) {
libxsmm_sparse_csc_asparse_innerloop_scalar(io_generated_code, i_xgemm_desc, l_k, l_z, i_row_idx, i_column_idx);
}
}
}
/* C fallback code */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#else\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* loop over the columns of A */
for ( l_z = 0; l_z < l_column_elements; l_z++ ) {
if ( (i_row_idx[i_column_idx[l_k] + l_z] < (unsigned int)i_xgemm_desc->m) ) {
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " C[(l_n*%u)+%u] += A[%u] * B[(l_n*%u)+%u];\n", (unsigned int)i_xgemm_desc->ldc, i_row_idx[i_column_idx[l_k] + l_z], i_column_idx[l_k] + l_z, (unsigned int)i_xgemm_desc->ldb, l_k );
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
l_flop_count += 2;
}
}
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "#endif\n\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, " }\n");
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
/* add flop counter */
l_code_length = LIBXSMM_SNPRINTF(l_new_code, l_max_code_length, "\n#ifndef NDEBUG\n#ifdef _OPENMP\n#pragma omp atomic\n#endif\nlibxsmm_num_total_flops += %u;\n#endif\n", l_flop_count * i_xgemm_desc->n);
libxsmm_append_code_as_string( io_generated_code, l_new_code, l_code_length );
}
LIBXSMM_API_INTERN void libxsmm_generator_gemm_rm_ac_soa_avx256_512( libxsmm_generated_code* io_generated_code,
const libxsmm_gemm_descriptor* i_xgemm_desc,
const char* i_arch ) {
unsigned int l_soa_width = 0;
unsigned int l_max_reg_block = 0;
unsigned int l_n1_range = 0;
unsigned int l_n2_range = 0;
unsigned int l_n1_block = 0;
unsigned int l_n2_block = 0;
libxsmm_micro_kernel_config l_micro_kernel_config;
libxsmm_loop_label_tracker l_loop_label_tracker;
libxsmm_gp_reg_mapping l_gp_reg_mapping;
/* select soa width */
if ( LIBXSMM_GEMM_PRECISION_F64 == LIBXSMM_GETENUM_INP( i_xgemm_desc->datatype ) ) {
if ( strcmp(i_arch, "knl") == 0 ||
strcmp(i_arch, "knm") == 0 ||
strcmp(i_arch, "skx") == 0 ||
strcmp(i_arch, "clx") == 0 ||
strcmp(i_arch, "cpx") == 0 ) {
l_soa_width = 8;
l_max_reg_block = 28;
} else {
l_soa_width = 4;
l_max_reg_block = 14;
}
} else {
if ( strcmp(i_arch, "knl") == 0 ||
strcmp(i_arch, "knm") == 0 ||
strcmp(i_arch, "skx") == 0 ||
strcmp(i_arch, "clx") == 0 ||
strcmp(i_arch, "cpx") == 0 ) {
l_soa_width = 16;
l_max_reg_block = 28;
} else {
l_soa_width = 8;
l_max_reg_block = 14;
}
}
/* define gp register mapping */
libxsmm_reset_x86_gp_reg_mapping( &l_gp_reg_mapping );
/* matching calling convention on Linux */
#if defined(_WIN32) || defined(__CYGWIN__)
l_gp_reg_mapping.gp_reg_a = LIBXSMM_X86_GP_REG_RCX;
l_gp_reg_mapping.gp_reg_b = LIBXSMM_X86_GP_REG_RDX;
l_gp_reg_mapping.gp_reg_c = LIBXSMM_X86_GP_REG_R8;
/* TODO: full support for Windows calling convention */
l_gp_reg_mapping.gp_reg_a_prefetch = LIBXSMM_X86_GP_REG_RDI;
l_gp_reg_mapping.gp_reg_b_prefetch = LIBXSMM_X86_GP_REG_RSI;
#else /* match calling convention on Linux */
l_gp_reg_mapping.gp_reg_a = LIBXSMM_X86_GP_REG_RDI;
l_gp_reg_mapping.gp_reg_b = LIBXSMM_X86_GP_REG_RSI;
l_gp_reg_mapping.gp_reg_c = LIBXSMM_X86_GP_REG_RDX;
l_gp_reg_mapping.gp_reg_a_prefetch = LIBXSMM_X86_GP_REG_RCX;
l_gp_reg_mapping.gp_reg_b_prefetch = LIBXSMM_X86_GP_REG_R8;
#endif
l_gp_reg_mapping.gp_reg_c_prefetch = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_mloop = LIBXSMM_X86_GP_REG_R12;
l_gp_reg_mapping.gp_reg_nloop = LIBXSMM_X86_GP_REG_R13;
l_gp_reg_mapping.gp_reg_kloop = LIBXSMM_X86_GP_REG_R14;
l_gp_reg_mapping.gp_reg_help_0 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_1 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_2 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_3 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_4 = LIBXSMM_X86_GP_REG_UNDEF;
l_gp_reg_mapping.gp_reg_help_5 = LIBXSMM_X86_GP_REG_UNDEF;
/* define loop_label_tracker */
libxsmm_reset_loop_label_tracker( &l_loop_label_tracker );
/* define the micro kernel code gen properties */
libxsmm_generator_gemm_init_micro_kernel_config_fullvector( &l_micro_kernel_config, i_xgemm_desc, i_arch, 0 );
/* calculate the chunk size of current columns to work on */
if ( libxsmm_compute_equalized_blocking( i_xgemm_desc->n, l_max_reg_block, &l_n1_range, &l_n1_block, &l_n2_range, &l_n2_block ) ) {
LIBXSMM_HANDLE_ERROR( io_generated_code, LIBXSMM_ERR_N_BLOCK );
return;
}
/* open asm */
libxsmm_x86_instruction_open_stream( io_generated_code, &l_gp_reg_mapping, i_arch, i_xgemm_desc->prefetch );
