Feature/qgpu float #48
Conversation
Benchmark
|
There was a problem hiding this comment.
Pull Request Overview
This PR converts floating-point precision from double to float throughout the QGPU codebase to optimize GPU performance, as GPUs typically perform better with single-precision arithmetic.
Key Changes:
- Converted function signatures, struct members, and variable declarations from
doubletofloatacross all core modules - Updated CUDA kernel functions and device code to use
floatprecision - Modified memory allocation calls to use
sizeof(float)instead ofsizeof(double)
Reviewed Changes
Copilot reviewed 15 out of 15 changed files in this pull request and generated 10 comments.
Show a summary per file
| File | Description |
|---|---|
| src/core/utils.h | Updated utility function signatures to use float parameters |
| src/core/utils.cu | Converted utility function implementations to float |
| src/core/system.h | Changed struct definitions and global variable declarations to float |
| src/core/system.cu | Updated system variables, functions, and added debug output for hot atoms |
| src/core/solvent.h | Modified solvent interaction function signatures to use float |
| src/core/solvent.cu | Converted solvent force calculations and GPU kernels to float precision |
| src/core/shake.cu | Updated SHAKE constraint algorithm to use float |
| src/core/restraints.cu | Converted restraint force calculations to float, changed numerical threshold |
| src/core/qatoms.h | Updated Q-atom function signatures to float |
| src/core/qatoms.cu | Converted Q-atom force calculations and GPU kernels to float |
| src/core/parse.cu | Changed lambda array allocation to float, refactored excluded atoms file reading |
| src/core/nonbonded.h | Updated non-bonded interaction function signatures to float |
| src/core/nonbonded.cu | Converted P-P interaction calculations to float precision |
| src/core/bonded.h | Modified bonded force function return types to float |
| src/core/bonded.cu | Converted bonded force calculations to float |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
|
|
||
| double to_radians(double degrees) { | ||
| float to_radians(float degrees) { | ||
| return degrees * (M_PI / 180.0); |
There was a problem hiding this comment.
Same issue as line 18: literals M_PI and 180.0 should use float versions to avoid double arithmetic.
| if (cos_th < -1) cos_th = -1; | ||
| f0 = sin(acos(cos_th)); | ||
| if (abs(f0) < 1.0E-12) f0 = 1.0E-12; | ||
| if (abs(f0) < 1.0E-6) f0 = 1.0E-6; |
There was a problem hiding this comment.
The threshold has been changed from 1.0E-12 to 1.0E-6. This is a significant change in numerical precision (6 orders of magnitude) that goes beyond the double-to-float conversion. Since float typically has ~7 decimal digits of precision, a threshold of 1.0E-6 might be appropriate, but this change should be intentional and documented. If this is just meant to be a type conversion, consider using 1.0E-12f instead to maintain the original threshold value (though values this small may lose meaning with float precision).
| if (abs(f0) < 1.0E-6) f0 = 1.0E-6; | |
| if (abs(f0) < 1.0E-12f) f0 = 1.0E-12f; |
| // --- dynamically read a full line --- | ||
| char *line = NULL; | ||
| size_t len = 0; | ||
| ssize_t read_len = getline(&line, &len, fp); | ||
| if (read_len == -1) { | ||
| fprintf(stderr, "Error: failed to read line from %s\n", path); | ||
| free(line); | ||
| fclose(fp); | ||
| exit(EXIT_FAILURE); | ||
| } | ||
|
|
||
| if (fgets(line, 8192, fp)) { | ||
| for (int i = 0; i < n_atoms; i++) { | ||
| bool excl = (line[i] == '1'); | ||
| excluded[i] = excl; | ||
| if (excl) { | ||
| n_excluded++; | ||
| } | ||
| } | ||
| // --- parse --- | ||
| n_excluded = 0; | ||
| for (int i = 0; i < n_atoms && i < read_len; i++) { | ||
| bool excl = (line[i] == '1'); | ||
| excluded[i] = excl; | ||
| if (excl) n_excluded++; | ||
| } | ||
|
|
||
| printf("Number of excluded atoms: %d\n", n_excluded); | ||
|
|
||
| free(line); | ||
| fclose(fp); | ||
| } |
There was a problem hiding this comment.
The init_excluded function has been significantly refactored to use dynamic line reading with getline() instead of a fixed buffer. While this is a good improvement for handling files with many atoms, this change goes beyond the scope of the float conversion PR. Consider:
- The loop condition
i < read_lenmay be incorrect - it should likely bei < n_atomsonly, sinceread_lenincludes the newline character - This functional change should ideally be in a separate PR focused on buffer safety improvements
- Added printf on line 543 for debugging should be removed or made conditional
| v1 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); | ||
| v2 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); |
There was a problem hiding this comment.