/* m loop */
libxsmm_x86_instruction_register_jump_back_label( io_generated_code, &l_loop_label_tracker );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_mloop, 1 );
/* loop over n-blocks */
if ( l_n1_block == i_xgemm_desc->n ) {
/* no N loop at all */
libxsmm_generator_gemm_rm_ac_soa_avx256_512_kloop( io_generated_code, &l_loop_label_tracker, &l_gp_reg_mapping, &l_micro_kernel_config, i_xgemm_desc,
i_arch, l_soa_width, i_xgemm_desc->n );
} else if ( (l_n1_range > 0) && (l_n2_range > 0) ) {
/* reset n loop */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_mov_instruction, l_gp_reg_mapping.gp_reg_nloop, 0 );
/* we have two ranges */
/* first range */
libxsmm_x86_instruction_register_jump_back_label( io_generated_code, &l_loop_label_tracker );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_nloop, l_n1_block );
libxsmm_generator_gemm_rm_ac_soa_avx256_512_kloop( io_generated_code, &l_loop_label_tracker, &l_gp_reg_mapping, &l_micro_kernel_config, i_xgemm_desc,
i_arch, l_soa_width, l_n1_block );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_cmp_instruction, l_gp_reg_mapping.gp_reg_nloop, l_n1_range );
libxsmm_x86_instruction_jump_back_to_label( io_generated_code, l_micro_kernel_config.alu_jmp_instruction, &l_loop_label_tracker );
/* second range */
libxsmm_x86_instruction_register_jump_back_label( io_generated_code, &l_loop_label_tracker );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_nloop, l_n2_block );
libxsmm_generator_gemm_rm_ac_soa_avx256_512_kloop( io_generated_code, &l_loop_label_tracker, &l_gp_reg_mapping, &l_micro_kernel_config, i_xgemm_desc,
i_arch, l_soa_width, l_n2_block );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_cmp_instruction, l_gp_reg_mapping.gp_reg_nloop, i_xgemm_desc->n );
libxsmm_x86_instruction_jump_back_to_label( io_generated_code, l_micro_kernel_config.alu_jmp_instruction, &l_loop_label_tracker );
/* reset B pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code,
l_micro_kernel_config.alu_sub_instruction,
l_gp_reg_mapping.gp_reg_b,
i_xgemm_desc->n * l_micro_kernel_config.datatype_size );
/* reset C pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code,
l_micro_kernel_config.alu_sub_instruction,
l_gp_reg_mapping.gp_reg_c,
i_xgemm_desc->n * l_soa_width * l_micro_kernel_config.datatype_size );
} else if ( (l_n1_range > 0) && (l_n2_range == 0) ) {
/* reset n loop */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_mov_instruction, l_gp_reg_mapping.gp_reg_nloop, 0 );
/* we have one range */
libxsmm_x86_instruction_register_jump_back_label( io_generated_code, &l_loop_label_tracker );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_nloop, l_n1_block );
libxsmm_generator_gemm_rm_ac_soa_avx256_512_kloop( io_generated_code, &l_loop_label_tracker, &l_gp_reg_mapping, &l_micro_kernel_config, i_xgemm_desc,
i_arch, l_soa_width, l_n1_block );
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_cmp_instruction, l_gp_reg_mapping.gp_reg_nloop, i_xgemm_desc->n );
libxsmm_x86_instruction_jump_back_to_label( io_generated_code, l_micro_kernel_config.alu_jmp_instruction, &l_loop_label_tracker );
/* reset B pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code,
l_micro_kernel_config.alu_sub_instruction,
l_gp_reg_mapping.gp_reg_b,
i_xgemm_desc->n * l_micro_kernel_config.datatype_size );
/* reset C pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code,
l_micro_kernel_config.alu_sub_instruction,
l_gp_reg_mapping.gp_reg_c,
i_xgemm_desc->n * l_soa_width * l_micro_kernel_config.datatype_size );
} else {
LIBXSMM_HANDLE_ERROR( io_generated_code, LIBXSMM_ERR_N_BLOCK );
return;
}
/* advance A pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_a,
l_micro_kernel_config.datatype_size*l_soa_width*i_xgemm_desc->lda);
/* advance C pointer */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_add_instruction, l_gp_reg_mapping.gp_reg_c,
l_micro_kernel_config.datatype_size*l_soa_width*i_xgemm_desc->ldc);
/* close m loop */
libxsmm_x86_instruction_alu_imm( io_generated_code, l_micro_kernel_config.alu_cmp_instruction, l_gp_reg_mapping.gp_reg_mloop, i_xgemm_desc->m );
libxsmm_x86_instruction_jump_back_to_label( io_generated_code, l_micro_kernel_config.alu_jmp_instruction, &l_loop_label_tracker );
/* close asm */
libxsmm_x86_instruction_close_stream( io_generated_code, &l_gp_reg_mapping, i_arch, i_xgemm_desc->prefetch );
}