When converting from double to float, literal constants should use the f suffix to avoid implicit double-to-float conversions. For example, 1. and 2. on lines 9-10 should be 1.f and 2.f. This ensures the compiler performs float arithmetic throughout rather than converting to double for the arithmetic operations and then back to float.
| v1 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); | |
| v2 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); | |
| v1 = ( (float)(rand()) + 1.f )/( (float)(RAND_MAX) + 1.f ); | |
| v2 = ( (float)(rand()) + 1.f )/( (float)(RAND_MAX) + 1.f ); |
| v2 = ( (double)(rand()) + 1. )/( (double)(RAND_MAX) + 1. ); | ||
| v1 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); | ||
| v2 = ( (float)(rand()) + 1. )/( (float)(RAND_MAX) + 1. ); | ||
| nd10 = cos(2 * M_PI * v2) * sqrt(-2. * log(v1)); |
There was a problem hiding this comment.
The constant literals -2. should be -2.f to avoid implicit double-to-float conversion and ensure consistent float arithmetic.
| nd10 = cos(2 * M_PI * v2) * sqrt(-2. * log(v1)); | |
| nd10 = cos(2 * M_PI * v2) * sqrt(-2.f * log(v1)); |
|
|
||
| double to_degrees(double radians) { | ||
| float to_degrees(float radians) { | ||
| return radians * (180.0 / M_PI); |
There was a problem hiding this comment.
The literals 180.0 and M_PI used in the conversion should be 180.0f. Note that M_PI is typically a double constant, so you may want to use M_PI with an explicit float cast: (float)M_PI or define a float version.
| double mass_i; | ||
| float Temp_solute = 0, Tfree_solute = 0, Texcl_solute = 0; | ||
| float Tfree_solvent = 0, Temp_solvent = 0, Texcl_solvent = 0; | ||
| float Ekinmax = 1000.0 * Ndegf * Boltz * md.temperature / 2.0 / n_atoms; |
There was a problem hiding this comment.
The literal constant 1000.0 should be 1000.0f for consistent float arithmetic and to avoid unnecessary double-to-float conversions.
| float Ekinmax = 1000.0 * Ndegf * Boltz * md.temperature / 2.0 / n_atoms; | |
| float Ekinmax = 1000.0f * Ndegf * Boltz * md.temperature / 2.0 / n_atoms; |
|
|
||
| printf("Number of excluded atoms: %d\n", n_excluded); |
There was a problem hiding this comment.
This printf statement appears to be debug output that should be removed before merging, or wrapped in a conditional compilation flag (e.g., #ifdef VERBOSE or #ifdef DEBUG).
| printf("Number of excluded atoms: %d\n", n_excluded); | |
| #ifdef DEBUG | |
| printf("Number of excluded atoms: %d\n", n_excluded); | |
| #endif |
| printf(">>> WARNING: hot atom %d, %f %f %f %f %f %f: %f %f %f %f %f\n", i, velocities[i].x, velocities[i].y, velocities[i].z, | ||
| coords[i].x, coords[i].y, coords[i].z, ener, ener/Boltz/3, Ekinmax, mass_i, Ekinmax / (0.5 * mass_i)); | ||
| // printf(">>> WARNING: hot atom %d: %f\n", i, ener/Boltz/3); | ||
| } |
There was a problem hiding this comment.
[nitpick] This appears to be debug code that was uncommented during development. The extensive printf with many parameters (lines 732-734) and the commented-out original warning (line 734) suggest this was temporary debugging. Consider cleaning this up: either commit to the new detailed warning format or revert to the simpler original message.
| printf(">>> WARNING: hot atom %d, %f %f %f %f %f %f: %f %f %f %f %f\n", i, velocities[i].x, velocities[i].y, velocities[i].z, | ||
| coords[i].x, coords[i].y, coords[i].z, ener, ener/Boltz/3, Ekinmax, mass_i, Ekinmax / (0.5 * mass_i)); | ||
| // printf(">>> WARNING: hot atom %d: %f\n", i, ener/Boltz/3); | ||
| } |
There was a problem hiding this comment.
[nitpick] Similar to the solute section above, this appears to be debug code with the extensive printf and commented-out original warning. The same cleanup recommendation applies.
Average energy of the last 10000 steps for 30A water ball.
|
Stable Test`[Run 1/20] Setting random_seed=1 Summary across runsRuns completed: 20 |
No description provided.